Click here for : Part1 - Part 2 - Part 3 - Part 4 - Part 5 - this is Part 6

At its most generous estimate, the entire AI industry is running at $250 billion a year in revenue. [1] Yet, according to press reports of its forecast, Moody’s expects just the six builders it tracks to spend $785 billion on facilities and equipment this year, and about $1 trillion the next. [2]

Revenue covers probably only 25% of the spending. The estimate dates from August, and on the independent tally Part 1 used, revenue has grown at least 35% a quarter for four quarters [3] but the spending is growing too. For now the builders’ own businesses pay the difference: search, advertising, software, shopping and, increasingly, renting out computing. I expect this to end not as a spectacular overnight crash, but as a bust that builds in sequential stages as the funding pockets run dry.

The Four Pockets of Capital

Four sources of money matter most and each has a limit approaching fast.

• Spare cash is nearly gone. The businesses still pay for almost everything: the cash their operations generate covered 104% of Microsoft’s, Amazon’s, Alphabet’s and Meta’s capital spending in the June quarter, down from about 150% in 2025. At the end of 2025 they had about $60 billion a quarter left over after capital expenditures. By the June quarter of 2026, that cushion had vaporized to under $7 billion. [4] Analysts already expect the free cash flow of those four plus Oracle, after stock pay, to fall to about minus $265 billion on a trailing basis by the third quarter of 2027, and to recover to about $505 billion by 2030. [5] The problem? I do not think that trough gets funded on today’s terms.

• Debt has a ceiling. Since January 2025, the three builders that borrow in the bond market have raised about $280 billion. [6] Add Microsoft, and the four have also signed $836 billion of leases that have not started yet. Those payments are not yet on the balance sheet. [6] Held to the debt levels typical for companies rated BBB or BB, the four companies’ room to borrow more is anywhere from nothing to about a trillion dollars, depending on which test is used and whether the leases count. These are comfort levels, not limits. [7] On a range of assumptions about how much of their spending they borrow, that room runs out between 2027 and 2030, and in early 2028 in my middle case.

• Equity can fill only a little. Alphabet, the largest of them, has already begun selling shares, about $45 billion placed in June, partly to fund the build. [8] The two big labs’ listings would add more. They are the first test below, and they could go either way.

• Vendor credit is the weakest link. The companies selling the chips and the compute are now lending to, or standing behind, the customers buying them. Nvidia has agreed to cover up to $105 billion of a customer’s lease payments if that customer defaults. The guarantee starts with the first leases, expected from 2028, and OpenAI has agreed to reimburse Nvidia for anything it pays, which is thin comfort if OpenAI is the party that cannot pay. [9] Amazon has invested $50 billion into a lab that turns around and buys Amazon’s computing. [10]

The money is moving in a circle. It works beautifully—provided the weakest members of that circle, the independent AI labs, can continually raise outside capital. They have the thinnest cushion. OpenAI agreed a $122 billion round in March at an $852 billion value, and Anthropic raised $65 billion in May at $965 billion. [11] Neither publishes accounts, so I cannot say how many months that lasts.

There are other sources: private credit, joint-venture debt and customer prepayments. Each delays the date and a government could also decide the labs are strategic and fund them, although unlikely under a democratic majority.

There is no fraud here and every piece is disclosed, legal, and individually rational. Nvidia booked $164 billion of data-center revenue in its last two quarters and $117 billion of operating profit across the company. [9] That is real cash. The best reply to me is that this profit shows the product earns its price. It shows buyers paid it. Whether they earn it back is the question of Part 1. Vendor financing is a discount that never shows in the list price or the reported margin until it is paid. A structure that pushes all the true costs into the future still hides them. In the year to July 2000, Cisco earned record profits selling the routers that built the internet. The next year it lost $1.0 billion and took a $2.25 billion charge for excess inventory. [12] Profit that exists because the spending continues is not proof that the spending will be profitable.

Buying Insurance on Monopolies

The bull case requires AI to become a utility everyone pays for, akin to a smartphone or streaming service. It models hundreds of millions of subscribers paying $100 a month by 2030, with infrastructure owners keeping 75 cents of every dollar, and capital spending slowing down to let returns catch up.

That narrative fails on contact with reality.

First, the price has to beat free. Google, Meta, and Apple give their core assistants away and charge only for heavier use or add-ons. Even if 600 million people paid $30 a month, that yields $216 billion a year. That is more than Netflix, Spotify, Disney’s streaming, Amazon’s subscriptions, Microsoft’s home Office, and ChatGPT collect from consumers worldwide combined, about $162 billion. [13] A $100-a-month model defies consumer gravity. That is the consumer version of the story. The business version is the one Part 1 tested: replacing payroll is the only wallet of those I could size that is large enough, and it moves slowest.

Second, the margin profile is wrong. Large listed software companies spend 20 to 33 cents of each revenue dollar on delivering the service. [3] If landlords keep 75 cents, the math breaks. And finally, nobody is going to voluntarily slow down their spending first.

They aren’t spending because the math works, they are spending because they fear their oligopoly could break.

For a decade, big tech earned astronomical returns because their moats—search, social networks, enterprise software—were practically unbreachable. An AI assistant that writes the document or answers the query without requiring the user to view an ad threatens those moats. The executives say it openly. In July 2024 Alphabet’s chief executive told investors that the risk of under-investing was dramatically greater than the risk of over-investing. Meta’s chief executive said in September 2025 that misspending a couple of hundred billion dollars would be very unfortunate, but that the risk was higher on the other side. [14]

Alphabet’s and Meta’s cushion is search, advertising and social media. Those businesses depend on people arriving at their pages. If an AI assistant answers the question, the click does not happen. In a Pew Research study of 900 American adults in March 2025, only 8% of visits to a Google results page with an AI summary included a click on a traditional result. Without the summary, 15% did. [15] That must be pretty scary for Google. The study covers Google only, and the revenue shows no damage yet. Alphabet’s search and other revenue grew about 17% in the June quarter, and Meta’s advertising revenue grew 27.5%. [16] So it’s clearly not an immediate threat but more of a preemptive move when faced with a technology they can’t suppress.

This isn’t an investment evaluated on its cost of capital. It is an insurance premium. They are paying to ensure they own the new “front door” to the internet. So the return sums in Part 1 do not stop them. Two things can: the fear lifts, or the money runs short. I expect the money to run short first.

But as the mainframe era showed us, a cash machine doesn’t have to break to lose its value. In the early 1990s, corporate buyers delayed mainframe upgrades while waiting for client-server networks to mature. About a quarter of the mainframe demand that had been expected never arrived, before the new technology could do the work. [17] Search works the same way, just one layer removed: it is free to the user, but paid for by advertisers buying clicks. If an AI assistant answers the question directly, the click never happens. The moment advertisers and investors expect that traffic to shift, revenue growth stalls and the valuation multiple compresses—long before the core business actually breaks. The market already pays less for these four than it did. Remember, stocks peak before Capex.

The premium the market pays for these four, over the rest of the index, has been shrinking for most of ten years. Microsoft, Amazon, Alphabet and Meta together traded at about 2 times the S&P 500’s price for each dollar of operating profit in 2017 and 2018. A rebound took the figure back to about 1.7 in 2023. So far this year it averages about 1.2. It touched 1.07 at the end of June and was 1.25 on 9 October. [18] On reported net income they are already below the market, because gains on private stakes inflate Alphabet’s and Amazon’s profits. Their return on capital fell too, from about 35 cents on the dollar to about 23. [18] I do not think the market has priced the case where spending does not slow.

Racing to Destroy Their Own Scarcity

A luxury brand protects its pricing power by artificially capping production. The AI industry is heading for a trillion dollars a year of spending to destroy its own scarcity as fast as humanly possible, selling a product whose price per unit is falling fast. Part 1 set out the risk that the price falls faster than use grows.

The cleanest evidence is Ramp, a payments platform that publishes weekly token volumes and spending for the businesses that use its AI token-spend tool. From the last week of July to the first week of October, those businesses bought about 65% more tokens and spent about 2% more. On four-week averages, tokens are up about 65% and dollars are up about 5%. The average price paid per million tokens fell from $0.75 to $0.46, partly because buyers moved to cheaper models and partly because the same models got cheaper. [19] The price fell as fast as use rose, so spending did not. This does not yet show across the market. The two big labs’ own annualised revenue kept rising through the summer, on press-reported figures, and Ramp sees token spend only for the businesses that use its tool, so it cannot see seat licences or contracts billed outside it. [20]

Efficiency is the enemy of the landlord. Many everyday queries do not need the best model. Hazy Research tested small models running on a laptop-class chip. On a set of single-turn chat and reasoning questions, the best of them matched or beat the reference answer 71% of the time in 2025, up from 23% in 2023. [21] Routing software now sends each request to the cheapest model able to handle it. AT&T’s routing cut the cost of coding and some other tasks by as much as 56% for a 2% drop in quality, as reported. [22] China makes the point harder. Its labs give away models only a few months behind the best. [3]

Subscribe now

Two prices matter here. The price of a model’s answer is falling. The price of machine time is firmer, because chips and power are short: H100 rents at neocloud and marketplace providers slid through 2025 and have partly recovered in 2026, to about $2.83 an hour in October on Silicon Data’s index. Hyperscaler list rents for the same chip have kept falling. [23] The labs sit between the two. They rent scarce machine time and sell answers at falling prices.

A build-out of this size is financed on the assumption that machine time stays scarce and expensive. Efficiency is the enemy of the people who built the machines. If models keep doing more with less, the world needs fewer machines, and more of the machines already built will have to be rented out for whatever the market will pay.

The Fiber vs. The Routers

In the aftermath of the telecom crash, the physical infrastructure survived and some of its owners did not. Traffic on the internet kept growing, though far slower than forecast. Global Crossing wrote off $17.2 billion of assets in 2001, citing among other things overcapacity in the telecommunications market. Its shareholders were wiped out, but the network itself kept running under new owners. [24]

Fiber lasted because glass does not age. Its owners did not last: Global Crossing filed for bankruptcy in January 2002, less than two years after the telecom index peaked in March 2000. [24] [3]

The routers—the boxes moving the traffic—did not last as long. Level 3, one of the biggest builders, wrote off its operating equipment over three to seven years and its fiber over seven to twenty-five. [24] Capacity that arrives years ahead of its demand does not sit patiently on a shelf. The customers bought much less. Cisco’s router sales fell almost a third in two years. Its overall gross margin was higher in 2003 than in 2000, so this was not a margin squeeze, and its filings put the fall down to carriers spending less. Its share price still took a quarter-century to regain its peak. A supplier’s profit follows the spending, not the moat. [12]

This current build-out splits along the exact same fault line. The land, the grid connections, and the concrete shells are this cycle’s fiber. On most owners’ own schedules they last decades. They keep their value only if demand keeps pace, because efficiency cuts the power and buildings needed per unit of AI as it cuts the chips. The GPUs are this cycle’s routers, and rented compute hours are this cycle’s circuits. Capacity prices are what collapsed last time. The GPUs are where most of the capital has been spent, and I do not expect a financial payoff inside a warehouse of five-year-old chips.

The most likely ending is not a spectacular failure of artificial intelligence. The technology will succeed. The capacity will fill up—years after the payment schedules required it to. Prices will be set by an industry drowning in overcapacity. The machines will get used but a large part of the capital will be lost.

The Physical Speed Limit

Money is the first brake on this build. The physical world is the second, and it is already pressing.

Announced capacity is not delivered capacity. Eaton notes that while announced US data center projects have reached 307 gigawatts, only about a fifth will convert in the near term. [25] The rest is mostly slated for 2028 or later. [25] In Texas, data centers that came online between 2022 and 2024 drew, at their peak, an average of 49.8% of the power they had requested. [25]

As you look further out, the bottlenecks shift. Through 2027, the chip makers say packaging and memory are the constraint, and Micron expects memory to stay short through 2028. From 2028 the evidence points to power and electrical equipment as well. A large gas turbine ordered today ships in 2030 or 2031 and is running on site in 2032 or 2033, by GE Vernova’s own account. [26] Money can buy its way through these limits, but slowly.

This physical speed limit matters for three reasons:

• It slows the break. Capital that cannot be deployed sits in construction and inventory. This is why I expect a bust that builds in stages, not a sudden crash.

• It shifts the rent. While memory and turbines remain scarce, their manufacturers capture the margin. The AI labs are trapped in the middle, renting scarce machine time to sell tokens at falling prices.

• It moves the dates. In the contracts I can read, billing starts when the capacity is delivered and accepted. A late campus delays the invoice. In the one data-center compute contract I have read in full, a long enough delay by the supplier lets the customer walk away, while delays from chip shortages or force majeure only move the date. [27]

When these physical limits finally clear, a massive wave of capacity will hit the market within the same few years, after prices have been falling. A speed limit merely postpones the reckoning; it does not cancel it.

The Sequence of the Bust

The mainframe era provides the blueprint for who loses first. In the 1960s, independent lessors bought IBM machines and rented them out, most of them amortizing the hardware over ten years while IBM used six. [28] When IBM announced a new generation of machines, by IBM’s own account in court the ten-year schedules proved too long. Five of the lessors wrote off more than $200 million between them from 1970 to 1974. [28] IBM’s own grip broke more than a decade later.

Today’s lessors are the owners of the machines who rent them out: the smaller cloud operators like CoreWeave, who amortize their servers over six years, and the lenders and vehicles behind them. [29] The labs are their tenants. In the 1960s IBM’s own customers could hand machines back on 30 days’ notice, so IBM carried the volume risk. The independent lessors leased on one- and two-year terms. [28] The labs have signed longer commitments than either, so the risk moves to whether they can keep paying. If GPUs age faster than that schedule, the owners inherit the 1970s lessor problem, and the tenants’ payments are what has to carry it. The labs are squeezed first, between scarce hardware and falling prices. The owners who rent to them are hit through the labs. The platform with lock-in breaks last.

I expect the stress to cascade through five distinct tests. The dates are the earliest ones. The speed limit can push them out:

1. The Listings: Both OpenAI and Anthropic have reportedly filed confidentially to go public, and OpenAI’s chief executive has said it will not list in 2026. [30] A listing replaces private belief with public quarterly testing. A strong listing is the bulls’ best outcome: it raises money, and a satisfactory credit rating would end Nvidia’s guarantee. It would count against me. A listing that is delayed, withdrawn or priced below the last private round would count for me.

2. The Step-Ups: Watch the private funding rounds. The moment a leading lab raises capital at a flat valuation, or grants heavy downside protection to secure the cash, the private ladder has stopped climbing. Flat means a priced round at or below the previous valuation. A smaller step-up than the last round is the earlier warning. A bigger round at a higher step-up counts against me.

3. The First Amendment: Customers have placed roughly $2.4 trillion in orders with four cloud companies, a total that covers all of their business and not only AI. [31] It is also the strongest demand evidence in the market, with a catch. Amazon’s backlog rose from $364 billion to $496 billion in one quarter, and in the same quarter Anthropic, one of the two weakest buyers, announced a commitment of more than $100 billion over ten years. [31] The filing does not say how much of the rise that is. How much of the whole belongs to the two big labs is the open question. At some point one large compute contract will be publicly rescheduled. One public amendment would signal the vulnerability of the rest. An exit because a supplier was late tells me about the supplier, not about demand, so it does not count. A customer cutting or delaying its own commitment does.

4. The 2027 Revenue Test: Compute contracts start billing when the capacity is delivered. I cannot see the labs’ contracts and I assume they are take-or-pay. If so, deals signed in 2025 and 2026 start to produce fixed monthly invoices from 2026 to 2028, with 2027 my central case. If the contracts are rigid, the labs’ cash reserves will be drained rapidly. If the first heavy invoices are paid on schedule through 2028, I am wrong.

5. The 2028 Hardware Test: Equipment makers (Nvidia, Broadcom) are paid by the clouds, not the labs, so a lab’s default does not touch a sale. It does touch the guarantees they have written: Nvidia’s starts with the first leases in 2028, and Broadcom has a backstop of its own. [9] What costs them most is the hyperscalers deciding they don’t need the next batch of machines.

Columbia Business School’s Shivaram Rajgopal ran a severe, 24-month scenario on eight companies exposed to this spending. It assumes cloud customers use 20% less compute and pay 15% less, equipment orders fall 35% with a 10% price drop, and investors pay 30% to 45% lower multiples for what is left. On those assumptions the eight lose $11.3 trillion, more than half their combined market value. It is a scenario, not a forecast. [32]

Cisco’s sales to service providers broke in the winter of 2000 to 2001, about a year after the telecom index peaked in March 2000. [12] [3] The test is which shows first. This is the order in which the evidence arrives, not the order in which share prices move. Prices move on expectations, and a funding scare or one big order cut could hit the equipment makers first.

My Forecast and What Would Prove It Wrong

Stocks go sideways for now, peak before the spending does, and then fall quite hard. The build has another year or two to run, because compute is still short. [23] [26] Stocks peaked first in 2000: the technology index topped in March [33] and the equipment makers’ sales broke a year later. [12] If the build peaks in 2028, stocks peak in 2027 and we have a bear market.

This is more than a sector correction. Seven companies in the build are about 30% of the S&P 500. [3] In 2000 technology fell about 51% while the rest rose about 6%, and the rest fell about 10% in 2001. [33] This time the rest falls with it, because the wealth is larger. US households hold about $74 trillion of equities, against $13.7 trillion in 2000: 227% of GDP against about 137%. [39] The top half of earners carries spending, and the bottom half is already squeezed. The response will be weaker than in 2001, with inflation above target since early 2021 and a federal deficit near 6% of GDP. [40]

The bubble is in earnings more than in multiples. Corporate profits after tax are 11.9% of GDP, the highest since 1947, against 10.1% in 2019. [35] At the 2019 share, profits would be about 15% lower on the same economy. Supplier margins are high because supply is short, and they fall as supply arrives. But because the bubble is in earnings, unlike 1999, I expect a lengthy bear market that will be the ultimate pain trade for an entire generation that has learned to “buy the dip”.

I will continue to monitor memory prices first, because the shortage is tightest there. [26] Then token prices: Ramp’s weekly business data shows volume up 65% since July while spending is flat, because the price paid per token has fallen by more than a third. [19] Then the five tests in this piece, and credit spreads, which the forty-year piece named as the first sign of the scare. [36] Long yields are high because the build absorbs savings. They will ease when the spending turns, and that is where I would add long duration Treasuries.

Subscribe now

What would prove me wrong or extends the cycle?

• Revenue arrives. If AI-funded advertising, government contracts or faster cloud growth show up as large disclosed revenue, the arithmetic softens. Part 1 sized the need at $420 billion to $560 billion of extra annual revenue for every $1 trillion of building. Azure, AWS and Google Cloud accelerated from 39%, 24% and 48% growth to 43%, 37% and 82% between the December 2025 and June 2026 quarters, the last including its own chips. [37] That pace cannot last as long as the bill needs. Look at the cloud companies’ quarterly revenue breakdowns and published government awards.

• Chips last longer. Seven-year server lives instead of five cut the yearly cost by about 18%. [3] That softens the call and does not end it. Look in annual reports.

• Margins hold. Profits after tax averaging above 11% of GDP through 2028 (11.4% on the last four quarters) would make today’s earnings the norm. The Bureau of Economic Analysis publishes the series every quarter. [35]

• Hardware cracks first. If Nvidia cuts its sales forecast before the cloud backlogs shrink or the labs run short of cash, my order is backwards.

The Collision of Fear and Imagination

Paul Kedrosky says the largest financial bubbles of the past two hundred years tended to share a few things: loose credit, a great technology story, and sometimes a real estate component and a policy angle. His claim is that this is the first to sit at the intersection of all of them. [38]

I read the ingredients in this cycle as follows. The loose credit is lenders underwriting the rent on these leases as a spread over Treasuries, and the vendors’ own guarantees. The technology is a genuine exponential. The real estate is the land, power, and data centers. The policy is the tax code, which now lets companies deduct the full cost of qualifying equipment in the year they buy it. [38]

When the history of this cycle is written, the villain won’t be fraud, nor will it be a failure of the technology itself. It will be recorded as a collision between the two oldest forces in markets.

At the top, fear. An oligopoly of the world’s most profitable companies—businesses that spent a decade buying back stock because they had nothing left to build—suddenly reversed course. They became convinced that this new platform would decide their survival. And nothing on earth is less price-sensitive than a company fighting for its life.

Underneath, imagination. Humans do not possess an intuitive grasp of exponential math. When a real exponential curve appears, the market eventually stops doing the arithmetic and simply chooses to believe.

Fear signed the contracts. Imagination priced them. And the economy, which feels neither, will deliver the revenue at its own speed.

References

Figures are from company filings and announcements, court records and published studies, except where a reference marks them as press-reported. The $250 billion revenue ceiling is Aswath Damodaran’s. The room-to-borrow arithmetic, the valuation series and the 2027 and 2028 split of the stress test are my own.

[1] Aswath Damodaran, “AI’s Bar Mitzvah Moment: From Hype & Hope to Business Questions!”, Musings on Markets, 20 August 2026 (updated 21 August): current AI revenue “caps out at about $250 billion” on the most upbeat estimates.

[2] Moody’s Ratings hyperscaler capital-expenditure forecast, press-reported (CNBC and DataCenterDynamics, 24 July 2026; Moody’s report not independently obtained): Microsoft, Amazon, Meta, Alphabet, Oracle and CoreWeave, $785 billion in 2026 and close to $1 trillion in 2027.

[3] Stray Narratives, Issue 33 (Part 1): reference 22 (independent tally of annualised AI revenue, global excluding China: $175 billion in June 2026 after four quarters of growth of 35% or more); reference 20 (cost of revenue as a share of revenue: Datadog 20%, Salesforce 22%, ServiceNow 22%, Snowflake 33%); reference 14 (Epoch AI: seven-year server lives against five cut the yearly cost by about 18%); reference 7 (State Street SPDR S&P 500 ETF holdings, 29 September 2026: seven companies about 30%); references 5 and 28 (Epoch AI on the falling price of a given level of performance; open-weight models about three months behind the frontier on average (Epoch AI, October 2025); the DeepSeek, Qwen and Kimi models are published free of charge (model cards, as in reference 28)); reference 8 (Couper, Hejkal and Wolman, Federal Reserve Bank of Richmond Economic Quarterly, Fall 2003: the Nasdaq telecom index peaked on 10 March 2000).

[4] Free cash flow (operating cash flow less purchases of property and equipment), Microsoft, Amazon, Alphabet and Meta, from EODHD quarterly data tied to the companies’ Forms 10-Q and 10-K: about $60 billion in each of the third and fourth quarters of 2025, $21 billion in the first quarter of 2026, $6.7 billion in the second. Operating cash flow covered 104% of the four companies’ capital expenditure in the June quarter of 2026 and about 150% in 2025 (EODHD cash-flow statements; author’s calculation).

[5] Counterpoint Global (Michael Mauboussin and Dan Callahan), Morgan Stanley Investment Management, “To Free or Not to Free (Cash Flow),” 17 September 2026 (FactSet consensus at 31 August 2026, as reported; the forecasts cannot be re-derived): trailing free cash flow of Amazon, Alphabet, Microsoft, Meta and Oracle bottoms near -$265 billion in the third quarter of 2027 and recovers to nearly $505 billion by 2030, all after stock-based compensation; consensus operating profit of Nvidia keeps rising to 2030 and of Micron to 2029.

[6] Gross bond and note issuance from 1 January 2025 to 9 October 2026, from each company’s Forms 8-K, prospectus supplements and 10-Q debt notes: Alphabet about $114 billion, Amazon about $113 billion, Meta $55 billion, total $282 billion (Microsoft issued none). Leases signed but not yet commenced, undiscounted, at 30 June 2026 (Forms 10-Q; Microsoft Form 10-K): Alphabet $91.0 billion, Amazon $137.2 billion, Meta $279.0 billion, Microsoft $329.1 billion, total $836.3 billion.

[7] Author’s calculation, not a forecast. Room: additional debt that would bring the four companies to 70% debt-to-equity, 7 times operating income to interest, or 20% to 25% debt-to-assets, which a subscription research firm gives as medians for BBB and BB rated large companies (comfort levels, not limits; the four are rated far above BBB). From nothing to about $1 trillion depending on the test and whether leases count. Months until new borrowing at 25% to 40% of capital spending uses the room: April 2027 to November 2030. Central case ($500 billion of room, one third of spending borrowed, company guidance): March 2028.

[8] Alphabet Inc., Form 8-K filed 4 June 2026 ($84.75 billion announced, of which $44.75 billion placed and a $40 billion at-the-market programme, mainly to meet tax obligations on employee equity awards; none sold by 30 June); Form 10-Q, quarter ended 30 June 2026.

[9] NVIDIA Corporation, Form 8-K filed 17 August 2026 (Item 1.01, residual value guaranties on leases for which an affiliate of OpenAI is tenant: payment obligation cumulatively capped at $105 billion; effective on lease commencement, ready-for-service expected beginning in 2028 and the first phase in Nvidia’s fiscal 2029; OpenAI has agreed to reimburse); Form 10-Q, quarter ended 26 July 2026 (six-month Data Center revenue $164.3 billion; operating income, whole company, $117.3 billion). Broadcom Inc., Form 10-Q, quarter ended 2 August 2026 (Financial Guarantee: backstop of a customer’s lease obligations, maximum potential liability about $29 billion).

[10] Amazon.com, Inc., Forms 10-Q, quarters ended 31 March and 30 June 2026 (Note 2, OpenAI and Anthropic).

[11] OpenAI, “OpenAI raises $122 billion to accelerate the next phase of AI,” 31 March 2026 ($852 billion post-money valuation; figures as reported, the company page was not retrievable; Amazon’s Form 10-Q for June 2026, Note 2, corroborates the structure). Anthropic, “Series H,” anthropic.com/news/series-h, 28 May 2026: $65 billion raised at $965 billion post-money, against $380 billion at the previous round; run-rate revenue above $47 billion.

[12] Cisco Systems, Forms 10-K for fiscal 2000 to 2003 (July year-ends): net income $2.67 billion in 2000, net loss $1.01 billion in 2001, excess-inventory charge $2.25 billion; router net sales down 31.5% from fiscal 2001 to 2003. Share price, EODHD (CSCO.US, split-adjusted): closing high 27 March 2000, first close above it 10 December 2025. Total company gross margin 64.4% in fiscal 2000 and 70.1% in fiscal 2003 (no router-level margin is disclosed). Form 10-Q for the quarter ended 28 April 2001: sales to service providers “were adversely affected during the second and third quarters of fiscal 2001”.

[13] Consumer subscription revenue, each on its latest full year or run rate: Netflix, revenue, fiscal 2025 (Q4 2025 shareholder letter); Amazon, subscription services net sales, 2025 (Form 10-K); Disney, Entertainment direct-to-consumer revenue, fiscal 2025 (Q4 release, 13 November 2025); Spotify, Premium revenue, first half of 2026 doubled (Form 6-K); Microsoft, Microsoft 365 Consumer, fiscal 2026 (Form 10-K); ChatGPT consumer, at most 60% of OpenAI’s $2 billion a month run rate (OpenAI, 31 March 2026; author’s estimate). Total about $162 billion.

[14] Sundar Pichai, Alphabet Q2 2024 earnings call, 23 July 2024: “the risk of under-investing is dramatically greater than the risk of over-investing for us here” (transcript as published by The Motley Fool). Mark Zuckerberg, Access podcast, 18 September 2025, press-reported (Business Insider): “I actually think the risk is higher on the other side”.

[15] Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results,” 22 July 2025 (900 US adults, March 2025; 8% of visits to a results page with an AI summary included a click on a traditional result, 15% without).

[16] Alphabet Inc., Form 10-Q, quarter ended 30 June 2026 (“Google Search & other” revenue $63,271 million against $54,190 million, up 16.8%); Meta Platforms, Inc., Form 10-Q, quarter ended 30 June 2026 (advertising revenue $59,363 million against $46,563 million, up 27.5%).

[17] Timothy Bresnahan and Shane Greenstein, “The Competitive Crash in Large-Scale Commercial Computing,” NBER Working Paper 4901, October 1994, pp. 15 to 19 and 44 to 45: share of mainframe sites with a persistent increase in installed processing power (MIPS) over a three-year window, 1984-86 against 1989-91, fell from 33% to 25%; “About one fourth of total expected mainframe demand has gone away”; the decline “is not actual switches to client/server”.

[18] Stray Narratives calculations on EODHD prices and company filings: summed market value over summed operating income of Microsoft, Amazon, Alphabet and Meta, divided by the same ratio for S&P 500 members at each month end. Averages 2.03 (2017), 2.08 (2018), 1.70 (2023), 1.17 (2026 to 9 October); 1.07 on 30 June 2026 and 1.25 on 9 October. On GAAP net income 1.49 (2023) and 0.86 (9 October 2026), because of gains on private stakes. Return on capital about 35 cents (2023) and about 23 cents (twelve months to June 2026).

[19] Ramp Economics Lab, AI Index (ramp.com/data/ai-index), pulled 15 October 2026: weekly token volume and spend of a sample of Ramp businesses that use its Token Spend Management product (not representative of Ramp’s total AI spend, per Ramp; the sample changes over time and earlier weeks have been restated). Week of 27 July 2026 $33.2 million and 44.1 trillion tokens; week of 5 October $33.7 million and 72.5 trillion tokens (spend +1.5%, tokens +65%; four-week averages tokens +65%, spend +5%); blended price per million tokens about $0.75 to $0.46 (-38%).

[20] Lab run rates, press-reported. Anthropic: above $47 billion in May 2026 (company-stated, Series H announcement) and above $65 billion at the end of July (company statement to investors, 17 August 2026, via Bloomberg). OpenAI: about $2 billion a month in March (company announcement, as reported); above $40 billion by mid-August (Bloomberg, 13 August 2026; Aswath Damodaran, 20 August, citing OpenAI’s own estimate).

[21] Jon Saad-Falcon et al. (Hazy Research), arXiv 2511.07885 (v7, 1 October 2026): share of 1 million single-turn chat and reasoning queries on which the best local model matched or beat frontier models (win/tie rate, partly judged by an LLM), 23% in 2023 (Mixtral-8x7B) and 71% in 2025 (GPT-OSS-120B on an Apple M4 Max). Multi-turn, agentic and long-context work is excluded.

[22] The Information, “AT&T is Using Open Source Models to Curb Anthropic Bills,” August 2026 (paywalled; figures as attributed in PYMNTS, 20 August 2026). Press-reported.

[23] Silicon Data, SDH100RT H100 rental index (daily, neocloud tier), silicondata.com: $2.83 per GPU-hour on 11 October 2026. Marketplace median $2.58 in the second half of 2024 and $1.95 in the second half of 2025; hyperscaler median $9.34 in the second half of 2024 and $6.26 in the second half of 2025, with hyperscaler rates falling in every region since the first quarter of 2025 (“H100 Rental Prices by Region,” 31 July 2026).

[24] Global Crossing Ltd., Form 10-K for fiscal 2002, filed 8 December 2003 (the company filed no 10-K for fiscal 2001): asset impairment charges of $17.181 billion in 2001, of which $8.573 billion goodwill and intangibles; holders of common and preferred shares receive nothing under the plan. FCC Order DA 03-3121, 8 October 2003. Level 3 Communications, Form 10-K for 2001: operating equipment 3 to 7 years, network infrastructure including fiber 7 to 25 years. Traffic growth: Couper, Hejkal and Wolman, Federal Reserve Bank of Richmond Economic Quarterly, Fall 2003, as in Issue 33 reference 8. Chapter 11 petitions filed 28 January 2002; the impairment is attributed in the filing, among other things, to overcapacity in the telecommunications market.

[25] Eaton Corporation plc, Q2 2026 earnings call, 31 July 2026 (307 GW is a tracked pipeline of announced and planned projects; “only roughly 20%” converts near term; the majority 2028 and beyond). North American Electric Reliability Corporation, 2025 Long-Term Reliability Assessment (January 2026), ERCOT: data centres energised 2022 to 2024 reached 49.8% of requested megawatts at peak.

[26] TSMC, Q2 2026 earnings call, 16 July 2026 (packaging capacity “so tight that now it limits my customers’ growth”); Micron, fiscal fourth-quarter 2026 prepared remarks, 30 September 2026 (memory supply “constrained” through 2027 and 2028); GE Vernova, Q2 2026 webcast, 22 July 2026 (gas turbines “mostly sold out through ’30”; chief executive: heavy-duty units “shipped in ’30, ’31 ... commissioned in ’32 and ’33 at site”).

[27] IREN Limited, Partner Statement of Work with Microsoft, 2 November 2025 (Exhibit 10.3, Form 10-Q filed 5 February 2026): firm minimum commitment, with a termination right and refund on late delivery by the supplier; delays from force majeure (including industry-wide shortages) or a failure by Nvidia to deliver chips extend the date and do not open the termination right. CoreWeave, Forms 10-Q for 2025: OpenAI may terminate on “repeated failure to meet availability of service requirements”; the sentence appears in the 2025 filings and not in later ones. Order forms are not filed.

[28] Greyhound Computer Corp. v. International Business Machines Corp., 559 F.2d 488 (9th Cir. 1977): Greyhound depreciated IBM equipment at 10% a year, IBM over six years. United States v. IBM (S.D.N.Y.), Defendant’s Exhibit 14971, “Historical Narrative,” 1980, an advocacy document prepared for IBM’s defence: lessors’ ten-year schedules too long once System/370 arrived; five lessors wrote off more than $200 million, 1970 to 1974 (the total is the author’s sum). Telex Corp. v. IBM, 510 F.2d 894 (10th Cir. 1975): IBM’s rental customers could return equipment on 30 days’ notice, while competing lessors’ equipment was generally leased for terms of one, two or more years.

[29] Depreciable lives of servers and network equipment, each company’s Form 10-K: Alphabet six years; Amazon five to six; Meta five to five and a half; Microsoft two to six; Oracle and CoreWeave six.

[30] Press-reported: Sam Altman, Fortune, 12 September 2026 (“not 2026”); OpenAI confidential draft registration, CNBC, 8 June 2026; Anthropic confidential draft registration, NPR, 1 June 2026; Anthropic timing, Fortune and TechCrunch, 28 to 29 September 2026. The registrations are non-public.

[31] Remaining performance obligations, total company and not AI-only: Microsoft Form 10-K (year ended 30 June 2026, Note 12; $684 billion); Oracle Form 10-Q (quarter ended 31 August 2026; $664 billion); Alphabet Form 10-Q (30 June 2026; $519.5 billion); Amazon Form 10-Q (30 June 2026; about $496 billion). Sum about $2.4 trillion (author’s sum). Amazon’s figure was about $364 billion at 31 March 2026 (Form 10-Q, first quarter). Amazon Form 10-Q (June 2026, Note 2): AWS and Anthropic expanded their commitment by more than $100 billion over 10 years in the quarter.

[32] Shivaram Rajgopal (Columbia Business School), “What an AI Slowdown Could Cost Investors,” 15 September 2026: eight-company severe scenario over 24 months, $11.3 trillion lost of about $19.8 trillion; scenario analysis, not a forecast; the loss includes assumed cuts of 30% to 45% to valuation multiples as well as the volume and price shocks in the text. The $6.2 trillion and $5.1 trillion split and its dating are the author’s allocation, not Rajgopal’s.

[33] S&P 500 Information Technology sector index and S&P 500 price index (EODHD SPLRCT.INDX, GSPC.INDX), 24 March to 29 December 2000: technology about -51%, S&P 500 about -14%. The rest of the index is the author’s calculation, chain-linking the two series with technology at about 34% of the index in March 2000 (about +6%, a range of +5% to +8% for weights of 33% to 36%); 2001: about -10%.

[34] EODHD, IBM.US adjusted close: peak 15 January 1970 to low 13 August 1970 (about -41%); peak 13 February 1973 to low 13 September 1974 (about -57%; on price alone about -42% and -58%). IBM Archives, “IBM Highlights, 1970-1984” (net earnings $934 million in 1969, $1.01 billion in 1970, $1.27 billion in 1972, $1.83 billion in 1974). S&P 500 price index (EODHD GSPC.INDX): 91.68 to 74.76 (-18.5%) over the 1970 window and 116.78 to 65.20 (-44.2%) over the 1973-74 window. Jeremy J. Siegel, “Valuing Growth Stocks: Revisiting the Nifty Fifty,” AAII Journal, October 1998 (price-earnings ratio of 35 in the early 1970s; “worth only 17.1 times earnings” on the following 26 years).

[35] US Bureau of Economic Analysis via FRED, all US corporations including financials, corporate profits after tax with inventory valuation and capital consumption adjustments (CPATAX) divided by nominal GDP, vintage of 30 September 2026: 11.9% in the second quarter of 2026 (previous high 11.4%, second quarter of 2021); average 6.9% for 1947 to 2019, 10.1% for 2019 and 10.0% for 2010 to 2019; 11.4% on the four quarters to the second quarter of 2026. Profits at the 2019 share would be about 15% lower than at 11.9% (10.1 divided by 11.9; author’s calculation).

[36] Stray Narratives, Issue 26, “The 40-Year Tide is Turning (And the First Move Might Fool Most),” 7 August 2026 (the deflationary scare as the decoy; what ends a long bull; sequence over date).

[37] Cloud revenue growth year over year, December 2025 and June 2026 quarters: Microsoft Azure +39% (earnings release, 28 January 2026) and +43% (29 July 2026); Amazon AWS +24% (earnings release, 5 February 2026) and +37% (Form 10-Q, June 2026); Google Cloud +48% (earnings release, 4 February 2026) and +82% (Form 10-Q, June 2026; includes sales of Google’s AI chip systems from the second quarter of 2026). Total cloud revenue, not AI-only.

[38] Paul Kedrosky, The Meb Faber Show, episode 648, 28 August 2026: bubbles “tend to have a bunch of things in common”, “maybe loose credit”, a great technology story, “sometimes a real estate component” and “sometimes a policy angle”; “this moment is the first one that sits at the intersection of all of the forces that created the largest bubbles in US history”. One Big Beautiful Bill Act (2025), section 70301: permanent 100% bonus depreciation for property acquired after 19 January 2025. The mapping to this cycle is the author’s.

[39] Board of Governors of the Federal Reserve System, Financial Accounts of the United States (Z.1), households and nonprofit organizations, directly and indirectly held corporate equities, market value (via FRED, series BOGZ1LM153064475Q, release of 11 September 2026): $74.0 trillion at the end of the second quarter of 2026 and $13.7 trillion at the end of the first quarter of 2000. Ratio to nominal GDP (BEA via FRED, GDP): 227% and 137% (author’s calculation). Indirect holdings include mutual funds, pension and insurance assets.

[40] US Bureau of Economic Analysis via FRED, vintage of 30 September 2026: PCE price index (PCEPI) and PCE excluding food and energy (PCEPILFE), year over year, never at or below 2.0% in any month from March 2021 to August 2026 (3.4% and 3.0% in August 2026). Federal deficit: FRED FYFSGDA188S, federal surplus or deficit as a percent of GDP, fiscal 2025, -5.8%; BEA federal receipts less expenditures over GDP (FGRECPT, FGEXPND, GDP), -5.2% in the second quarter of 2026 (author’s calculation).

Companies that sell chips do not usually guarantee their customers’ rent.

In August, Nvidia agreed to guarantee up to $105 billion on leases for data centres being built for an affiliate of OpenAI. [1] If OpenAI defaults and the space cannot be re-let or sold for enough, Nvidia pays the shortfall. In return, the site will run only Nvidia’s systems. The seller has agreed to guarantee part of its own customer’s rent, and it is not the only seller doing something like this.

Why would a seller do that? My answer is that the money to pay for all of this has not arrived yet. This year, six companies will spend about $785 billion building data centres and filling them with machines, according to press reports of Moody’s forecast. Next year, it will be close to $1 trillion. [2] For all that spending, customers currently pay at most $250 billion a year for AI. [3]

The whole issue, summarised in one word, is sequencing.

Basically, I think the bill for this build-out will arrive before the money that pays for it, and that the realisation of this ends in a bust that builds in stages.

The best guide to why is electricity. It changed everything, and it still took decades to pay off. In 1899, electric motors supplied less than 5% of the mechanical power in American factories. Only in the early 1920s, four decades after the first power station opened, was slightly more than half of it electric. Factories first bolted motors onto their old steam-era layouts. The gains came only when they rebuilt the factory floor around the new power. [4]

Nor did the money stay with the companies selling the power. They ended up as regulated utilities, and from the late 1910s their rates fell sharply against other prices. [4] Electricity made companies more efficient, not more profitable: once everyone had it, nobody had an edge.

AI looks set to repeat both lessons, on price and on speed.

Start with price: Most sellers lose customers when they charge more. A few do not. A luxury handbag sells scarcity. Cigarettes sell a habit. And the spreadsheet your whole office runs on sells the pain of leaving it.

AI has none of that, at least not yet. Everyone is building the same thing at the same time, so it will not stay scarce. Most consumer AI is given away, so paying for it is not yet a habit. And models only a few months behind the best are free to download, so leaving is easy. So the price keeps falling, far faster than electricity’s ever did. Household electricity got about 5% cheaper a year for eight decades. The same level of AI performance has become about 13 times cheaper every year since 2023. [5]

Then there is speed: The buyers with the most money and the best reason to pay are employers, because a machine that does a job replaces a wage. But to get the gains, companies have to rebuild how they work, as factories did around the electric motor, and that takes years. Even OpenAI’s chief executive, Sam Altman, says so. Asked on a podcast published in August whether 2026 would be the year every business is “up for grabs”, he said he disagreed on the timeline. “The economy just has so much inertia,” he said, and “we’ve all been too ambitious on timelines.” [6]

That matters to you even if you have never used a chatbot. If you own an S&P 500 index fund, about $30 of every $100 in it sits in seven companies that are building AI, or selling to the companies that are. They are Nvidia, Microsoft, Amazon, Alphabet, Meta, Broadcom and Micron. [7]

This is part one of two. Here I do the arithmetic: how big the bill is, and why the money cannot arrive in time. Part two, “The Companies Selling AI Are Paying for It. For Now.”, follows the money: who is paying for all of this today, and the order in which I expect it to break.

Subscribe now

Nobody Has Sent the Bill Yet

If this much capital has been committed, why does everything still look fine?

Because the real test comes later. It did in the late 1990s too, when telecom companies laid fibre-optic cable across America to carry the internet. One of the boom’s mantras was that internet use doubled every three to four months. Between 1996 and 1999 the amount of fibre laid each year more than quadrupled. [8]

The demand was real. Internet traffic kept growing, roughly doubling every year. [8] But it grew far more slowly than the forecasts, and two other things went wrong. Everyone built the same thing at the same time, and the technology kept improving. A strand of fibre that carried 2.5 gigabits a second in 1996 could carry 100 by 2000, forty times as much from cable already in the ground. [8] Capacity became plentiful, and of course, its price fell. [8]

The Nasdaq’s telecom index peaked in March 2000 and had fallen 62% by the end of that year. [8] Global Crossing, which owned a fibre network reaching 27 countries, filed for bankruptcy in January 2002. [9] The network kept running. A new owner put in $250 million and took control, creditors got part of the new company, and under the bankruptcy plan the shareholders got nothing. [9]

AI has some of the same ingredients. The demand is real and everyone is building the same capacity at the same time. And as the technology improves and providers compete, the price of a given level of performance keeps falling. [5] So far that has not hurt: cheaper AI is being used more, and revenue is still growing fast. But when a price falls that fast, usage has to grow almost as fast just to keep revenue where it is. The fibre lesson is what can come next. Once capacity outgrows use, prices can fall faster than use grows. The risk is not that customers stop coming. It is that nobody keeps enough pricing power to charge what the machines cost. Computing has been here before. Between 1994 and 1998, IBM shipped 30% to 60% more mainframe computing power every year, and its mainframe revenue fell in four of those five years, because the price of each unit fell faster. And this time the equipment has a short life. The companies' own accounts write servers off in five to six years, [10] so they have to earn their keep now, not in some later decade.

CoreWeave, which sells AI computing, makes almost all its revenue on what the industry calls “take-or-pay” contracts. [11] The customer books a slice of computing capacity for years and owes the money whether it uses it or not. Other companies’ contracts are private, so I cannot see how strict they are. I assume most work the same way, and that they are strict. Only a small part of the money is paid upfront: across CoreWeave’s contracts, the deposit has averaged 15% to 25% of the contract’s value. [11] The big payments start only when the capacity is switched on, and a data-centre campus takes years to build, so most of the bill is still in the future.

Nvidia’s guarantee only takes effect when the first leases are ready for service, expected from 2028. [1] Oracle has $664 billion of contracted work and expects to turn only 13% of it into revenue in the next twelve months. [12] A machine’s cost only reaches reported profit once it is switched on. A large order book in 2026 says little about whether the money arrives in 2028.

Part of the reason everything looks fine is where the money comes from. Most of it is still the big cloud companies’ own cash, earned from search, advertising, software and online shopping. Some is borrowed or raised from outside investors. Some is backed by the sellers themselves, like Nvidia’s guarantee. But the balance is shifting fast. In the latest quarter the six companies spent about $200 billion, as much as their businesses brought in. [13] And much of the building never shows up in their spending at all. Microsoft, Meta, Amazon and Alphabet have signed about $830 billion of leases, mostly for data centres, that have not started yet. [13] Landlords build those sites with their own borrowed or raised money. That is the big difference from the fibre boom, much of which was built by new companies such as Qwest, Level 3 and IXC. [8] The closer precedent is the mainframe. IBM paid for its machines itself and rented them out, and its grip on big business computers lasted more than 25 years. When it broke, IBM's shares fell 77% in six years. I think this boom will also be slower to break, but not safer for shareholders. As long as that money keeps flowing, the orders keep coming, and the order books look like demand. Part two comes back to this.

The Trillion-Dollar Treadmill

What does a trillion dollars a year of machines have to earn? The answer takes three steps.

That $250 billion is the most generous estimate of what the AI industry earns from AI products and services. It comes from Aswath Damodaran, a professor of valuation at New York University. [3] It adds up the annualised revenue of the AI labs, OpenAI and Anthropic among them (recent revenue scaled up to a full year), plus what xAI, Microsoft and others earn from AI products. It leaves out the companies that sell the chips and equipment, because the question is what customers pay to use AI.

On the spending side, Moody’s expects close to $1 trillion a year by 2027. [2] Not all of it is AI. Some pays for ordinary cloud services and other buildings. My estimate is that about three quarters is AI, roughly $750 billion, but no company discloses the split. So I run the long-run bill twice. I show the working on the full $1 trillion, because the numbers are rounder, and give the three-quarters answer next to it. The year-by-year sums later on use three quarters, the kinder case. That is why the bill comes out as a range, $3.2 trillion to $4.3 trillion: the low end is the three-quarters case.

Step 1 - How big the stock of machines gets: The stock is all the servers, network equipment and buildings the companies own. Their own accounts treat servers and network equipment as used up over five to six years. [10] Buildings last longer. Epoch AI, which models what a data centre costs, assumes five years for the equipment, fourteen for the building, and about 70% of the cost in equipment. [14] On those lives, about 13% of the original cost of the whole stock is used up each year. That is an average life of about 7.7 years (100 divided by 13). Longer building lives would lower that rate. The stock would then grow bigger before it stops, and so would the bill.

Treat every machine as lasting that average. Spend $1 trillion a year. After one year you own $1 trillion of machines. After two, $2 trillion, and so on. The early machines start to wear out. After about eight years, $1 trillion of machines wears out every year, the same as the $1 trillion bought each year. From then on the stock stops growing. It has reached $7.7 trillion at what the machines cost new, and 13% of $7.7 trillion is $1 trillion.

You now own machines of every age, from brand new to nearly worn out. On average each is halfway through its life. Half of what was paid is already used up, and the replacement payments in step 2 cover it. The other half, $3.85 trillion, is still tied up in the machines, and that is the money that has to earn a return.

Step 2 - What that stock has to earn every year: The machines have to make two payments. The first is replacement: $1 trillion a year, every year, to replace what wears out. The second is a return for the people who put up the money. Over the twelve months to June 2026, Microsoft, Amazon, Alphabet and Meta earned about 23 cents a year after tax on each dollar of capital they put to work. That is about 28 cents before tax. [15] Capital here means the money that shareholders and lenders have tied up in the business. I assume the new machines must earn the same 28 cents, rounded up to 30 cents for simplicity. A 30% return on the $3.85 trillion still tied up in the machines is about $1.15 trillion.

Together the machines must produce $2.15 trillion of cash every year: $1.00 trillion to replace what wears out, plus $1.15 trillion of return.

Step 3 - Turning that cash into revenue: What customers pay is revenue. Staff, electricity, sales teams and overhead take their share. What is left is the cash the machines produce. I assume half of each revenue dollar is left as cash. If half is left, the machines need $4.3 trillion of revenue a year to produce the $2.15 trillion. If only three quarters of the spending is AI, the same sum gives $3.2 trillion. If instead only 30 cents of each revenue dollar is left as cash, and all of the spending is AI, they need $7.2 trillion.

Across those three cases the answer runs from about $3 trillion to $7 trillion. It never gets small. The 30 cents is what these companies already earn. Even if lenders and shareholders accepted 10 to 12 cents, a normal cost of capital before tax, the bill would still be $2.1 to $2.9 trillion, eight to twelve times what AI earns today. [15] Even with no return for investors at all, the machines still have to replace themselves. That is $1 trillion of cash a year. If half of each revenue dollar is left as cash, it takes $2 trillion of revenue. If only three quarters of the spending is AI, it takes $1.5 trillion.

Subscribe now

The Size of It

At $3.2 trillion to $4.3 trillion a year, the bill is about 2.7 to 3.6 times everything the US federal government spends on defence. [16] That compares a world total with one country’s spending.

America is about a quarter of world output. [17] If it carries a quarter of the bill, in line with its share of world output, America’s share is $0.8 trillion to $1.1 trillion a year. That is about as much as everything American businesses invest in software in a year, and up to a third more. [16]

That is every year, not once. I know of no industry that has grown from $250 billion a year to trillions in a few years. Money on this scale has to come out of some other giant pool of spending.

A Race Between the Bill and the Revenue

The $3.2 trillion to $4.3 trillion is where the bill ends up, in the 2030s. The race is about how fast it climbs on the way there. Every year the companies keep building, the revenue AI needs goes up, and revenue has to chase it.

Start with what is already bought. By my count from company spending to the end of June, taking three quarters of it as AI, the industry has built or is building roughly 24 gigawatts of AI computing capacity since 2023. [18] A gigawatt is the output of a large nuclear reactor, and the industry counts capacity in gigawatts because power is the physical ceiling on growth. Epoch AI’s own estimate is about 30 gigawatts, but it measures total facility power including cooling, so it is not directly comparable. [19]

Epoch AI puts the cost of owning and running one gigawatt at about $8.5 billion a year. That spreads the up-front cost of servers, network and buildings over their lives, with a charge for the money tied up in them, and adds about $0.9 billion of operating costs, electricity included. [14] The charge for the money is the return the computer services industry requires, much lower than the 30 cents in step 2. So the machines already built or being built cost roughly $204 billion a year to own and run.

Of each dollar customers spend on AI, the owners of the computers can expect to keep roughly 30 to 40 cents, as payment for running the models and training the next ones. Software companies spend 20 to 33 cents of each revenue dollar on all their costs of delivering the service, hosting and support staff included. [20] So my 30 to 40 cents does not overstate what the machines need. The labs hand over more today: Anthropic projected in May that its computing cost would fall from 71 cents of each revenue dollar to 56 cents. [21] I do not think a business that lasts can keep handing over that much.

So the machines already paid for need $510 billion to $680 billion of revenue a year. Today the industry brings in at most $250 billion, so the owners of the computers receive $75 billion to $100 billion, against $204 billion of cost.

Machines do not stop working when the accounts say they are used up, and older chips can keep earning on lighter jobs. If servers last seven years instead of five, Epoch’s figures [14] cut the yearly cost by about 18%. The revenue needed for the machines already paid for then falls to roughly $420 billion to $560 billion. That is still 1.7 to 2.2 times what the industry earns today. It shrinks the gap but does not close it.

And the building has not stopped!

On Moody’s forecast, the six companies still have about $440 billion to spend in the second half of this year, and close to $1 trillion next year. [2] [18] Each $1 trillion buys nearly 20 gigawatts and needs another $420 billion to $560 billion of revenue a year. Some of it is already sold on the long take-or-pay contracts I described above, so the cost is locked in. The revenue that has to pay for it is not.

This is a Red Queen’s race: you have to run as fast as you can just to stay in the same place. The bill climbs like a staircase. If the building goes on at next year’s pace, the revenue AI needs reaches about $0.7 trillion to $0.9 trillion by the end of this year. It reaches $1.1 trillion to $1.5 trillion by the end of 2027, and $2.0 trillion to $2.6 trillion by the end of 2029.

Now set revenue against it. An independent tally, global excluding China, puts annualised AI revenue at $175 billion in June, after four quarters of growth of 35% or more. [22] Add one more quarter at that pace and it is about $236 billion today.

Even if revenue keeps growing 35% a quarter, it is still short of the bill at the end of 2027. It only overtakes in 2028, after growing six to eight times. If it doubles every year, which is faster than any of the three big cloud businesses is growing today, it reaches the low end of the bill only in the second half of 2029, and still not its top end. If it grows as fast as Google Cloud, the fastest of the three at 82% a year, it is still short in 2029. [23]

And growth would have to stay extraordinary after that. The full bill arrives in the 2030s. Revenue of $4 trillion would be about fourteen times all US digital advertising, which was $294.6 billion in 2025, [24] though that sets a world figure against one country’s market. It would also be more than 3% of everything the world produced in 2025. [17]

Each year of building adds $420 billion to $560 billion to the revenue needed. In the last year, AI revenue added about $170 billion. [22] So the new money customers add each year has to roughly triple, and then stay that high for as long as the building goes on. I think that is very unlikely.

Three Wallets, None Can Pay in Time

Revenue is price times volume. The fibre story is why I think the price will not hold. The wallets show who could pay, and why the money cannot arrive in time.

Where would $0.8 trillion to $1.1 trillion a year of American money come from, and fast enough? To keep pace, worldwide AI revenue has to add roughly $0.4 trillion to $0.6 trillion every year while the build-out continues. Three wallets are big enough to discuss: households, software investment and payroll. I also size two smaller pools, advertising and governments.

Households. American consumers spend about $22 trillion a year. [16] If households alone paid America’s share of the bill, $0.8 trillion to $1.1 trillion a year, they would have to spend four to five cents of every dollar they spend on AI. The problem is pricing power. Household money moves fast: you can add a $20 subscription on a Tuesday. Paid plans exist at about that price. Meta’s Premium plan costs $19.99 a month. [25] But even 100 million people paying $20 a month is $24 billion a year, small next to a bill in the trillions. And most consumer AI is given away. Google’s Gemini has a free plan and 950 million monthly users. [26] Meta puts its assistant in WhatsApp for free. The paid plans it announced on 15 September add extras such as more media generation. Meta says the core experience stays free. [25] Apple includes its own at no extra cost, but it limits daily use of its server-based features and has said expanded access will be available for a fee. [27] DeepSeek and Alibaba release leading models that anyone can download for nothing. [28] A provider that charges a premium risks losing the household to a free alternative.

Advertising and governments. Two other pools could pay. Both are smaller than they sound. American digital advertising was $294.6 billion in 2025, and search alone is 38.8% of it. [24] Even if every dollar moved into AI, it would cover only a quarter to a little over a third of America’s share of the bill. And the whole market grew by about $36 billion in 2025. If AI is paying for itself through better advertising, it is doing so in tens of billions, not trillions. Published federal IT budgets, civilian and defence, add up to about $134 billion for fiscal 2026, and AI is only part of that. The Pentagon’s figure is its request. [29] I have not sized state and local government. Part two comes back to both.

Software investment. American businesses invest about $810 billion a year in software, as the national accounts count it: software bought, including software rented by subscription, and software built in-house. It leaves out what businesses pay for cloud computing capacity. [16] America’s share of the bill is all of it, and up to a third more. Software contracts reset when they renew, so this money moves fast. It is too small on its own. I have not sized what businesses pay for cloud computing capacity, or for IT services and outsourcing beyond custom software.

Payroll. American employers pay $16.3 trillion a year in wages and benefits. [16] If employers alone paid America’s share of the bill, it would take five to seven cents of every dollar they spend on wages and benefits. This is the one wallet where a buyer has both the money and a reason to pay: a machine that does a job replaces a wage. It is also where AI could finally build the lock-in that gives a seller pricing power. Once a company rebuilds its work around a model, leaving could become hard.

But payroll moves slowest. Taking a share of it is not a software purchase. It is a reorganisation: rewriting job descriptions, rebuilding workflows, retraining staff and accepting legal responsibility for what the machine gets wrong. A tool that is right 95 times out of 100 is a productivity aid, not an employee, because the company answers for the other five. Regulators, insurers, auditors and unions set the pace, and they work in years. Altman called it inertia. It is not in his models. It is in his customers.

And then there is the elephant in the room. Five to seven cents of every payroll dollar is roughly the pay of 8 million to 11 million American jobs, about one job in every fifteen to twenty. [30] If AI really takes that share of payroll, it does so by replacing work that people are paid to do today, through layoffs or through jobs that are never filled. I do not think any government will watch that happen quietly. The better AI gets at the one thing that would pay its bill, the more it invites the politics that slow it down: rules, taxes and limits on what companies may automate.

The money that can move fast is too small, or won’t pay. The money that is large enough cannot move fast enough, and moving it has dire social consequences.

Subscribe now

Why I Think This Ends Badly

Even at 35% a quarter, revenue does not catch the bill until 2028. After that, revenue has to add $420 billion to $560 billion a year for as long as the building goes on. Of the wallets I could size, only payroll can feed growth like that, and payroll moves slowly. So growth slows. It has a second reason to slow. Much of today’s growth comes from programmers, as the investor Paul Kedrosky points out. [31] Code is unusual: a short request explodes into thousands of lines that get rewritten again and again. Most office work goes the other way, turning a long inbox into a short summary. And much of the coding boom is a one-off clean-up of old code. Growth led by programmers is a poor guide to the rest of the economy. But the builders do not slow down when growth does. Each of these companies believes the next platform decides who survives, and nobody concedes a platform to protect a margin.

The bill wins the race if growth slows before the machines are paid for. I think it will. So the most likely outcome is that the technology works and the machines get used, but the money that built them does not come back in time to pay the bills.

The contracts run longer than the servers do. OpenAI expanded its commitment to Amazon by $100 billion over eight years, and Anthropic by more than $100 billion over ten. [32] Amazon treats its servers as used up in five to six years. [10] By my reading, not the filing’s, the servers that serve the early years are likely to need replacing before those contracts end. Even a take-or-pay contract protects the owner only while the customer can pay. If the customers’ revenue is late, the owner faces the cost of replacing the servers, and the shortfall lands on whoever lent or invested against that revenue.

And it does not stay inside AI. The seven companies named at the top of this piece are about 30% of the S&P 500. If the money arrives late, their profits take the hit, and every fund that tracks the S&P 500 takes it with them. [7]

I expect it to end in stages rather than in one crash. Sequencing, remember? Part two sets out the stages.

What Would Change My Mind

I am wrong if customers start spending the money, and paying a price that holds. What customers spend is the industry’s revenue, so I would see it at three levels: who is spending, the total, and the price.

The first test is payroll. If American employers are spending five cents of every payroll dollar on AI within two years, about $0.8 trillion a year, payroll is moving fast enough. That is the pay of about 8 million jobs, close to the 8.7 million jobs America lost between the end of 2007 and early 2010, in the slump around the financial crisis, moving to machines within two years. [30] It would be a revolution, in the workplace and in politics, and I think the politics would push back hard. Households, software buyers, advertisers and governments could also spend more than I expect. I would need to see it in disclosed figures, not forecasts. The best place to see it is the AI revenue that the large cloud and business-software companies report from business customers in their quarterly results.

The second test is the total. If AI revenue keeps growing 35% a quarter for the next seven quarters, it passes $1.9 trillion. That means growing about eight times in under two years, from a base that is already about $236 billion. It would take AI from smaller than all US digital advertising to more than six times its size, though that again sets a world figure against one country’s market. [24] I know of no industry of this size that has ever grown that fast. But if it happens, revenue is above what the machines already paid for and the 2027 build together need, even counting the spending since June. Then I am wrong about the race. I would check three sources. Microsoft, Alphabet and Amazon each report cloud revenue every quarter. The large cloud and business-software companies report AI revenue from business customers. Anthropic and OpenAI announce annualised revenue figures.

The third test is price. If AI providers can raise the price of a given level of performance without losing customers, or that price stops falling, the pricing-power problem is smaller than I think. The check is the prices the labs publish for their models, and independent trackers such as Epoch AI. [5] The hardest thing for this test to survive is China. Its labs give away models only a few months behind the best. [5] [28] Its companies have done this before: they make more than 80% of the world’s solar panels, at the lowest cost anywhere. [33] I expect them to do the same to AI and turn it into a commodity sold close to cost. If that happens, there is no pricing power left to find.

Where This Goes Next

That is part one. It leaves one question. If the sums are this hard, who is paying for all of this today?

Not mainly the customers. This year the six builders Moody’s tracks are spending about $785 billion on machines and buildings, on Moody’s forecast. [2] Customers pay at most $250 billion a year for AI. [3] The rest is coming from somewhere else, and some of it is coming from the sellers themselves.

Part two follows the money. It starts with the guarantee at the top of this piece. It shows who else is lending to the buyers. It sets out the best case for AI and why I think it fails. It gives the order in which I expect it to break, and what that does to the stock market.

Figures are from company filings, official statistics and the sources listed below. References marked press-reported or estimated have not been checked against a primary document. The $250 billion revenue ceiling is Aswath Damodaran’s. The rest of the arithmetic is my own.

Figures are from company filings, official statistics and the sources listed below. References marked press-reported or estimated have not been checked against a primary document. The $250 billion revenue ceiling is Aswath Damodaran’s. The rest of the arithmetic is my own.

References

[1] NVIDIA Corporation, Form 8-K filed 17 August 2026 (Item 1.01, residual value guaranties; accession 0001045810-26-000069); Form 10-Q, quarter ended 26 July 2026 (Note 10; accession 0001045810-26-000075)

[2] Moody’s Ratings, hyperscaler capital-expenditure forecast, press-reported (Moody’s report not independently obtained): CNBC, “Moody’s says ‘unprecedented’ AI spending threatens credit quality of Amazon, Meta, Alphabet and others,” 24 July 2026; DataCenterDynamics, “Moody’s: Hyperscaler capex forecasts marked up by $85bn, to close in on $1trn by 2027,” 24 July 2026. Capital expenditure across Microsoft, Amazon, Meta, Alphabet, Oracle and CoreWeave forecast at $785 billion in 2026 and close to $1 trillion in 2027 (Moody’s treats AWS capex as an estimate)

[3] Aswath Damodaran, “AI’s Bar Mitzvah Moment: From Hype & Hope to Business Questions!”, Musings on Markets, first posted 20 August 2026, updated 21 August (current AI revenue “caps out at about $250 billion” on the most upbeat estimates, built from the labs’ annualised run rates plus what xAI, Microsoft and others earn)

[4] Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” American Economic Review 80(2), Papers and Proceedings, May 1990, pp. 355 to 361: in 1899 electric motors were less than 5% of factory mechanical drive and electric lighting reached 3% of US homes; slightly more than half of factory drive was electrified in the early 1920s, four decades after the first central power station; regulated regional utility rates were lowered substantially relative to the general price level after 1914 to 1917; early retrofits added motors to belt-and-shaft layouts, and the productivity gains came with redesigned factories

[5] Epoch AI, Luke Emberson and David Roodman, “The plunging price of thought,” epoch.ai/publications/the-plunging-price-of-thought, 22 September 2026 (updated 23 September): the cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13 times a year; 7 to 10 times a year on game puzzles, 16 to 19 times on maths; fastest when a level is first reached (75 times a year), slowing to 4.7 times a year two years later; the most intense price competition near the state of the art has been between closed-weight models. Epoch AI, Luke Emberson, “Open-weight models lag state-of-the-art by around 3 months on average,” 30 October 2025. An earlier Epoch data insight (Cottier, Snodin, Owen and Adamczewski, 12 March 2025) found per-token price declines of 9 to 900 times a year. The same report: the average US residential price of electricity fell by a compounded 1.05 times a year between 1892 and 1973

[6] David Senra, “Sam Altman, OpenAI,” episode of the David Senra podcast, page dated 23 August 2026 (davidsenra.com/episode/sam-altman); the transcript is machine-generated and the audio was not reviewed

[7] State Street Global Advisors, SPDR S&P 500 ETF Trust daily holdings, as of 29 September 2026 (downloaded 1 October 2026): Nvidia 8.3%, Microsoft 5.7%, Amazon 3.7%, Alphabet 5.5% (two share classes), Meta 2.5%, Broadcom 2.6%, Micron 1.8%, together about 30%. The author’s selection of companies that are building AI or selling to those who do

[8] Elise A. Couper, John P. Hejkal and Alexander L. Wolman, “Boom and Bust in Telecommunications,” Federal Reserve Bank of Richmond Economic Quarterly, Fall 2003: the mantra that internet use doubles every three to four months; traffic roughly doubling every year since 1996 (Coffman and Odlyzko); annual fibre deployment more than four-fold, 1996 to 1999; fibre capacity of about 2.5 gigabits a second in 1996 and 100 by 2000; demand for long-haul capacity grew more slowly than forecast; ‘capacity prices are dropping’ (America’s Network, February 2001); Nasdaq telecom index peaked on 10 March 2000 and fell 62% by the end of 2000

[9] Federal Communications Commission, Order and Authorization DA 03-3121, In the Matter of Global Crossing Ltd. (Debtor-in-Possession), adopted 8 October 2003: network in 27 countries; Chapter 11 petitions filed on 28 January 2002; under the reorganisation, Singapore Technologies Telemedia invests $250 million in New GX, Global Crossing relinquishes all of its equity and voting interest in New GX, and its creditors receive New GX shares. Global Crossing Ltd., Form 10-K for fiscal 2002 (filed 8 December 2003), Item 1: holders of common and preferred shares “will not receive any distribution under the plan”

[10] Alphabet Inc., Form 10-K for fiscal 2025 (servers and network equipment generally six years); Amazon five to six years; Meta five to five and a half; Microsoft two to six; Oracle and CoreWeave six (each company’s Form 10-K, Note 1 or equivalent). The filings give lives for servers and network equipment as a class, not for AI accelerators specifically

[11] CoreWeave, Inc., Form 10-K for fiscal 2025: customers purchase a specified amount of capacity on a take-or-pay basis over the contract term, payment due regardless of the level of utilisation; contracts typically include a prepayment, and the weighted-average prepayment across all active contracts was 15% to 25% of total contract value at the end of each of 2023, 2024 and 2025; over 98% of 2025 revenue came from committed contracts

[12] Oracle Corporation, Form 10-Q, quarter ended 31 August 2026, filed 11 September 2026 (remaining performance obligations $664 billion; about 13% expected to be recognised as revenue in the next twelve months, 37% in months 13 to 36, 34% in months 37 to 60)

[13] Stray Narratives calculations on EODHD cash-flow data, spot-checked against Microsoft’s and Amazon’s filings: Microsoft, Amazon, Alphabet, Meta, Oracle and CoreWeave together, four quarters to June 2026: operating cash flow $714bn, capital expenditure $607bn, net borrowing $237bn, new shares $30bn; quarter to June 2026: capital expenditure about $200bn against operating cash flow about $196bn. Leases signed but not yet commenced, mostly for data centres: Microsoft $329.1bn (Form 10-K, fiscal 2026), Meta $279.0bn, Amazon $137.2bn and Alphabet $85.2bn of future payments (Forms 10-Q, second quarter of 2026), about $830bn in all

[14] Epoch AI, Amelia Michael and Ben Cottier, data insight on data-centre costs, epoch.ai/data-insights/ai-datacenter-cost-breakdown, 14 May 2026: about $38 billion up front and $0.9 billion a year to operate per gigawatt of IT capacity; about 69% of the up-front cost is servers and network equipment and about 30% the facility; lives of five years for IT equipment and fourteen for the facility; annualised to about $8.5 billion a year at the industry’s cost of capital (Epoch’s own sensitivity: $12 billion at a three-year server life, $7 billion at seven years). The author’s calculation on Epoch’s split and lives: the whole stock is used up at about 12.8% of its cost a year, an average life of about 7.8 years (rounded in the text to 13% and 7.7 years). With server lives of 5.5 to 6 years the same split gives 11.8% to 12.3%; with building lives of 25 to 40 years it gives 6.0% to 8.9%, which would raise the stock and the bill

[15] Stray Narratives calculations on EODHD data tied to the companies’ Forms 10-Q: Microsoft, Amazon, Alphabet and Meta, return on capital = after-tax operating profit ÷ (equity + long-term debt − cash and short-term investments); twelve months to June 2026 about 28.4 cents before tax and 22.7 cents after tax (effective tax rate about 20%). Amazon’s long-term debt of $128.9 billion is taken from its Form 10-Q because the data feed omits it. Cost of capital before tax: Aswath Damodaran, Cost of Capital by Sector (US), January 2026, after-tax cost of capital 7.8% for computer services and 9.3% for software (system and application), grossed up at the 20% effective tax rate to about 10% to 12%

[16] US Bureau of Economic Analysis via FRED, second quarter 2026, seasonally adjusted annual rates, vintage of 30 September 2026: federal national defense consumption expenditures and gross investment FDEFX, about $1.19 trillion; private fixed investment in software B985RC1Q027SBEA, about $810 billion (software bought, including hosted software rented by subscription, and software built in-house; fees for cloud computing capacity are outside it; BEA, NIPA Handbook chapter 6, and Rassier, Kornfeld and Strassner, “Treatment of Data in National Accounts,” May 2019); compensation of employees COE, about $16.3 trillion; personal consumption expenditures PCE, about $22 trillion (average of the three months)

[17] World Bank, GDP (current US$), NY.GDP.MKTP.CD, 2025: United States $30.77 trillion of a world total of $118.35 trillion, 26.0%

[18] Stray Narratives calculations on EODHD data: quarterly capital expenditure, Microsoft, Amazon, Alphabet, Meta, Oracle, CoreWeave and Nebius, Q1 2023 to Q2 2026 (Oracle’s fiscal quarters mapped to the nearest calendar quarter), about $1.18 trillion; three quarters of the total assumed AI infrastructure; $38 billion a gigawatt of IT capacity per Epoch AI; the result is about 23.4 gigawatts to the end of June 2026, shown as roughly 24 (spending since June adds more). Capital expenditure counts spending including construction in progress, so the capacity is built or being built. First half of 2026 at the six companies Moody’s covers: about $344 billion, leaving about $440 billion of Moody’s $785 billion 2026 forecast for the second half (author’s calculation from the companies’ filings)

[19] Epoch AI, “Global AI power capacity is now comparable to peak power usage of New York State,” epoch.ai/data-insights/ai-datacenter-power (read 1 October 2026): about 30 gigawatts of total facility power in the fourth quarter of 2025, calculated by multiplying chip power by 2.5 to include cooling and other overhead; not directly comparable with a count of IT capacity

[20] Cost of revenue as a share of revenue, latest fiscal year, from annual reports: total cost of revenue over total revenue: Datadog (2025) 20%; Salesforce (year to 31 January 2026) 22%; ServiceNow (2025) 22%; Snowflake (year to 31 January 2026) 33%; each from the company’s Form 10-K. These lines include hosting and support staff, not only cloud bills. Datadog and Snowflake run mainly on the large public clouds; ServiceNow and Salesforce also operate some of their own infrastructure

[21] Anthropic’s computing cost per revenue dollar, press-reported (Anthropic does not file accounts): internal projection shared with investors, Wall Street Journal, 20 May 2026, via PYMNTS, 21 May 2026, and PitchBook News, 10 June 2026 (71 cents in the first quarter, projected 56 cents in the second). A confidential draft registration reported on 29 September 2026 repeats the 71 cents; I have not seen a second-quarter ratio

[22] Exponential View (Azeem Azhar), “The State of the AI Economy,” 25 June 2026 (slide 7: $175bn annualised run rate, $110bn banked over the trailing twelve months, global excluding China, built from public disclosures and Exponential View’s own models; slide 11: quarter-on-quarter readings of 35%, 41%, 35% and 35%, “equivalent to 3.2x annually”). The four readings compound to 3.5 times, so annualised revenue was about $50 billion in June 2025. Taking the readings as listed, oldest first, it was about $68 billion in September 2025, against about $236 billion today (June’s $175 billion plus one more quarter at 35%): an increase of about $170 billion over the year (author’s calculation)

[23] Cloud revenue growth year over year, quarters ended June 2026 against June 2025: Microsoft, Azure and other cloud services +43% (earnings release); Amazon, AWS net sales +37% (Form 10-Q); Alphabet, Google Cloud revenue +82% (Form 10-Q). Total cloud revenue, not AI-only; Google Cloud’s growth includes sales of TPU systems recognised from the second quarter of 2026

[24] IAB and PwC, Internet Advertising Revenue Report: Full Year 2025, April 2026 (US internet advertising revenue $294.6 billion, up 13.9% on 2024; search 38.8% of the total)

[25] Meta, “Meet Your New Assistant: Meta AI, Built With Llama 3,” 18 April 2024 (free in its apps); Meta, “Introducing Meta One,” 15 September 2026 (about.fb.com/news/2026/09/introducing-meta-one-subscription-service-more-features-ai/): paid plans from $2.99 to $499 a month, Premium $19.99, and “the core experience across our apps and Meta AI has always been free, and that’s not changing”

[26] Google, Gemini subscription plans page (read 26 September 2026); Alphabet Q2 2026 earnings, CEO remarks, 22 July 2026 (Gemini app, 950 million monthly active users)

[27] Apple Newsroom, “Siri AI, a profoundly more capable and personal assistant, is here,” 14 September 2026 (daily usage limits on server-model features; “Expanded access to such features will be available for a fee in the future”); Apple Newsroom, “Apple Intelligence is available today on iPhone, iPad, and Mac,” 28 October 2024 (introduced as a free software update)

[28] Model cards and licences on Hugging Face (read 26 September 2026): DeepSeek (MIT License); Alibaba Qwen (Apache 2.0 for mid-size models; the flagship Qwen3.8-2.4T-A95B under the Qwen3.8-Max License, free of charge, with a separate licence for large hosted-service businesses); Moonshot AI Kimi K3 (Kimi K3 License, “free of charge”)

[29] Office of Management and Budget, Analytical Perspectives, Budget of the US Government, Fiscal Year 2027 (April 2026), chapter 5 (civilian federal IT spending $67.9 billion in fiscal 2026, Chart 5-2; excludes the Department of War); US Department of Defense, Information Technology and Cyberspace Activities, FY 2026 President’s Budget Request overview, August 2025 (request $66.1 billion). Together about $134 billion; intelligence-community IT outside the Department of Defense is not included

[30] US Bureau of Labor Statistics via FRED, total nonfarm payroll employment PAYEMS, September 2026: 159.0 million jobs. Average compensation per job about $103,000 (compensation of employees, about $16.35 trillion, divided by payroll jobs); $0.8 trillion to $1.1 trillion is the pay of about 7.8 million to 10.7 million jobs (author’s calculation). The same series fell from 138.4 million in December 2007 to 129.7 million in February 2010, a loss of about 8.7 million jobs

[31] Paul Kedrosky (SK Ventures), interview on Between Two COOs with Michael Koenig, “The AI Bubble Has a Due Date. It’s 2029,” 22 September 2026 (transcript machine-generated): early AI demand is dominated by coders, whose token use is expansive, while most office use is compressive; much coding work is a one-time refactoring of old code

[32] Amazon.com, Inc., Form 10-Q, quarter ended 30 June 2026 (Note 1): OpenAI’s commitment to AWS expanded by $100.0 billion over 8.0 years, Anthropic’s by more than $100.0 billion over 10.0 years

[33] International Energy Agency, Special Report on Solar PV Global Supply Chains, July 2022, executive summary: China’s share in all the manufacturing stages of solar panels (polysilicon, ingots, wafers, cells and modules) exceeds 80%; China is the most cost-competitive location for every part of the supply chain, with costs 10% lower than India, 20% lower than the United States and 35% lower than Europe

I bought Xunlei (NASDAQ: XNET) in May at $6.33, anchoring the thesis on an 80% discount to its net asset value. That value was built around a 7.8% stake in Arashi Vision, the maker of Insta360 cameras.

In July, I set a tripwire: a decisive break below 130 yuan on Insta360 would force me to reassess the trade. It broke that level and Xunlei shares are now trading around $5, down over 20% from entry. Insta360 has halved, falling roughly 6 times further than the STAR 50 index it trades against.

What remains is a bet on a single document, a specific date, and the price of management’s willingness to act.

The number that has not moved

Cash and short-term investments, less bank debt, come to $220.4 million, or $3.46 per ADS [4]. At the time of this writing, the Insta360 stake is worth $509.6 million, or $7.99 per ADS converted at that day’s exchange rate [3][4]. Then there is the operating business, which lost money in the second quarter. Take the $4.90 price, subtract the $3.46 of net cash and what is left is the $1.44 per ADS that the market is paying for the stake and the operating business together. Now let’s assume the operating business is worth nothing. The whole $1.44 becomes the price of the stake, against $7.99 of Insta360 shares, that is 18 cents on the dollar. When I opened the position, that number was 17 cents on the dollar.

There are two different reasons the market might only give partial credit for the stake:

Reason one: the market thinks Insta360’s Shanghai price is too high, and is quietly betting it’s worth less.

Reason two: the market doesn’t have an independent view on Insta360 at all. It’s just applying a fixed haircut to whatever the stake is quoted at, because the real question isn’t “is Insta360 cheap or dear” but “will Xunlei’s management ever actually sell it and hand you the cash.”

The fact that the stock has kept a fixed hair cut means its reason 2.

Subscribe now

A Deteriorating Collateral and a Broken Stub

Insta360 halved on more than just sentiment. First-half profits collapsed to a loss of 90.1 million yuan. Gross margins fell to 38.3% from 51.2% a year earlier. Operating cash flow plunged into negative territory as the company spent aggressively to stockpile chips. Inventory doubled. Short-term borrowing rose eight-fold [6].

This is not the same business trading at a lower multiple; this is a business changing shape under the pressure of a margin-crushing price war.

Xunlei’s own operating business, which I called a “cheap stub” in July, is no longer a bargain either. It posted a GAAP operating loss of $4.8 million in the second quarter [5]. If you price the Insta360 stake at its Shanghai value, the market is valuing Xunlei’s operating business at roughly minus $420 million.

The Illusion of the Clock

The case for a forced sale rested on the Investment Company Act. It is the US statute that can classify a company as an investment company once too much of its balance sheet sits in securities of other companies. Past that threshold, a company is no longer treated as an operating business. I wrote in July that this "still points in one direction" and that roughly $446 million of selling was required to reach management's stated 45% ceiling [2].

That statutory obligation is looking increasingly like a management option.

The 45% figure and the line “we never intend to be an investment company” are not in Xunlei’s most recent annual filing. They are earnings-call language, said twice this year, both times preceded by “if needed” [7]. The filing itself, made two weeks before I bought the stock, says something different. Even if the statutory test is met, the company believes it can rely on a separate exclusion. That exclusion is available to a business “primarily engaged” in an operating trade, its cloud and consumer services, rather than in holding investments [8]. On the August call, the company’s own words were “going through certain procedures to get ready to reduce our stake.” In the same answer: “we have no specific plans for the share disposal for the time being” [7].

Read together, the obligation is not a statutory clock. It is a management option. Its writer has already told the regulator it does not think it has to exercise it. Selling the stake down would still cost tax on the gain. I published that tax at 10% in July. It should have been a range then too. The rate is zero if the Hong Kong holding entity qualifies under the cross-border tax treaty that applies below a 25% ownership threshold — Xunlei’s stake in Insta360 is 7.84%. It is 10% by default. It is 25% if the entity is treated as mainland-resident [9].

I Am Holding. Here is What Has to be True.

I am holding the position. The downside is heavily buffered by $3.46 per ADS in net cash, and the stake itself remains highly liquid. However, the expected return now relies entirely on management breaking their own stated inertia.

If nothing is filed and Insta360 keeps de-rating toward 50 yuan, the value in two years is around $3.33 per ADS — a loss of about 17% a year. Net cash alone today is $3.46, so this is close to a floor case, not a tail case. If a plan is filed around the November results, and Xunlei sells down to its stated ceiling through 2027 at prices that allow Insta360 to keep falling, the value is around $6.25. That is a gain of about 13% a year. If the sale happens on the same timetable and Insta360’s fortunes turn instead, the value is around $9.65, a gain of about 40% a year, on legs that are not yet in motion [4].

Let’s be honest, would I put on this position today? No. But the asymmetry is like an option so I am keeping it a little longer. That’s not a great argument by the way and my patience is running out.

Three specific events will invalidate this stance:

• Upside Surprise: A reduction-plan pre-disclosure appears on the Shanghai exchange before Xunlei’s mid-November Q3 results. This is the only event that turns the discount into cash. If it prints, I was wrong about their timeline, and the market’s price was wrong.

• Structural Failure: Insta360’s Q3 report (due late October) shows total net margins remaining negative, or gross margins staying below 35%. This proves the camera business is losing a structural fight, making a much lower valuation the new normal.

• Market Capitulation: Xunlei’s shares close below net cash ($3.46). This would signal the market has begun pricing the operating business as a terminal liability.

References

[1] Stray Narratives Issue 11, An 80% Discount to NAV, Four Weeks to Close, published 14 May 2026. Entry $6.33; Insta360 stake, net cash, operating business as published there.

[2] Stray Narratives Issue 19, The Discount That Did Not Close, published 5 July 2026. Re-underwritten thesis, no clock, mechanical exit on an acquisition-shaped capital move, review tripwire at a decisive break below 130 yuan on Insta360.

[3] EODHD eod/XNET.US: $6.33 close 13 May 2026 (entry); $4.90 close 3 September 2026. Sina Shanghai daily bars, sh688775: ¥220.44 close 13 May 2026; ¥108.89 close 3 September 2026, the lowest close since Insta360 listed in June 2025; the STAR 50 index over the same window, −9.0%, from the same feed. EODHD carries no coverage of the Shanghai STAR Market.

[4] Xunlei Limited Form 6-K, Exhibit 99.1, unaudited results for the quarter ended 30 June 2026, filed 13 August 2026: cash $125.8m, short-term investments $151.0m, bank borrowings $56.5m, net cash $220.4m; 318,702,251 ordinary shares outstanding, 63,740,450 ADS at five ordinary shares per ADS. Insta360 stake: 31,437,542 shares (Arashi Vision Inc. 2026 half-year report, top-ten shareholder table, 30 June 2026) at ¥108.89, converted at EODHD’s 3 September USD/CNY close of 6.7181. Scenario arithmetic (bear/base/bull) per the same balance sheet, tax at 10% in the base case, a 25% discount applied to every non-cash component in all three scenarios.

[5] Xunlei Limited Form 6-K, Exhibit 99.1, quarter ended 30 June 2026: revenue $102.7m (+38.9% y/y); subscription revenue $44.5m (Q1 2026: $45.0m); gross margin 55.8% (Q2 2025: 63.0%); GAAP operating loss $4.8m.

[6] Arashi Vision Inc. (Insta360, Shanghai STAR Market: 688775), 2026 half-year report, disclosed on the Shanghai exchange’s disclosure site 27 August 2026 (unaudited): profit before tax −¥90.1m (H1 2025: +¥540.4m); second-quarter gross margin 38.3% (Q2 2025: 51.2%); operating cash flow −¥2,761m (H1 2025: +¥241m); inventory ¥6,212m (+112.8% in six months), 43.7% of total assets; short-term borrowings ¥1,061m (from ¥123m); bank facility ceiling raised from ¥6.5bn to ¥12.0bn (disclosure 27 August 2026); other non-current liabilities ¥1,795m (from ¥496m), an equity-repurchase obligation to minority investors in the company’s drone subsidiary, which raised ¥1,303m from those investors in the first half. Trailing earnings multiple 99x at the 3 September close (¥45.66bn market capitalisation), against roughly 57x at the mid-July check.

[7] Xunlei Limited, first-quarter 2026 earnings call (28 May 2026) and second-quarter 2026 earnings call (13 August 2026), management commentary (secondary transcripts; primary source is the call itself): “if needed, we will gradually seek to adjust our holdings so that Xunlei will hold investment securities with a value not exceeding 45% of the company’s total assets”; “we never intend to be an investment company”; “we are going through certain procedures to get ready to reduce our stake in the company. But at this time, we have no specific plans for the share disposal for the time being.”

[8] Xunlei Limited, Form 20-F for the fiscal year ended 31 December 2025, filed 28 April 2026: “We believe that, even if we were determined to be an investment company under Section 3(a)(1)(C), we are entitled to rely on the exclusion from investment company status afforded by Section 3(b)(1) of the Investment Company Act.” No “45%” self-declared target and no “never intend to be an investment company” language appears in this filing; both are earnings-call statements only.

[9] Cost basis for the Insta360 stake: $20.0m carrying amount at 31 December 2024 (Form 20-F, filed 28 April 2026), a book figure, not a confirmed tax basis. Tax treatment: Article 13(5) of the Mainland–Hong Kong tax arrangement taxes a share-sale gain in mainland China only where the seller held 25% or more of the underlying company’s capital in the prior twelve months; Xunlei’s holding entity holds 7.84% of Insta360. Withholding at 10% is the default rate for a non-resident enterprise absent treaty relief; 25% applies if the holding entity is deemed mainland-resident by place of effective management. Treaty-residency status of the holding entity is not established in the public record.

The Trade in 30 Seconds:

• The Flaw: America’s trillion-dollar deficit is funded by foreign portfolio capital. This is financial Airbnb, not a thirty-year mortgage.

• The Trap: When the stock market crashes, that foreign capital will dump US equities and sell dollars to go home. The dollar will drop right alongside the S&P 500. Good luck hiding there.

• The Fix: I am buying the Japanese yen against the Swiss franc. It is the rare crash hedge that actually pays you a yield while you wait for the sky to fall.

Quick trade update first: I am taking profit at the open on my SK Hynix position. It is up by over 15% in less than a month and as I mentioned in the article, it was a tactical trade, not a strategic position.

Now to this month article which is a very strong long term conviction of mine.

Months after the United States finally cut the dollar loose from gold in 1971, Treasury Secretary John Connally reportedly told his foreign counterparts that the dollar was America’s currency but their problem.

He was absurdly right for half a century but I think that sentence is starting to point in the opposite direction and the bill is finally arriving. I went long the Japanese yen and short the Swiss franc, added to the book at the open today.

This is the first in a series of hedges I will be layering into the portfolio over the coming months. As I have started describing in my new series on where I think we stand in this bull market, I don’t have the pretension of trying to forecast the market top but I think we can, with humility, better understand where we stand in this cycle. Valuations are stretched, market concentration is historic, and retail euphoria is palpable. When the music finally stops, you want your insurance policies signed. Too late to buy insurance when volatility is rising. This yen trade is policy number one.

The core claim here is narrow. It is a bet on market plumbing rather than American decline. I am not saying the dollar is finished.

For forty years, the market reflex was completely automatic: when panic hits, you buy dollars. I believe that reflex is now conditional, and that specific condition is currently failing.

The Problem With Rented Money

The entire argument sits in the accounting.

When a country imports more than it exports, it runs a current account deficit. Someone abroad has to fund that gap. That funding generally arrives in one of two forms. The first is direct investment and is a foreign company building a manufacturing plant, buying a domestic business, or putting something in the ground they intend to keep. The second is portfolio flow. This is foreigners buying stocks and bonds.

Direct investment is patient capital. It moves in and it stays. Portfolio flow is a tourist. It is money that can be sold on a Tuesday and wired back home on a Wednesday.

The US current account deficit is currently running near 3% of GDP on a four-quarter basis, which translates to just under a trillion dollars a year. The issue is not the sheer size of the deficit. It has been large before. The vulnerability lies in what is financing it.

Direct investment finances essentially none of it, and never has. The deficit is covered by the world buying American securities. According to the Federal Reserve’s own flow-of-funds data, foreigners now hold roughly $19.4 trillion of US equities and around $15.9 trillion of US debt. What has grown is not a willingness to build infrastructure in America, but the actual stock of America that the world is invested in. And who would blame us? There aren’t that many markets that have annualized over 20% per year for the past 10 years like the Nasdaq 100 has!

A Broken Reflex

Here is how the trap actually springs.

When a national deficit is funded by patient direct investment, the currency and the domestic stock market are only loosely linked. The money financing the country is fundamentally different from the money trading in the index.

But when that exact same deficit is funded by portfolio flows, they become the exact same money. The foreign capital covering the American external deficit is the foreign capital propping up the S&P 500.

In a serious US equity drawdown, that capital does not simply accept the markdown, hold hands, and sing Kumbaya. A significant portion of it goes home and to go home, they have to sell their dollars.

So the currency and the stock market fall together. The dollar turns pro-cyclical. It weakens in the exact risk-off scenario where it is historically supposed to strengthen, because the capital outflow crushing equities is the exact same capital outflow crushing the currency. It looks a lot like the late 1960s, right before Connally had his famous meeting.

If this reading is correct, the standard market crash hedge will fail right when you need it most. The risk-off bid will not magically vanish during a crash, it will migrate and flow to the countries that run surpluses and are owed money by the world.

Japan is standing at the very front of that queue.

Why the Yen Broke (And What Snaps It Back)

For three years, trading the yen was idiot-proof. You just tracked the gap between American and Japanese ten-year yields. A wider gap meant a weaker yen. It was entirely mechanical. Through 2022, the correlation was a flawless 0.88. If you knew the spread, you knew the currency.

That relationship is dead. Through 2025, it decayed to zero, and this year it violently inverted to negative 0.56. The single best historical indicator for the yen now points in the exact opposite direction.

Today, the yen sits near ¥156 to the dollar, hovering around a forty-year low. It has become the world’s favorite ATM: borrow yen at zero, buy absolutely anything with a pulse that yields, and pocket the difference. The entire financial world is crowded on one side of this boat. Speculative positioning is as aggressively short the yen as it has been in a decade.

The immediate assumption from the peanut gallery is that the yen is weak because Japan is a walking fiscal time bomb carrying the largest government debt in the developed world.

The accounting violently disagrees. Japan actually runs a massive current account surplus of roughly 4.9% of GDP. Its private sector saves more than enough to fund both private investment and the government’s deficit. On the IMF’s Fiscal Monitor, Japanese net public debt has actually fallen from a peak near 161% of GDP in 2020 to about 137%. Japan's external assets also exceed its liabilities by roughly 79% of GDP. A country the world owes money to is not the same thing as a country about to have a debt crisis. The yen is not weak because Japan is broke.

It is weak because the Bank of Japan is running negative real interest rates directly into rising wages and expanding credit. That is a textbook recipe for overheating an economy. Because the central bank is visibly behind the curve, rising Japanese inflation expectations are currently making the yen weaker, not stronger. A central bank that sleeps through its own inflation problem is not defending its currency; it is quietly devaluing it.

The politics have moved too. Through the summer the government tolerated a weak yen. It no longer does. The Takaichi administration now says it can live with higher rates and a firmer currency. The reason is domestic. Inflation imported through a weak yen is the kind Japanese households hate most. The Bank of Japan has less independence than the Fed or the ECB, and it does not tighten against a government that wants low rates. A government that wants a stronger yen removes the last excuse.

But every month the BOJ waits, the gap between where rates are and where inflation says they should be gets wider. This ends in a violent snapback. The yen is weak for reasons that can all rapidly reverse, and I only need one of these three roads to play out:

1. The American Crash. As described above, in a portfolio-led bust, the dollar falls and the safe-haven bid immediately moves to the surplus bloc. The yen is the deepest, most liquid expression of that exact trade. It pays out exactly when the rest of your portfolio is bleeding.

2. The Hawkish Capitulation. The inflation I was waiting for has started to print. Trimmed-mean inflation is back at 2%. Core inflation is running above 3% on a three-month annualised basis. Services producer prices, which lead services inflation, are up 3.6%. Nominal wages are up 3.6% and real wages 1.9%. Manufacturers' output-price expectations in the Tankan are the highest in five decades. A central bank at a 1% policy rate with core inflation above 3% is not tight. It is loose, and it knows it. At some point the BOJ stops being able to look away and is forced to hike directly into a massively crowded short position. That point has a date. The Bank of Japan meets on 18 September. A hike at that meeting is priced as near certain. The hike is not the question. The guidance is. If the Bank says it will keep raising rates roughly once a quarter until inflation settles, the yen stops being a funding currency. If the Bank calls September enough, the yen gives back ground and the curve steepens again. The carry pays me while the Bank makes up its mind. The guidance, not the 25 basis points, is what I am waiting to read.

3. The Great Repatriation. Japan’s public pension system holds roughly $1.2 trillion abroad, essentially unhedged. The weak yen handed these funds a massive translation gain, but that process runs in reverse the moment the currency turns. Japanese life insurers have already started dumping foreign bonds because hedging costs and new solvency regimes made them too expensive to hold. While this repatriation is traditionally a slow structural tailwind rather than a sudden trigger, it represents a massive wall of domestic money waiting to come home. There is a second flow behind the pension money, and it is annual rather than one-off. Japan earns income on its foreign assets worth about 6.3% of GDP a year. The Ministry of Finance publishes the primary-income line every month. For years that income stayed abroad, because bringing it home into a falling currency looked like throwing money away. A credible central bank changes that sum. Hedging works the same way. The 79% of GDP I mentioned above is mostly unhedged. When the yen turns, the rational response is to hedge, and hedging a foreign asset means buying yen. The income comes home first. The hedges follow. The pension money is the slowest and largest layer, and a leveraged carry unwind is what speeds all of it up when volatility rises.

It is also worth noting that on the OECD’s purchasing power parity estimate, fair value for the yen sits near ¥97 to the dollar. That metric does not tell you the exact timing, but looking at a rubber band stretched that far is incredibly persuasive.

Why the Swiss Franc?

The natural instinct here is to buy the yen against the dollar. That is a trap, and it is the exact mistake that has drowned every yen bull for the last decade.

Against the dollar, you pay the interest rate differential every single day you wait. You are taking a direct, bleeding bet on the exact timing of the American turn. Nobody ever gets that timing perfectly right.

But against the Swiss franc, the arithmetic changes in three beautiful ways.

I get paid to wait. The Bank of Japan is at 1% and climbing. The Swiss National Bank is at zero. That spread gives me positive carry of roughly one percent a year. On a macro trade where the biggest historical risk is the cost of holding it long enough to be proven right, flipping the carry from paying to collecting is everything.

I am shorting the reluctant haven. The franc is the only real rival to the yen as a panic currency, but the Swiss National Bank actively despises a strong franc. They lean against it heavily through direct market intervention. The asymmetry here is brilliant. Capping your own currency is something a central bank can do forever because it just requires printing your own money, and the SNB has plenty of ink. Holding a currency up requires spending foreign reserves, which are finite. I will gladly bet against the guys with the unlimited printing press who actively want to lose.

I keep the convexity. We just had a live rehearsal of this. At the end of July 2024, the BOJ raised rates and the yen spiked 14% against the dollar in five weeks, unwinding massive amounts of carry trade leverage and crushing global equities. Against the franc, the yen rose far less because the franc rallied alongside it as a fellow haven. So by shorting the franc, I capture roughly half the yen’s flight-to-safety surge, plus the daily positive carry, plus whatever rate hikes the BOJ delivers that the Swiss cannot match.

A partial hedge that pays you to hold it always beats a full hedge that bleeds you dry while your broker sends you polite margin call emails.

Intellectual Honesty: What Breaks This Trade

I have spent more time trying to destroy this thesis than build it. Some of the risks are not just theoretical; they are playing out on the tape right now. There is a reason there is something called the “Japan widow-maker trade”. For over two decades hedge funds and macro traders have repeatedly tried to make money betting against Japan. None of the strategies have ever worked, so who am I to try?

The Oil Shock. Japan imports its energy. A supply-driven crude spike creates a risk-off environment where the yen can actually fall because their terms of trade get crushed at the exact wrong moment. Crude spiked more than 50% in July on the escalation with Iran and the Hormuz disruptions, and the yen weakened straight into it, printing a fresh forty-year low. Brent has since settled back near $92 with the strait still contested, and the yen has recovered around three per cent off that low. This hedge is built for a demand-led deflationary bust and it has just been stress-tested by a supply-led inflationary one. It survived. That is not the same as being vindicated. This hedge is built for a demand-led deflationary bust, but it is currently staring down a supply-led inflationary shock. I am actively leaning into that headwind.

The Dirty Entry. Japanese authorities are suspected to have intervened heavily to support the yen on July 30, and again in the first days of September. So the level I am buying at is artificial. My entire argument for shorting the franc is that holding a currency down is infinite, while holding one up requires finite reserves. Japan is currently burning finite reserves to prop up the yen, and the arithmetic is not kind. Japan holds about $1.3 trillion of reserves. Each round of yen buying costs something between $50 billion and $75 billion. Those are a macro research house’s estimates; the Ministry of Finance publishes reserves monthly and intervention totals at each month-end, with daily detail each quarter. A finite reserve against a repeated bill is a bluff the market will eventually call. The yen also tends to weaken in the weeks after an intervention, as traders test how serious the authorities are, so the tape may move against me before the Bank meets. I am ignoring the official support because it only smooths the speed of a move, not its direction. I hold the position for the Bank, not for the Ministry.

The Fiscal Trap. The objection I hear most is that Japan cannot afford to raise rates. The arithmetic runs the other way. The government pays about ¥10.3 trillion of interest a year and receives about ¥9.2 trillion, so the net interest bill is near ¥1.1 trillion, roughly 4% of revenue. The Ministry of Finance budget documents carry the primaries. What matters is which part of the curve each side answers to. Payments follow Japan’s long yields, because that is where the debt is issued. Receipts mostly follow short yields abroad, because the biggest asset is $1.3 trillion of reserves sitting in short-dated dollar paper. So a steepening Japanese curve hurts the public finances on the only side Tokyo controls. A hike that flattens the curve helps them. The Japanese 30-year to 2-year spread flattened by about 44 basis points this summer on hawkish speeches alone, and the Ministry of Finance publishes the daily yield curve if you want to check. The cost of the debt is a long-end problem, and a credible central bank is what brings the long end down. That is why I do not treat the debt as a reason the BOJ stays on hold.

Finally, dollar bears have been early for so long that being early is functionally indistinguishable from being wrong. If the US deficit adjusts the normal way through a standard economic slowdown, the dollar will likely rally just as it has in most post-war downturns. I have been actually been rather bullish the dollar all year so much the positioning against it was one sided.

Subscribe now

The Playbook

I added long yen and short Swiss franc to the book at the open today, at 192.70. I am sizing it big, as big as my conviction level on this one. It is large enough to matter when the carry trade unwinds over the next few quarters/years, but small enough that I can comfortably sit through a longer wait.

• Base Case: The BOJ is forced to tighten through the winter. The cross grinds from roughly 193 toward the 175 to 180 area over the next year, and I collect a point of carry while it happens. We continue on that path every year ultimately settling around 150.

• Bull Case: The US equity market cracks. The global carry trade unwinds violently, the yen vastly outruns the franc as capital flees, and the cross slices through 150 in 2027.

• Bear Case: The BOJ hikes on 18 September and calls it enough. The yen gives the summer’s gains back and settles at the level the authorities appear to defend. Or the US slowdown closes the trade gap normally and the dollar rallies, or the current oil shock deepens dramatically. The cross drifts back toward 200 and the position bleeds slowly against the carry I am earning.

• The Call: Long yen, short Swiss franc. It is the rare crash hedge that actually pays you a yield while you wait for the storm.

Wrong If: The yen fails to appreciate against the dollar during the next sustained US equity drawdown. That correlation is the empirical bedrock of this entire trade. If the US market crashes and the dollar still rallies, the thesis is fundamentally broken. My stop loss is at the all time high a few months ago, 204.

Last year, Alphabet spent $45.7 billion buying its own shares. This year, its filings report a different number. Zero. Zero, in the first quarter, and zero again in the second [8].

Meta spent $26.3 billion last year. This year: zero, and zero [8]. Oracle bought $0.1 billion across its entire fiscal year. Amazon has bought nothing since 2022. Of the five giants building the AI era, only Microsoft still buys its own stock, at about $4.6 billion a quarter, and even that is shrinking [8].

The investors who know these businesses best, the companies themselves, have stopped buying their shares. And they have not merely stopped. Alphabet sold $18 billion of new stock in June and set up a programme to sell $40 billion more, quietly, over time [9]. The best-informed buyer in the market has crossed to the other side.

The largest buyer nobody talks about

Ask people who the biggest buyer of American shares is and they will guess pension funds, or index funds, or foreigners. The answer, for the past decade, was none of these. It was the companies themselves. Buybacks across the S&P 500 ran at a trillion dollars in the most recent twelve months on record, a fresh all-time high [12]. That is more than my entire estimate of the automatic retirement bid, and larger than any other source of demand I can size. For ten years, the largest buyer in the American stock market was the American stock market.

And the buying was heaviest exactly where the prices rose most. In 2021, the five giants that pay for the data centres, Microsoft, Amazon, Alphabet, Meta and Oracle, retired a net $144 billion of their own stock in one year [6].

Meanwhile the money coming in from outside changed character. Every month, retirement contributions flow into index funds that buy every stock automatically, in proportion to size, at whatever the price happens to be. I estimate that automatic bid at roughly $500 to 800 billion a year, built from government retirement-plan filings and industry fund-flow data [7]. It is set by payrolls and demographics, not by opinions. And because the index buys by size, roughly a third of every automatic dollar lands on the handful of giant stocks at the top [7].

Now put the machine together:

• A price-blind bid of hundreds of billions a year flowed in.

• The largest buyer in the market, the companies, was removing shares at the same time.

Demand that never asks the price, chasing a supply that was shrinking. Prices melted up, fastest at the top. Valuation never had to clear anything. Scarcity did the work.

You might object that these companies never looked expensive the way bubbles look expensive, because they are among the most profitable enterprises ever built. True, and that is exactly why the machine went unnoticed. The profits were real but the profits were also the fuel. Cash earnings funded the buybacks, the buybacks shrank the share count and a shrinking share count makes earnings per share grow faster than earnings. The price climbed, earnings per share climbed with it, and the ratio between them stayed respectable. A reasonable-looking valuation was not evidence against the machine. It was the machine’s product.

The switch

They did not stop because they wanted to but because the five spent $330 billion on data centres and equipment in the first half of this year, a pace near $660 billion a year, roughly 96 cents of every dollar their operations generate in cash [8]. Add the dividends they still pay and the outflow exceeds the cash coming in. Two of the five, Amazon and Oracle, already outspend their operations outright [8].

So the gap is being filled the way it always eventually is: with paper. About $90 billion of new borrowing sits in the half-year filings, and market data indicates $150 billion or more across the five [8][9]. And the equity, from the companies’ own documents [9]. Alphabet’s programme totals $84.75 billion: $18 billion of common stock and $16.75 billion of convertibles sold at once, the $10 billion Berkshire placement, and the quiet $40 billion to come. Together, by press tallies, the largest American follow-on on record. Meta, $25 billion of bonds out to forty-year maturities. Oracle, a $45 to 50 billion financing plan, with $25 billion of bonds and a $20 billion share-sale programme launched. Amazon, $67 billion of long-term debt in six months [8]. Alibaba, $10 billion of new shares, by press tallies Hong Kong’s largest-ever follow-on.

Notice who is missing. Nvidia, the company at the very centre of the boom, repurchased $39 billion of its own stock in the same six months, more than four times what the five builders managed together [8]. The five spend the capex whilst Nvidia collects it.

The whole switch, in one number. In 2021 the five removed about $95 billion a year of stock from the market, net of what they issue to employees. This year they are adding stock at a pace of roughly $140 billion a year, counting the buybacks that remain, the shares issued to staff, and the new shares sold [8]. From minus $95 billion to plus $140 billion: a swing of some $235 billion a year, concentrated at the very top of the index, where the automatic bid lands heaviest.

The valuations that looked sensible on a shrinking share count now need to look sensible on a growing one.

The other side of the trade

So who buys what the best-informed holders are selling?

If you have a retirement account, the answer includes you. Part of every paycheck buys these exact stocks automatically, in proportion to their size, at whatever the price is that day. That is not a criticism; the automatic bid is the sane default for most people. But see the trade clearly. For a decade, your monthly contribution competed with the companies themselves to buy a shrinking pool of shares, which is a wonderful position to be in. From this year, your contribution is absorbing a growing pool of shares, some of it sold to you directly by the companies, which is a different position entirely.

And the queue on the selling side is forming. SpaceX went public this spring and raised $75 billion at a $1.7 trillion valuation, the largest deal ever; it rose 19% on the first day [10]. OpenAI confidentially submitted a draft listing document in May, by its own announcement [11]. Anthropic is widely reported to have done the same in June, at a reported target valuation around $2 trillion. None of that is confirmed by the company; the reports are press [11]. American initial public offering proceeds stand at $146 billion with four months of the year remaining, more than triple all of last year. Follow-on offerings are up by two-thirds [10].

Set the two sides against each other. The automatic bid: roughly $500 to 800 billion a year, held there by payrolls [7]. The supply now meeting it: initial offerings near $146 billion already this year, follow-ons at $92 billion in the first half, the giants’ own float growing at a $140 billion yearly pace [8][10]. Gross new supply is running on the order of $400 to 500 billion a year, already closing on the lower bound of the bid. The trillion-dollar absorber that used to soak up supply has lost its largest engines exactly as issuance triples [12]. Both blades of the scissors are moving, in the same direction, at the same time:

Subscribe now

You have seen this movie

In 1919 a company was formed to own the future of communication, and for a decade it did. RCA controlled radio the way today’s giants control computing. The technology delivered everything promised: by the end of the twenties, radio was spreading through millions of American homes, remaking news, entertainment, politics, everything it touched. RCA became the most actively traded stock on the New York exchange. In September 1929 it peaked around $114. Three years later it traded near $2.50, down 98%, while the number of radios in American homes kept growing [4]. The technology never stopped winning. The shares never durably saw their 1929 price again for decades.

That is not a story about radio. It is the recurring shape of every transformative technology, documented across two centuries in Alasdair Nairn’s history Engines That Move Markets [1]. Britain built the world’s densest railway network while the shareholders of 1845 lost most of their money, with railway investment at the peak near 7% of GDP [2]. Electricity conquered the world while Edison lost control of his own company in the 1892 merger that created General Electric [3]. Seven hundred American car makers became a handful [5]. The engine wins. The financiers of its deployment mostly lose. And the mechanism is always the same: the deployment is paid for by selling paper into peak enthusiasm, until the paper outruns the enthusiasm.

For a decade this boom looked like the exception, because its giants bought paper back instead of selling it, but that is no longer the case.

The 2008 rhyme, and the difference

Readers with long memories will hear another echo, this one from 2008. The financing looks familiar. The location of the risk does not, and that second point matters more.

The familiar part first. In 2006, lenders financed the buyers of their own product, and the same house was counted at every link of a securitisation chain. Today the chip and cloud sellers finance their own customers, and the same promised dollar of computing shows up three times: in one company’s backlog, in another’s guarantee, in a third’s valuation. The obligations carried dates then too. Roughly a trillion dollars of adjustable mortgages were scheduled to reset across 2007 and 2008, at $30 to 40 billion a month at the peak, and everyone could see the schedule [13]. The computing contracts signed in 2025 and 2026 begin billing in 2027 and 2028; the one start date actually written in a filing is 2028 [14]. In both cases the bills were visible years in advance, and ignored while everything still looked healthy.

Now the difference. In 2008 the risk sat inside the banking system. The five big investment banks ran leverage as high as forty to one. Bear Stearns carried $11.8 billion of equity against $383.6 billion of liabilities, and borrowed as much as $70 billion overnight, inside a $13 trillion shadow-banking system [15]. A 3% move in the collateral could wipe out a firm’s equity. When it moved, the payment system seized.

Today’s borrowers are investment-grade industrial companies carrying $565 billion of cash and short-term investments between them [8]. Microsoft still holds the triple-A rating it has held since 2008. When these bets go wrong, the losses land on shareholders and bondholders, not on the plumbing that clears everyone’s payroll. The banks are middlemen this time, not warehouses. They originate data-centre loans and pass them onward, and their direct lending to the sector is measured in tens of billions, not trillions [16]. That is why I expect this cycle to end by deflating rather than seizing. The next piece in this series examines how.

One caveat. Moving the risk out of the banks does not move it out of the system. The fastest-growing lender to this boom is private credit, a pool of $1.5 to 2 trillion. The Financial Stability Board, citing private-sector research, expects it to fund roughly $800 billion of AI infrastructure by 2028, financed increasingly by retail vehicles and by insurers that private-equity firms own [18]. That changes how fast an accident spreads and how far it reaches but it does not change the arithmetic.

So: can the bid clear it?

So far, yes, and that cuts against the doom-mongers. Every raise has filled. SpaceX priced and rose. Alphabet’s offering was taken. Alibaba’s book was reported three times covered [10]. The bid has cleared everything thrown at it.

But read what the machine is now being asked to do. The automatic bid is roughly fixed; payrolls and demographics do not care about excitement. The supply meeting it has swung by hundreds of billions a year and is scheduled to grow: the share programmes drawing quietly, the bond calendars, the listing queue. The race is not about any single raise. It is one run rate against another: supply growing toward a bid that payrolls hold roughly fixed. Scarcity has already ended, and for the first time in a decade the marginal buyer of the biggest stocks in the world will have to be someone who may ask the price.

That is how this boom is the same as every boom before it. The railway, the dynamo, the radio and the car all reached this exact moment: the technology proven, the enthusiasm intact, and the paper beginning to flow. From roughly this point on, the engines kept winning and their financiers started losing. The buyer’s largest engines have left the market they made. The paper is flowing. The next few quarters will answer it: can the bid clear it.

Buyback, issuance, capex and cash figures are from company filings, cross-checked against regulatory data; where a figure rests on press reporting it is labelled as such. The bid estimate is built from retirement-plan and fund-flow filings and is stated as a range. Sources below:

References

1. Alasdair Nairn, “Engines That Move Markets: Technology Investing from Railroads to the Internet and Beyond,” 2nd edition, Harriman House, 2018. The historical figures below are carried on independent sources, not the book.

2. Railway Mania: A. Odlyzko’s academic work on 1840s railway investment (~7% of UK GDP at peak); Campbell and Turner, financial-history studies

3. Formation of General Electric, 1892: standard corporate history

4. RCA share price 1929–1932: contemporaneous NYSE records via standard market histories

5. S. Klepper, entry and exit studies of the US automobile industry

6. Megacap buyback series 2021–2025: company cash-flow statements via SEC filings and market data services

7. The bid estimate: US Department of Labor Form 5500 (retirement-plan contributions, ~$652bn 2024); Investment Company Institute retirement and fund-flow data; index weights per S&P 500 constituent data. Stated as a range; components not summed where they overlap

8. Q1 and Q2 2026 Forms 10-Q (Microsoft, Amazon, Alphabet, Meta), FY2026 Form 10-K (Oracle), and NVIDIA’s Form 10-Q for the half to 26 July 2026 ($39.0bn of repurchases): repurchase, issuance, stock-compensation, capex and operating-cash-flow lines via SEC XBRL as filed. Amazon and Meta debt issuance per the filings; the Alphabet and Oracle debt-change lines and the cash-and-short-term-investments aggregate are market data (EODHD), pending filing tags

9. Alphabet Forms 8-K and pricing prospectuses, June 2026 ($18bn common + $16.75bn mandatory convertibles + $10bn placement + $40bn at-the-market programme); Meta Form 424B2, May 2026 ($25bn notes); Oracle financing-plan release and prospectus supplement ($25bn bonds, $20bn ATM); Alibaba Form 6-K, August 2026 ($10.2bn placement); Nebius Forms 6-K ($5.75bn convertibles); CoreWeave releases (>$20bn of 2026 facilities and notes)

10. Renaissance Capital, US IPO market reviews, 2026 (SpaceX $75bn at $1.7tn, +19% day one; YTD proceeds $145.8bn) and H1 follow-on totals per ECM reviews — specialist and press-grade throughout; Alibaba book coverage per Bloomberg

11. OpenAI, announcement of confidential draft S-1 submission, 22 May 2026 (primary for the fact of filing). Anthropic listing preparations: press reports, unconfirmed by the company

12. S&P Dow Jones Indices, quarterly buyback release, December 2025: $249.0bn for Q3 2025; record $1.020tn for the twelve months to September 2025. No later release located at source as of writing

13. Financial Crisis Inquiry Commission, Final Report (2011): subprime originations $600bn in 2006 (23.5% of all originations); private-label securitisation ($1tn across 2007–08, $30–40bn/month peak): the contemporaneous Credit Suisse schedule, reprinted by the IMF; press-canonical

14. NVIDIA Corporation, Form 8-K, 17 August 2026: lease obligations “expected beginning in 2028.” The billing mechanics are treated in the next piece in this series

15. FCIC Final Report: broker-dealer leverage up to 40:1; Bear Stearns equity and liabilities; the $13tn shadow-banking system

16. Federal Reserve, Financial Stability Report, May 2026 (bank commitments to nonbank financials $2.6tn; no data-centre lending quantification); bank data-centre lending and syndication figures: press reports of industry data

17. S&P Global Ratings: Oracle downgraded to BBB− (from BBB, negative outlook), outlook stable, July 2026; Moody’s Baa2 negative via press; Meta prospectus ratings disclosure; Moody’s sector warning, July 2026, via press

18. Financial Stability Board, “Vulnerabilities in Private Credit,” May 2026 (global private credit $1.5–2tn; ~$800bn projected into AI infrastructure by 2028, citing private-sector research); Federal Reserve FSR Box 4.1 (US private credit ~$1.4tn; semi-liquid retail vehicles); NAIC (private-equity-owned insurers’ invested assets $704.3bn)

That trick is the whole of artificial intelligence. AI is not a brain. It is the marble jar, industrialised — a very, very good averaging machine, trained on roughly everything humanity has ever written down, handing you the middle of it on demand.

And the middle, it turns out, is excellent. Well-organised, grammatically immaculate, and almost always fine. That is exactly the problem. AI isn’t making us wrong. It is making us all the same shade of plausibly right — and it is doing it to everyone at once.

Subscribe now

But the machine plays a beautiful move

You are about to object, and you are right to. In 2016 a machine called AlphaGo played a move in a game of Go that no human would have played, a stone so strange the commentators assumed a glitch. It was not a glitch. It was beautiful, and it won. A few years later AlphaFold solved the shapes of nearly every protein known to science, a fifty-year problem, correct, at a scale no laboratory could touch. If your complaint about the averaging machine is that it only ever hands back the mushy middle, these are your counter-examples, and they are good ones. Drop the lazy version of the argument.

But look at what kind of new this is. Go is a closed world: fixed rules, fixed board, a fixed way to win, all of it handed to the machine before it starts. Inside that world there are more possible games than atoms in the universe, so a good enough search will always turn up a beautiful place no human has stood. AlphaGo found the move. It did not invent Go, and it could never tell you whether Go was worth playing. There is finding a new move, and there is drawing the board. The machine does the first better than we ever will. It has not once done the second.

That is the averaging, refined. The machine is not too timid to be brilliant inside a frame you have drawn for it. What it cannot do is draw the frame, choose which game is worth the candle, or feel that the whole board is wrong. The averaging that bites is not of the answer. It is of the judgment one layer up, the taste that picks the question, and that is the layer the world is now quietly outsourcing.

The jar only works if nobody can see the other guesses

Read the fine print on that marble jar. The magic depends entirely on the guesses being independent. The wild over-counters and the timid under-counters cancel out, and the truth shines through the noise. Let the guessers see each other’s answers first, and the whole thing collapses: everyone shuffles toward the loudest number in the room, and the “average” stops measuring the marbles and starts measuring the crowd’s confidence in itself.

There is even a formula for the magic, written down by the social scientist Scott Page: the crowd's error equals the average person's error minus the diversity of the guesses [9]. Read it twice, because the second term is the whole game. What makes the jar work is not that anyone is right — it is that everyone is differently wrong. And the formula carries a bomb in its pocket. Shrink the diversity, and the crowd gets stupider even while every guesser in it gets smarter. The market can get dumber while every investor in it gets sharper. Nothing else in this piece matters more than that sentence.

Cliff Asness of AQR made the obvious, devastating point: markets are not a sealed jar [2]. Investors can all see each other. They read the same headlines, hold the same positions, and repeat the same story until the story moves the price and the price seems to prove the story. The wisdom of crowds needs the crowd to think independently.

It is worth being precise here, because averaging is not the villain. Done right, it is close to magic. Philip Tetlock spent a decade proving it. His Good Judgment Project recruited teams of hobbyists — a retired programmer, a pharmacist, a ballroom-dancing instructor — and set them against professional intelligence analysts who had access to classified intercepts. The hobbyists won, comfortably [3]. Not because any one of them was a genius, but because the project pooled their genuinely independent, genuinely different views. A good jar, kept carefully sealed.

Now feed that crowd a tool trained on its own back catalogue. We all prompt similar models, trained on similar data, parsing the same feeds — a healthy slice of which were themselves written by AI. The answers converge. This is measured, not suspected. Researchers put the same open-ended prompts to 102 people and 22 different AI models: the models mirrored each other far more than humans mirror humans, even after controlling for the obvious confounders [10]. Another team ran 26,000 real user queries across dozens of models and won a best-paper award for documenting the same collapse — different companies' machines, strikingly similar answers [11]. And the material the machines learn from is already partly their own: by late 2024, up to a quarter of corporate press-release text was machine-written, a figure the study's authors call a lower bound [12]. Dozens of sources say the identical thing, each one apparently confirming the others, and we mistake the echo for a chorus. We have automated groupthink and given it a warm, confident, faintly human tone of voice.

Everyone’s a genius now, which is the giveaway

You may remember the Dunning-Kruger effect: the least competent people are the most confident, because they don’t know enough to know what they’re missing. It was a comforting law. It meant overconfidence was at least a signal — a red flag you could watch for.

AI broke the flag. A 2026 study in Computers in Human Behavior, flagged by Baillie Gifford’s Tom Slater, found that AI use produces uniform overconfidence across every skill level [4]. Novice and expert walk away equally certain they understand more than they do. The cruel twist: the people with the highest AI literacy were the worst calibrated of the lot. They had mistaken fluency with the tool for mastery of the subject.

If that sounds familiar, it should. Tetlock had already catalogued the species in the wild. Tracking tens of thousands of predictions by celebrated experts, he found the average one forecast about as well as “a dart-throwing chimpanzee” [5]. Worse: the more famous and quotable the expert — the ones with the book deals and the green-room makeup — the lessaccurate they tended to be. Confidence and television charisma were inversely related to being right. AI, in this light, is simply the dart-throwing chimpanzee finally handed a publishing contract and a soothing voice. Always available, always certain, reliably average.

We have, in other words, all become the chap at the dinner party who skimmed one article and now has Opinions — except the article is beautifully written, we have twelve of them, and they all agree. As Slater puts it, the winners “won’t be those who use AI most, but those who can still think without it” [4].

So what is actually scarce?

Whatever cannot be averaged.

The philosopher Michael Polanyi gave us the phrase sixty years ago: we know more than we can tell [6]. The surgeon’s hands, the trader’s bad feeling about a deal that looks flawless on paper, the way a great manager reads a room before anyone has spoken — none of it was ever written down, which means none of it is in the machine. It is tacit knowledge, and it is the one asset that does not depreciate the moment it is shared, because it cannot be shared. You had to be there. For years.

My favourite illustration comes from the investor Carson Block, who described his proudly Luddite father on a recent podcast: up 70% in a year holding Nvidia and CrowdStrike, without — by his son’s account — knowing what either company actually does [7]. The punchline: “If our algorithm ever broke down, I’d just ask my father what to buy.” A lifetime of pattern recognition, sitting quietly above the spreadsheet, waiting for the day the spreadsheet fails.

That is the part no model can reach. AI is built to produce an answer. It cannot sit comfortably in uncertainty, and it certainly cannot laugh at the gap between what the data says and what experience knows.

The people who beat the jar

Here is the part that should change how you think about your own edge. What made Tetlock’s superforecasters exceptional was not raw brainpower, and it was not a secret data feed. It was a method — and every ingredient of it is the precise opposite of how a machine produces an answer.

They are comfortable being uncertain. They think in fine-grained probabilities — not “yes” or “no” but “sixty-eight per cent, and here’s what would change my mind.” They treat every forecast as a draft, nudging it in small steps as new evidence trickles in, a habit Tetlock calls being in “perpetual beta.” And they deliberately gather clashing viewpoints and hold them in tension before deciding — the “dragonfly eye”: many lenses, one synthesis [3].

Read that list again and notice what it is. It is a human being doing, slowly and on purpose, the very thing the averaging machine does instantly and for free — except the machine skips the part that matters. It synthesises, but it cannot doubt. It aggregates, but it cannot hold two ideas in tension and feel which one is wrong. It will never tell you “sixty-eight per cent, and here’s what would change my mind,” because it was built to sound sure. The superforecaster’s edge was never information. It is judgment under uncertainty — and that is exactly what grows scarcer, and dearer, the more the world floods with confident, average answers.

The consensus engine cannot sell you the unpopular truth

Here is the structural catch, and it is fatal to the dream that AI will hand you an edge. AI is, by design, a machine for surfacing what most people already believe. Contrarian insight — the kind that spots the turn before the crowd — cannot come out of a system trained on the crowd. You cannot average your way to a non-consensus view. That is what “non-consensus” means.

In August 1979, BusinessWeek ran a cover declaring “The Death of Equities” [8]. It was the perfect distillation of everything sensible people believed, beautifully argued, and it landed almost exactly at the start of the greatest bull market in living memory. A 1979 AI would have agreed with that cover enthusiastically, in flawless prose, with footnotes. The median view is most comfortable precisely when it is most expensively wrong.

Where this leaves your money, and your mind

The investment translation is short, because the principle is. The average is now free. The edge is whatever can’t be averaged.

And the market was rigging itself for sameness long before the machines arrived. In the late 1950s the average American share changed hands roughly every eight years; by 2020, measured the same way, every five and a half months [13]. Roughly two thirds of the money in US stock funds is now passive — it buys everything, weighted by size, and asks nothing [13]. After Europe changed how research is paid for, 334 listed companies lost analyst coverage entirely; nobody is even looking [13]. You can see the sameness on the tape. Ten names are 37 per cent of the S&P 500 [14]. In 2025 alone, 79 per cent of active large-cap managers underperformed the index; over twenty years, 93 per cent have [15]. Everyone expresses their personality in the small positions and hugs the same ten giants with the real money. And sit with the loop inside that fact: the ten giants are the AI trade. The largest concentration in market history is a machine-averaged consensus about the averaging machines themselves.

So the premium shifts to the un-averageable: judgment built over cycles, relationships that took a decade to earn, the on-the-ground pattern recognition that still decides outcomes in real assets, private markets, and physical things — domains where being in the room beats reading about the room. Call it the cognitive-purity premium. It accrues to the people and the businesses that can still think without the machine, in a world where almost nobody bothers to try.

There is a warning tucked in here for every organisation merrily automating. It is gloriously easy to cut the explicit layer — the drafting, the modelling, the codifiable work — and bank the savings. It is much harder to notice you have also quietly stopped producing judgment, because the junior who used to earn it by doing the boring work now just prompts for it. Automate the explicit, by all means. Just don’t strip-mine the tacit, or in twenty years you will run a beautifully efficient company full of people who have never had an original thought, all agreeing with each other, fluently.

One last thing about the middle

When Tetlock’s team built their forecasting machine, they ran into something with a name that belongs on the cover of this piece. The crowd’s averaged forecast, even a good crowd’s, came out chronically too timid — huddled near the middle, hedged half to death. So they did something counterintuitive. They took the average and shoved it back outward, toward the extremes, toward conviction. They called it “extremizing,” and it made the forecasts more accurate, not less [3].

Sit with that. The best forecasters alive found that the average pulls you too far to the middle, and that the fix is to deliberately lean back out.

AI will write your emails, draft your code, and summarise your meetings. Let it — hand over the whole commodity middle with my blessing. Just don’t outsource the one thing it was never able to do: disagree with everybody, sit with the discomfort of being maybe-wrong, and turn out to be right.

A note of thanks before the close. A day before this went out, Ahmed Husain arrived at much the same fear by a different road, with evidence I had not found and a phrase I wish I had coined: a weirdness premium [16]. Two people reaching the same conclusion independently is, of course, precisely the diversity this piece says is disappearing. Long may we disagree.

The future, as ever, belongs to the weirdos who still think for themselves.

This article is 100% free to read, so please share it to pay it forward.

Thanks!

Subscribe now

References

[1] Francis Galton, “Vox Populi,” Nature (1907) — the original “wisdom of crowds” experiment: at a 1906 country fair, the median of ~800 independent guesses of an ox’s weight came within ~1% of the true figure. Popularly retold as guessing marbles or jellybeans in a jar.

[2] Cliff Asness (AQR Capital Management) on the “wisdom of crowds” failing in markets — the “poll the audience” analogy and a narrative-to-flows-to-price feedback loop. Asness is a public commentator; argument deployed here in paraphrase.

[3] Philip E. Tetlock and Dan Gardner, Superforecasting: The Art and Science of Prediction (2015). The Good Judgment Project — teams of amateur forecasters outperforming intelligence-community analysts with classified access; the superforecaster traits (”perpetual beta,” the “dragonfly eye,” probabilistic thinking); and the “extremizing” of aggregated forecasts to improve accuracy.

[4] Tom Slater, Investment Manager, Scottish Mortgage Investment Trust (Baillie Gifford), “AI Isn’t Coming For Your Job. It’s Coming For Your Mind,” 2026 — citing a 2026 study in Computers in Human Behavior finding AI use produces uniform overconfidence across skill levels, with the most AI-literate the least well-calibrated.

[5] Philip E. Tetlock, Expert Political Judgment: How Good Is It? How Can We Know? (2005). The roughly two-decade study in which the average expert’s forecasts proved “roughly as accurate as a dart-throwing chimpanzee,” and the most media-prominent experts were among the least accurate.

[6] Michael Polanyi, The Tacit Dimension (1966): “we know more than we can tell.” The foundational statement of tacit versus explicit knowledge.

[7] Carson Block (Muddy Waters Research), interview on the Financial Times Unhedged podcast, 2026.

[8] “The Death of Equities,” BusinessWeek cover, 13 August 1979 — the canonical example of consensus conviction arriving at a major market turning point.

[9] Scott E. Page, The Difference (Princeton University Press, 2007) — the Diversity Prediction Theorem: collective squared error = average individual squared error − prediction diversity. An algebraic identity, not an empirical claim.

[10] Emily Wenger and Yoed N. Kenett, “Large language models are homogeneously creative,” PNAS Nexus 5(3), March 2026 — 102 human participants vs 22 LLMs; model responses mirror other models far more than humans mirror humans.

[11] Liwei Jiang et al., “Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond),” NeurIPS 2025 (Best Paper Award) — 26,000 open-ended queries; pronounced inter-model homogeneity across 70+ models.

[12] Weixin Liang et al., “The Widespread Adoption of Large Language Model-Assisted Writing Across Society,” Patterns (2025) — 537,413 corporate press releases, Jan 2022–Sep 2024; up to 24% of text LLM-generated by September 2024; described by the authors as a lower bound.

[13] Market structure: Reuters/NYSE turnover data (holding periods, 2020 analysis); Investment Company Institute, “Active and Index Investing,” June 2026 (63.8% of US domestic equity fund assets indexed); Fang, Hope, Huang and Moldovan, “The Effects of MiFID II on Sell-Side Analysts, Buy-Side Analysts, and Firms,” Review of Accounting Studies (2020) — 334 EEA firms lost coverage entirely.

[14] S&P 500 constituent weights, official fund holdings file, 24 August 2026 (top-10 constituents = 37.0%).

[15] SPIVA U.S. Scorecard, Year-End 2025, S&P Dow Jones Indices — 79% of active large-cap funds underperformed the S&P 500 in 2025; 93% over 20 years.

[16] Ahmed Husain, “The Weirdness Premium,” The Curious Mind, Research Letter No. 2, 26 August 2026 — thecuriousmind.org/p/the-weirdness-premium.

And here is the audio podcast version (done with Gemini Notebook):

Everyone is hunting for the crash in data centers. Or in the hardware stocks. Or the debt markets.

They are looking in the wrong place entirely.

Saying that AI is a bubble has become the standard consensus view across Wall Street. But in finance, the exact moment an observation becomes consensus is the exact moment it loses all utility. The question is never whether we are in a bubble. The question is which kind of bubble it is, and which layer of the stack gets slaughtered when it pops.

Infrastructure booms throw off two fundamentally different kinds of bubbles. They behave nothing alike and they reward opposite investors:

• The Early-Cycle Bubble arrives early, as its name implies. It builds something magnificent that the next generation inherits for free. It crashes hard, but leaves the physical asset standing. Britain’s Railway Mania bankrupted its investors, but left six thousand miles of track. The dot-com buildout laid tens of millions of miles of fibre, lit barely ten percent of it, went bust, and quietly funded the next twenty years of the internet [1].

• The Late-Cycle Bubble arrives late. The technology has obviously won. Financing turns aggressive, and the cycle’s cheap input gets wildly over-consumed until the economics break. It feels gentler on the way up because real earnings sit underneath it. It is far nastier on the way down, because what it leaves behind is not a public commons but stranded capital.

We are in the second kind of bubble. AI is not a new industrial platform; it is a late-cycle efficiency input.

Inputs like that do not keep their profit margins. But the money isn’t vanishing into thin air either. It is migrating to the two extreme ends of the pipe: to the platforms that control customer distribution at the top, and to the physical bottlenecks that supply energy and land at the bottom.

What gets stranded is the layer in the middle. So you have to own the ends and be terrified of the middle.

LLMs Are Memory Chips, Not Processors

The right historical reference point for Large Language Models is not the railway or the internet. It is the microprocessor [2].

The microprocessor made the entire global economy exponentially more productive while the price of computation itself plummeted relentlessly toward cost.

Some inputs manage to keep their margins. Intel held fat rents on x86 architecture for decades, and Nvidia does the same today with CUDA. Why? Because a massive wall of software was built specifically for their design and could not easily move.

An input keeps its rent only when it locks its users in.

Large language models do the exact opposite. They are accessed through near-identical interfaces and can be swapped out in an afternoon. They are not the processor; they are the memory chip, a pure commodity where whichever supplier is cheapest and good enough wins.

The tell that a technology has entered its late cycle is who adopts it first. New paradigms start on the fringe as crude toys that serious executives ignore. This generation of AI was seized instantly by incumbent megacaps, because the enterprise rails were already laid. When the breakthroughs come from multi-billion-dollar corporate research budgets rather than a garage, you are looking at a mature cycle, not a wild new frontier [3].

The Great Price Split

The headline narrative says the price of machine intelligence is falling across the board but the data says otherwise.

The price didn’t fall, it split:

Since mid-2024, the cheapest capable models have anchored at roughly two and a half cents per million tokens. Meanwhile the most expensive flagship on the market today costs $465 per million tokens on a blend of 30% input and 70% output, against roughly $51 for the flagship of three years ago on the same blend [5]. The top of the market raised its prices roughly ninefold while the bottom collapsed to near-zero.

Today, roughly 43% of all available models cost under a dollar per million tokens. The gap between the most expensive and the cheapest is an absurd 19,000 times [5].

This structural split ruins the middle layer. A commodity layer has emerged at the bottom, where good-enough intelligence is practically free, while a hyper-specialised niche holds at the top. What is disappearing is the middle class of AI: the assumption that a very good proprietary model is a durable, broadly monetisable business.

And the proof keeps compounding. In July 2026 a Chinese open-weight model, Kimi K3, took the top spot on a public coding leaderboard ahead of every Western frontier model, and its creators released the weights for free download [4]. Days later, OpenAI cut prices on its smaller models, taking roughly 80% off the output price of one of them.

The repricing did not stop there. DeepSeek, the most watched of the budget providers, is reported to have raised its prices by roughly 205% over the same weeks. The reported gap between OpenAI’s cut model and DeepSeek’s flagship collapsed from 21 times to 1.4 times. The cheap end is no longer a race to zero; it is converging on a single commodity band.

A free or cheap model doesn’t need to defeat the frontier everywhere. It just needs to be good enough for ninety percent of daily tasks.

Data Centers Aren’t Warehouses of Obsolete Chips

If model prices are commoditising, doesn’t that mean the massive capital expenditure pouring into data centers is being incinerated?

Look at the operating numbers [6].

• Azure crossed $100 billion in annual revenue in Microsoft’s fiscal 2026, growing 43% year on year in the June quarter.

• Microsoft Cloud overall reached $214 billion.

• Google Cloud’s operating margin expanded from 20.7% to 35.6% in twelve months.

These are widening operating margins, not collapsing ones.

Paper valuation gains. Much of Big Tech’s net income spike stems from markups on private stakes in AI labs. The scale of it is visible in the index: FactSet’s blended second-quarter earnings growth for the S&P 500 was 50.4%, but 32.0% excluding Alphabet and Amazon [7]. That is why I look at cloud operating margins rather than net earnings.

Cash burn. Alphabet’s quarterly free cash flow turned negative for the first time since it listed, a $5.9 billion outflow. Amazon’s June quarter was a $7.7 billion outflow. Meta’s fell 91% to $784 million. Only Microsoft stayed firmly positive, at $19.6 billion. Together, on the companies’ own definitions, the four generated roughly $6.9 billion while capital expenditure surged, and their credit spreads widened.

So how to reconcile booming cloud operating margins with tightening cash? By understanding the two-clock option structure of data center construction.

Land, buildings and power grid connections require a twenty-four-month lead time. But the actual chips and accelerators, which are the majority of total cost, are ordered just three months before installation, only after real customer demand is visible.

As Amazon put it bluntly: “If the demand isn’t there, we won’t spend the capital.”

That flexibility is real, and it covers the chip orders. Only the chip orders. For the chips to be stranded, demand must vanish inside a single ordering window, and even then what stands empty is a building with a massive power link, not a warehouse full of decaying chips.

The other clock carries no such option. The filings already hold roughly $3 trillion of contracted commitments across the big spenders, around $2.3 trillion of it binding on the strictest reading [9]. A later piece in this series follows those signatures.

The Model is an App, Not the App Store

Wall Street committed a fundamental classification error. It priced AI models like two-sided platforms, assets that gain network effects and pricing power as more users join.

The model is not the App Store. The model is just an app. The hyperscaler is the App Store.

Consider what a distribution platform actually does

• It owns the end-customer relationship.

• It processes all telemetry and billing.

• It sees exactly which features get traction.

• And when a third-party app gets big enough, the platform builds its own version and defaults it onto the customer’s screen.

This playbook is currently being executed in AI with Hyperscalers owning the corporate billing accounts. They see which models are being called and at what price and when a third-party model charges a premium, Microsoft or Google simply deploys its own internal or open-weight alternative directly into the stack.

The rents do not belong to the model. They belong to whoever owns customer distribution at the top, and whoever owns scarce power and grid access at the bottom.

The $217 Billion Exposure: Where the Real Risk Lives

The strongest counterargument to this thesis is circular financing.

A chipmaker invests in an AI lab, which uses the funds to buy chips. A cloud platform invests billions in a lab, which turns around and commits those billions back to the platform in compute contracts. It echoes the circular accounting that unravelled telecom firms during the 2001 crash.

If the model labs are funded by the very platforms selling to them, what happens if the labs go under?

Microsoft disclosed two crucial numbers on the same earnings call [6]:

• Commercial contracted backlog grew +84% to $678 billion.

• Excluding OpenAI, backlog grew +25%.

Do the arithmetic: OpenAI alone accounts for at least $217 billion of Microsoft’s forward book. That is just under a third of it at the strictest floor, and plausibly closer to half. That figure is a floor, not an estimate: it assumes OpenAI had no backlog whatsoever a year ago, and any starting value above zero makes the share larger.

Current cloud revenue is highly diversified: nearly 90% of Microsoft Cloud revenue comes from customers outside the frontier laboratories. That measure also carries Office, LinkedIn and Dynamics, so the concentration inside the cloud infrastructure business alone is higher still. The forward contracted book is wildly concentrated.

If a laboratory fails to raise its next multi-billion-dollar equity round, it won’t trigger an immediate drop in today’s cloud computing usage. Instead it creates a severe credit and lease exposure on the forward books of the platforms. One rating agency has already acted on exactly this risk. In July 2026, S&P downgraded Oracle and wrote that “OpenAI remains a key credit risk”, estimating that OpenAI makes up “roughly half of the $638 billion” in Oracle’s contracted book [8]. Moody’s holds the same company on negative outlook, citing the counterparty risk of its largest AI infrastructure customer.

What Happens When the Buyer Changes?

If model laboratories stumble, won’t traditional enterprises and governments absorb the extra compute capacity?

Yes. But at a fraction of the price.

Every mile of dot-com fibre optic cable was eventually lit and used. But the equity owners who laid that fibre still went broke; the eventual returns went to distressed-asset buyers who bought the infrastructure for pennies on the dollar.

Why? Because the buyer changed:

• Laboratory demand is funded by capital markets. It demands peak hardware and maximum memory bandwidth at almost any price, to win the intelligence race.

• Enterprise demand is funded by corporate operating budgets. It demands good-enough inference performance at a cost that clears an internal committee.

When capital-market funding slows, the volume of compute demanded doesn’t disappear. What disappears is the buyer who would pay anything for the newest, most memory-dense configuration, replaced by one who will only pay for the cheapest arrangement that works.

Volume holds. Realised price per unit falls. That is the same bifurcation as the model-price chart, seen from the demand side.

Training wants maximum memory bandwidth per chip whatever it costs; inference wants the cheapest adequate configuration. So the memory content of each machine is not a constant. It is a function of the training-to-inference mix, and that mix is about to move.

The Investor’s Playbook

On its face this is not an absurd multiple bubble. Names like Nvidia, Broadcom and Meta trade at growth-adjusted multiples below 1.0. Keep in mind though that the earnings beneath those multiples carry the paper markups flagged above.

The risk is not that price-earnings multiples are crazy. The risk is that future earnings assumptions rely on middle-layer rents that cannot be sustained.

So that is where this Substack will spend its time. When I try to understand an AI related investment opportunity in the issues ahead, the work will sit at the two ends of the pipe: the platforms that own customer distribution at the top, and the physical bottlenecks that own power, land and memory at the bottom. The middle is what I will seek to avoid.

Subscribe now

Core takeaways

The moat is at the distribution layer. The platforms hold their position regardless of which model wins, because they own the customer relationship and the billing, they can see which models are actually consumed and at what price, and they can ship their own version of anything that works. That is not a forecast about technology. It is where the defences in this stack actually sit. One boundary belongs in the same breath: the rent is real, and the committed capacity is the cost of defending it, signed years ahead of the revenue it serves. The moat and the bill come together.

The second moat is physical, and narrower than it looks. Power, land, memory and storage are genuinely scarce and cannot be conjured. But memory content per machine is not a constant; it is a function of the training-to-inference mix. Anyone holding a bottleneck because “AI needs it” is really betting that the mix holds.

Be sceptical of the middle. Most of it is unlistable, be sceptical of anything priced as though it will one day charge a platform’s toll. Its margins will likely converge toward cost as open-weight models improve.

And note where the exposure actually sits. When a laboratory cannot raise its next round, today’s cloud usage does not stop. What moves is the forward book, and that is concentrated in a way current revenue is not. It arrives as marks and leases on platform balance sheets. It is a credit exposure, not a demand collapse. That is a different risk from the one most people are watching for, and it is the one already being acted on.

So here is the position and what would change my mind. The bubble sits in the model layer, not the infrastructure layer: model economics converge toward zero profit while the platforms above them and the bottlenecks below them keep the rent.

Two things would tell me I have this wrong. The first is a proprietary frontier model rebuilding real pricing power. I would see it in the same price capture behind [5], re-pulled each quarter: the top flagship’s premium over the cheapest capable tier widening again, and holding for some time. The second is a platform’s cloud growth stalling on ordinary enterprise demand rather than on laboratory commitments. Microsoft discloses that itself every quarter: the non-frontier share of Microsoft Cloud, and commercial remaining performance obligations excluding OpenAI. I am keeping a close eye on that.

Neither is happening yet and the price of intelligence has split rather than fallen. The floor has been flat since mid-2024 while the frontier costs around nine times what it did three years ago on a like-for-like blend. And the reported gap between OpenAI’s cheap tier and DeepSeek’s has collapsed from twenty-one times to 1.4. Remember that open source models are free in the sense that they are free to own, but the expenses to run them remain. Open-weight models lead at least one public capability board. Cloud revenue is growing above forty per cent, with nearly ninety per cent of Microsoft Cloud sold to non-frontier customers. The committed backlog is growing eighty-four per cent, or twenty-five excluding a single frontier customer.

Knowing that AI is in a late-cycle input bubble won’t tell you the exact date of the top. But knowing which layer the bubble sits in tells you exactly what to hold while you wait.

References

[1] Carlota Perez, Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages (Edward Elgar, 2002). The frenzy and maturity phases are Perez’s framework; reading the two as leaving a usable commons versus stranded capital is Singh’s extension of it (see [2]).

[2] Sameer Singh, “AI: The Wrong Kind of Bubble,” Breadcrumb, 2026. The microprocessor reference class and the input-not-platform reframe draw on this essay. This piece departs from it on where the stranding lands.

[3] Nicolas Colin, “Late-Cycle Investment Theory” (Drift Signal, 2025) and “The Beginning of the End for Big AI” (Drift Signal, 2026), for the tell that a paradigm has matured.

[4] Arena.ai Frontend Code leaderboard, July 2026, and Moonshot AI release materials for Kimi K3. Standings re-verified at publication.

[5] Model pricing from a full capture of a public model marketplace listing, 19 August 2026. Blended price defined as 30% input plus 70% output per million tokens; batch and free tiers excluded. The 2023 comparison is on the same blend: the then flagship, at $30 input and $60 output per million tokens, blends to $51. Prices move weekly; the capture is re-verified at publication and re-pulled quarterly thereafter to score the falsifier.

[6] Microsoft fiscal 2026 fourth-quarter results and earnings call, 29 July 2026, for Azure and Microsoft Cloud revenue, the share of cloud revenue from customers outside frontier model companies, and commercial remaining performance obligations. Alphabet, Amazon and Meta second-quarter 2026 results and calls for cloud margins, capital-spending commentary and free cash flow; free-cash-flow figures are the companies’ own reported definitions for the June quarter.

[7] FactSet Earnings Insight, August 2026, for blended S&P 500 second-quarter earnings growth of 50.4% and 32.0% excluding Alphabet and Amazon.

[8] S&P Global Ratings, “Oracle Corp. Downgraded To ‘BBB-/A-3’ From ‘BBB/A-2’ On Rising Business Risk And Weaker Cash Flow; Outlook Stable”, 9 July 2026, including the estimate that OpenAI makes up roughly half of Oracle’s $638 billion in remaining performance obligations. Moody’s Investors Service, affirmation of Oracle at Baa2 with negative outlook, February 2026, citing counterparty risk associated with its largest AI infrastructure customer.

[9] Purchase-commitment, lease and power-agreement disclosures in the largest AI spenders’ most recent annual and quarterly filings, at SEC EDGAR. The roughly $3 trillion contracted and roughly $2.3 trillion binding aggregates are my own reading of those notes; a later piece in this series shows the work.

[10] National Bureau of Economic Research, Business Cycle Dating Committee chronology; the NBER-based recession indicator is FRED series USREC. The February to April 2020 contraction is the only recession in the fifteen-year window.

Start with a desk.

Every computer has two places to put data, and they do completely different jobs.

The filing cabinet in the corner is the hard drive. It holds everything you own, cheaply, and it is painfully slow. You have to walk over and find the folder.

The desk in front of you is the memory. It holds only what you are working on right now. It costs vastly more per square inch, and it is instant.

The entire point of the desk is immediacy. When you buy memory, you are not buying storage. You are buying time.

For forty years the desk was the boring part of the technology industry. Then artificial intelligence arrived and demanded a desk the size of a football field.

Why AI is starving for memory

An AI model is not a search engine looking things up. It is a vast grid of numbers, hundreds of billions of them, which are the weights it learned during training. To answer a single question, the whole model has to be run through the processor.

This is the bottleneck almost everyone missed. The entire model has to be sitting on the desk at once. Not in the filing cabinet. Every single time anyone asks it anything.

A data centre packed with the fastest processors on earth is useless if there is nowhere to put the numbers those processors need. The industry spent three years working out that the binding shortage was not the graphics chips. It was the desk space next to them.

The answer is High Bandwidth Memory, or HBM: ordinary memory chips stacked twelve or sixteen layers high, drilled through, bonded, and placed hard against the processor. Data travels millimetres instead of centimetres. It is extremely difficult to manufacture, and it is now the single component holding up the global AI build.

The catch: an oligopoly that has never stopped overbuilding

Only three companies make HBM at scale: Samsung, SK Hynix and Micron. SK Hynix is currently the best at the hardest part of it.

In any normal industry, three suppliers controlling a bottleneck for a generational technology would mean you buy the shares, close your eyes, and collect the rent.

Memory has never been a normal industry. It has been a ruthless commodity:

• Buyers choose on price. Historically there has been no software lock-in and no ecosystem loyalty.

• When memory is scarce, prices triple and the makers print money.

• When it is plentiful, prices collapse and they lose money on every chip.

And here is the flaw that has defined the industry for four decades: they have never once stopped themselves from overbuilding. They do not even need new factories to flood the market. They convert existing lines and shrink the circuitry on the chips they already make. Samsung raised its output of memory by about 40% in a year, measured in bits rather than chips, and spent almost nothing extra doing it. Supply arrives without anyone having to fund it. That is why the good times end faster here than in copper, oil or shipping.

So the question is not whether AI needs memory. It plainly does. The question is whether this time the industry finally behaves differently.

There is now real evidence that it is, and that is what most of the market is still not pricing.

Why the demand panic is wrong

SK Hynix shares peaked at 2,987,000 won on 25 June and closed at 1,504,000. They are down 49.6% in seven weeks. The market is pricing the end of the AI build. The evidence does not support it.

1. Customers are paying up, and saying so out loud.

Amazon raised its 2026 capital spending from $200bn to $220bn and its chief executive named rising memory prices as the reason. In the same results it disclosed a contracted cloud backlog of $496bn, up $132bn in one quarter, and said it will not have enough capacity to meet demand in 2026 or 2027. A customer who raises his budget to meet your price is not a customer with weakening demand.

2. The product is now genuinely contracted, and this is new.

On 28 July the company said it has concluded long-term agreements with around ten customers, including its key customers, with further discussions running. On the same call it set out the terms. They run normally around five years. They carry long-term volume commitments. And they carry deposits, which the company says are there “to support contract fulfillment and enhance visibility” of demand. Pricing is deliberately not fixed: the company uses “a range of pricing mechanisms that can better respond to price volatility.”

Read what that structure actually is. The volume is committed and the price is not. For thirty years this industry sold neither. A business with contracted volumes and open pricing is not a commodity producer, it is a toll. That is the single most important change in this company’s history and it happened three weeks ago.

And the deposits matter more than they sound. A customer who puts cash down in advance is part-funding the capacity that will serve him. The oldest question to ask of any supplier is who pays for the next factory. When the answer is the customer, the supplier keeps the rent. That answer has never been available in memory before.

3. It is not just the frontier labs.

The loudest bear case is that the big model labs run out of money. But Microsoft has disclosed that about 90% of Microsoft Cloud revenue comes from ordinary corporate computing rather than frontier AI. That measure includes Office 365 and LinkedIn, so treat it as a direction rather than a precise number, but the direction is clear and it is growing on its own.

4. The supplier is not releasing supply. It is contracting it.

Conventional DRAM wafer capacity is flat in 2026 and shrinks 4.0% in 2027. The share of wafers going to AI memory is forecast to rise from 18% at the end of 2025 to 22% in 2026 and 30% in 2027. Nobody is flooding anything.

Management describes its expansion in terms this industry has never used. Equipment investment and production ramps will run “in phases while considering demand visibility, investment efficiency”, staying ”flexible in alignment with confirmed customer demand.” The historical failure was building against forecasts. This is building against orders.

5. Prices have not turned.

A benchmark chip traded at $42.61 on 12 August, having risen from $12.76 last November. Contract prices rose 93% to 98% in the first quarter and 58% to 63% in the second, with 13% to 18% forecast for the third. The rate of increase is collapsing. A slowing rate of increase is not a falling price.

6. Even the bears’ best evidence points the other way.

Nvidia cut the memory in its Rubin Ultra design from about 1TB to 192-256GB, which was widely read as weak demand. The industry’s own price reporters state the cause directly: the 2027 shortage means suppliers cannot deliver the volume. That is a buyer rationing itself in a shortage, not a buyer that needs less.

So why did the shares halve? A single fund running about $45bn, heavily concentrated in exactly these names, took margin calls at the end of July and was forced to sell most of its public book to Citadel, ending with roughly $10bn. The KOSPI fell about 33% from its June peak. That is a positioning event, not a demand event.

The real pressure point, and how it actually resolves

There is one genuine mechanism that could break this, and it deserves to be understood properly rather than waved at.

A memory factory makes both kinds of chip from the same silicon wafers. Ordinary memory and stacked AI memory. Same buildings, same wafers, a choice about what to run. AI memory yields roughly a third as many gigabytes per wafer because it is so much harder to make.

So AI memory has to sell at a large enough premium to justify giving up the ordinary memory those wafers could have made. For two years that was easy, which is exactly why ordinary memory went short and its price tripled. Not because anyone stopped wanting ordinary memory, but because the wafers that would have made it were making something else.

And the premium has been closing, because ordinary memory has risen far faster than AI memory. If it closes far enough, the wafers should flow back, ordinary supply would rise, and the price of the product that earns most of the profit would fall.

How violent would that be? The industry’s own history gives exact answers, and they are brutal:

• 1.38 points of surplus was enough to turn the price.

• 5.8 points of surplus broke it.

• One to two points of shortage is what tripled it.

Small moves in volume, violent moves in price. That is the whole industry in three numbers.

But the mechanism has now been tested in public for nine months and has not fired. The price reporters flagged it last October and reported capacity beginning to shift two months later. Since then the share of wafers going to AI memory has gone up, not down. Ordinary memory set a record price. Nobody blinked.

Here is why, and it is the part the bears have wrong:

• Leaving AI memory is easy. Coming back is not. Ordinary memory needs a wafer and nothing else. AI memory needs a wafer plus a second step that happens somewhere else: stacking, drilling, bonding, and a control chip bought in from TSMC. The wafers are shared between the two products. That stacking capacity is not, and it had to be built on purpose.

• So a maker who moves wafers out to chase today’s ordinary price may not get them back in when it matters, and would surrender its position in the most important hardware of the decade to do it.

• And the shortage itself is now the brake. The expected 2027 shortage limits how much capacity can go to AI memory at all, because ordinary memory is too short and too profitable to starve any further.

Which leaves one resolution, and it is the bullish one. If the premium is too thin and no wafers can move, then the premium has to widen from the other end. The industry’s price reporters expect exactly that: AI memory contract prices rising several-fold in 2027. The gap closes by AI memory repricing upward, not by ordinary memory collapsing. Nobody has to switch, and the shortage holds.

That is the reasonable base case, and the market is not paying for it.

What management actually did, and what it does not mean

In July SK Hynix sold $26.5bn of new shares on the American market. Within twenty-four days it approved roughly $41bn of new factory building.

This is usually presented as a warning sign. Read the actual filing and it is mostly not.

The prospectus states what the money is for, and it is specific: 45.5 trillion won for Fab 1 at the Yongin complex and the P&T7 advanced packaging plant at Cheongju, both completing at the end of 2030, plus 11.9 trillion won of EUV lithography scanners for delivery by December 2027.

And it says the raise does not cover it. That is a 57.4 trillion won programme against net proceeds of about 26.2 billion dollars, roughly 37 trillion won. The company writes that it expects to fund the difference from operating cash flow, borrowings and debt. So the common claim that they did not need the money is simply wrong on their own document. It is also worth knowing that quarterly free cash flow is about $12.5bn, not the $40bn of quarterly operating profit that gets quoted as though it were cash.

There are four reasons on the record, and three of them are ordinary:

1. A Nasdaq listing requires a primary raise. This was the largest first-time US share sale by a foreign company ever, seven times oversubscribed, priced at $149, up 13% on debut. You cannot list without selling something.

2. There is a published building programme, and the money does not cover it.

3. Washington is applying pressure. The US Commerce Secretary has been openly lobbying both Korean makers to build American fabs, at the same moment the two of them pledged over $550bn to Korean plants. A US listing and a US shareholder base is a cheap opening move.

4. Management may have thought the price was good. This is the fourth answer, not the best one.

But there is one genuinely interesting fact in the sequence, and it is not the raise. The prospectus filed on 10 July names Fab 1 and P&T7 and nothing else. The terms “M17” and “Fab 2” do not appear in it once. Eighteen days later the company’s own results release named M17, a second Yongin phase, and a new cluster. So they raised against one program and then approved a larger and different one almost immediately.

That is a company responding to demand it did not have visibility on three weeks earlier. It is not a company calling the top.

Valuation: pricing the cycle

Last quarter the company earned 60.5 trillion won of operating profit on 79.3 trillion won of revenue. At 1,418 won to the dollar, that annualises to about $224bn of revenue and $171bn of operating profit, on a 76% margin, against a company worth about $773bn.

So the whole company costs about 4.5 times one year of current operating profit. That is not a cheap business. It is a peak one. And note the basis: that is a multiple of pre-tax operating profit, not earnings, so the true price-to-earnings figure is roughly a third higher again.

One number has to be set aside, and it is not the one usually named. Last quarter’s bottom-line profit was 93.9 trillion won, half as much again as the 60.5 trillion the factories earned. The difference was not made by making memory. Most of it is a gain on the company’s convertible bond in the Bain-led vehicle that owns Kioxia, the Japanese memory maker, which converts into a stake in that vehicle rather than into Kioxia shares directly. Only about a third of the holding was actually sold. The rest is carried at what the market says it is worth, and Kioxia’s share price roughly halved in July. Every number in this piece uses the factory profit and ignores that gain entirely.

The three cases

How each case is built:

• The bear respects the company’s own history. Revenue fell 33% in 2019, with the margin going from 51.5% to 10.1%. It fell 27% in 2023, with the margin at minus 23.6%. Volume growth has never once prevented a revenue decline. A 25% fall is therefore milder than both precedents, deliberately.

• The bear margin is 25% rather than the 20% the industry averages through a cycle, and the contracts are the reason. The catastrophic margins of 2023 happened because price *and* volume collapsed together and the fabs ran half empty. Committed volumes with deposits are precisely what prevents that. A price-only downturn does far less damage than a price-and-volume one.

• The base assumes today’s 76% margin is the anomaly it obviously is and halves it, while volume growth offsets falling prices to hold revenue flat.

• The bull needs the multiple to change, and now has a reason. Thirteen times instead of ten is what you pay for a business with contracted volumes rather than a pure cyclical. Before those agreements existed that was a hope. It is now a mechanism.

• The multiple is a judgement, not a fact. At 8x instead of 10x, every number falls by a fifth.

The risk that trades at nothing

Two of the company’s plants sit in China: Wuxi, which makes legacy DRAM, and Dalian, the former Intel NAND plant. Their US export waivers were revoked at the end of 2025 and replaced by an annual site-license review. The company’s access to American equipment for a meaningful part of its capacity is now renewed once a year, at Washington’s discretion. That is currently priced at nothing.

The bottom line

Demand for high bandwidth memory is not weakening, and the 50% fall was a forced seller in a crowded trade rather than a change in the business. What has changed, and what the market has not absorbed, is that this company now sells contracted volume at open prices. That is a different business from the one whose forty-year range everyone is quoting at it. The pressure point on wafers is real, and on the evidence it resolves through AI memory repricing upward rather than through ordinary memory collapsing.

I am treating this as a tactical position with a rerating case attached rather than a buy-and-forget.

Wrong if:

1. The share of wafers going to AI memory starts falling while prices hold. That would mean the supply response of the past forty years is finally arriving.

2. The long-term agreements turn out to carry no deposits and no real volume commitment. The structure is the whole rerating case. If the terms are soft, the case is soft.

3. A benchmark contract price prints a quarter-on-quarter fall. Not a slower rise. A fall.

Status:

1. On its last filed accounts, 99.87% of 2025 revenue and 99.94% of first-quarter revenue was recognised at a point in time, with no backlog and no take-or-pay. The agreements announced on 28 July are not yet in any filing.

2. The share of wafers going to AI memory is still rising, not falling.

3. The most recent benchmark spot price is a new high, not a fall.

None of the three conditions is met.

References

1. SK hynix Inc., Prospectus (Form 424B4), filed 10 July 2026. SEC CIK 0002120882, accession 0001193125-26-299963. Use of proceeds, capital expenditure table, revenue recognition by timing, and China facility risk factors.

2. SK hynix Inc., “SK hynix Announces 2Q26 Financial Results”, 28 July 2026. Revenue, operating profit, net profit, long-term agreements, and mid-to-long-term investment plans including P&T7 and M17.

3. Amazon.com Inc., second-quarter 2026 results and earnings call, 30 July 2026. Capital expenditure guidance and AWS backlog.

4. Microsoft Corporation, quarterly disclosure of Microsoft Cloud revenue composition.

5. TrendForce, Memory Spot Price Update, weekly, November 2025 to August 2026. DDR4 1Gx8 3200MT/s spot prices.

6. TrendForce, DRAM contract price forecasts, quarter on quarter, 2026, published 3 July 2026.

7. TrendForce, “DRAM Supply to Remain Tight in 2027, Prompting NVIDIA to Lower HBM Configurations for Rubin Ultra”, 4 August 2026.

8. TrendForce, HBM wafer allocation and 2027 contract price outlook, 2 June 2026.

9. US Bureau of Industry and Security, “Revocation of Validated End-User Authorizations in the People’s Republic of China”, Federal Register, 2 September 2025, effective 31 December 2025.

10. EODHD market data. SK hynix (000660.KO) closing prices and USD/KRW, to 12 August 2026.

Click here for : Part1 - This is Part 2

Most market commentators spend their lives playing a trivial game: guessing where the S&P 500 will close in December or predicting when a recession will occur. The track record of point forecasting is atrocious, recessions also. The people who sound the most certain are almost always the ones who have forgotten their long history of being wrong.

I take a different approach. I focus on locating where we sit inside a much longer macro rhythm. If you understand the tide, you do not need to predict every individual wave. You simply need to know what assets are worth holding, which risks matter, and what noise you can safely ignore.

The market cycle everyone used to watch is the business cycle, a brief oscillation of expansion and contraction lasting a few years. But even that cycle is breaking down. Historically, the United States suffered an economic contraction roughly every six years. Over the last fifteen years, we have seen exactly one recession, and it lasted all of two months during the 2020 lockdown.

Beneath the business cycle runs a far more powerful engine: the debt and inflation super-cycle. It is a generational tide where the cost of money falls for thirty to forty years, and then rises for the next thirty to forty.

The Arithmetic of the Tides

You can date the turn in 1981 with precision when Volcker raised rates. The recent trough sits somewhere between 2012 and 2020, depending on how you account for emergency pandemic intervention. But the broader pattern holds regardless of the exact dating. The current rising leg is young and could still look like a cyclical bounce, but the fiscal mechanics behind it suggest the secular tide has officially turned.

These patterns persist because the underlying forces move at generational speed. Debt loads accumulate over decades. Demographics shift over half-centuries. Inflation psychology, the quiet assumptions that workers and companies carry into salary negotiations and corporate contracts, takes an entire generation to learn and another to unlearn.

We only have two and a half full regime transitions in modern US history:

• The post-war inflation run

• The Volcker reversal of 1981

• The structural turn unfolding right now

When you have a sample size that small, you cannot rely on statistical overfitting. You have to look at the current, real-world evidence: valuations, credit conditions, and of course fiscal policy.

What Wins When the Tide Flips

Where you sit in this cycle dictates which assets outperform.

Era 1: Falling Cost of Money

When money gets cheaper across a multi-decade span, financial assets dominate. Equities surge because a lower discount rate makes future earnings worth more today. Growth stocks perform best of all because their projected earnings sit furthest out in time. Bonds rally alongside stocks. Gold, energy, and real assets lag behind.

Era 2: Rising Cost of Money

When money gets dearer, the order completely inverts. Commodities, tangible assets, and gold lead the board. Bonds lose real money as yields climb. The most expensive growth stocks get crushed as the same discount-rate arithmetic that helped them begins to hurt them.

The transition between these two regimes is where we sit today. The assets that led for forty years are now concentrated inside passive market indices and institutional portfolios to an extreme degree (I will be writing in coming issues about passive and how it is impacting market stability). That concentration is precisely where the most capital will be surrendered as the regime shifts.

How We Got Here: The Anatomy of a Late-Stage Bull

The secular bull market that most investors view as “normal” was built on a single tailwind: relentless disinflation.

From the early 1980s onward, the cost of money fell almost continuously. Volcker broke 1970s inflation, globalization injected a billion lower-cost workers into the global economy to suppress wages, and the discount rate applied to corporate earnings plummeted. A huge portion of what passed for investment genius over the last forty years was simply this discount-rate math at work, aggressively amplified by passive index flows and corporate stock buybacks.

Dated from the 2009 low, this bull market is about seventeen years old. It can be measured against post-war secular bulls that historically lasted sixteen to eighteen years, so if history is any guide, this one is pretty old.

The valuation metrics look even more stretched than the calendar count:

• Shiller CAPE Ratio: Sitting above 40, more than double its historical average.

• Index Concentration: The ten largest names make up roughly 40% of the entire index, double their weight from a decade ago.

• Profit Margins: Put simply, at historic highs.

• Retail Allocation: Household equity holdings sit near record territory, the five highest readings in the history of the series are all from the last eighteen months.

An expensive market, operating on peak profit margins, owned by almost everyone: this is the exact cocktail that has preceded every major secular top in history.

Every great secular bull relies on a narrative technology at its core. It was radio in the 1920s, the internet in the 1990s, and artificial intelligence today. That technology explains the enthusiasm fueling the move, but it tells you nothing about when you have to stop dancing and the music stops.

Crucial Distinction: A high valuation is not a sell signal on its own.

The conditions that historically end secular bull markets are a tightening credit environment and a central bank out of ammunition. Neither is present today. The yield curve is positively sloped, credit conditions are loosening, earnings are growing, and the Federal Reserve has ample room to cut rates. The massive AI capital expenditure boom continues to inflate corporate earnings while concentrating systemic fragility.

My positioning reflects this reality: I remain invested through this final leg. The returns during a final melt-up are massive, but the exit door will be narrow and violently crowded. This is the exact moment to quietly structure hedges, because the odds of timing the top are close to zero (but so exhilarating to try).

The Inflationary Aftermath

When this bull market eventually terminates, the regime that follows will be defined by recurring bouts of inflation.

The driving force is political dominance. When sovereign debt expands beyond what economic growth or taxation can service, governments historically inflate away the real burden rather than default or enforce fiscal austerity. Modern political incentives across both major US parties favor spending over consolidation. At the same time, the world is re-arming and re-shoring supply chains, swapping a supply-abundant world for a supply-constrained one. Persistent deficits meeting structural supply constraints is the classic recipe for volatile inflation.

Addressing the Three Main Objections

1. “Look at Japan”

Critics argue Japan ran massive deficits for thirty years and generated nothing but deflation. This comparison fails for three clear reasons:

• Balance Sheet Dynamics: Japan’s government spending was filling a crater created by a private sector aggressively paying down debt. US deficits today run alongside a private sector whose balance sheet repair is finished and whose debt service is well below 2007 levels. US deficits are net additive to demand.

• Demographic Isolation: Japan aged first and alone, offshoring production to cheap foreign labor. A world aging simultaneously across all major economies cannot use that trick.

• Creditor vs. Debtor: Japan ran its experiment as the world’s largest creditor nation, funded by captive domestic savings. The US is a twin-deficit debtor dependent on foreign capital. A debtor country engaging in fiscal excess sees the pressure hit its currency and domestic price levels.

2. “AI is Deflationary”

A technology driving the cost of intelligence toward zero is undeniably a supply-side disinflationary force, much like China entering the global trade system. But AI is a supply-side shock, while our macro regime is driven by demand.

Cheaper production does not stop fiscal deficits from flowing directly into economy-wide spending. In fact, if AI displaces workers on a large scale, the political pressure for government transfers and deficits will increase dramatically. The technology that makes goods cheaper to produce simultaneously strengthens the political imperative to print and spend.

3. “The Wage-Price Spiral is Dead”

Labour lacks the unionization and contractual indexing it held in the 1970s, so price shocks burn out faster rather than compounding automatically into higher wage rounds.

This mechanism is real, which is why inflation will arrive as a series of sharp, unpredictable steps rather than a smooth vertical climb. But a broken wage-price spiral only limits how fast inflation accelerates; it does not stop it from persisting. When the inflationary impulse is fiscal, driven by government transfers into a supply-constrained real economy during every downturn, the persistence comes from the deficits, not the wage negotiations.

The Trap: Expect the Decoy First

Here is where most macro thinkers get caught out: sequencing.

The transition to a high-inflation regime might not happen in a straight line. A secular bull market so heavily dependent on a concentrated capital-spending boom usually ends in a sudden deflationary scare:

1. Risk assets sell off sharply.

2. Credit spreads widen rapidly.

3. Capital flees into long-dated US Treasuries.

4. The Federal Reserve cuts interest rates aggressively.

This first phase will probably look and feel like 2008. It will fool a generation of investors into thinking the old deflationary playbook still works, tempting them to pile back into long-duration paper and growth tech.

That deflationary crash will be the decoy. The massive fiscal stimulus deployed to counter that shock will travel directly into the real economy rather than getting trapped on bank balance sheets. That is when the secular turn asserts itself, cementing a decade defined by volatile, step-function inflation.

Add to this a massive, completely untapped powder keg of demand: trapped home equity. Years of relentless real estate appreciation have handed households a gigantic pool of wealth, but it currently sits frozen behind high mortgage rates. Owners' equity in US real estate is $34.9 trillion as of Q1 2026, with the 30-year mortgage at 6.69%. That is larger than the entire US economy. That number is roughly the size of US GDP. When the Fed inevitably slashes rates to arrest that initial deflationary scare, they will inadvertently light the fuse on this dormant savings pool. Millions of homeowners will finally refinance, cash out, and spend. That newly unlocked wealth will flood directly into the real economy, throwing a massive secondary wave of fuel onto the recovery.

Where This Leaves Us

The long-cycle framework does not attempt to pinpoint the exact Tuesday the market tops out. It tells you where you are standing.

We are late in a forty-year decline in the cost of money, navigating the final leg of an equity bull market that is expensive, narrowly concentrated, and widely owned at peak corporate earnings.

I am actually creating a full “Regime Board” to help you, my readers, keep track of the indicators I am following and estimate how close we might be from the edge of the cliff. I will be sharing it with you when it is ready.

Subscribe now

The path ahead involves a likely sharp deflationary scare first, where capital runs to traditional safety, followed by a far more volatile inflationary era. Because those two phases reward completely different asset classes, getting the sequence right matters far more than guessing the exact date.

Play the final leg of the bull market, but stay close to the exit. The regime change is coming, and it will not announce itself nicely.

When I wrote about Magnum back in May, I left one big question open. I called it the load-bearing one: does the listed company structure put a permanent ceiling on Magnum’s margin compared to what Froneri can reach?

Froneri is the Nestlé and PAI joint venture making ice cream in the exact same countries, buying the exact same dairy, cocoa, and packaging. Everyone uses them as the yardstick so I did too.

But as I do for every name I discuss here, I went back to dig more through Froneri’s accounts and made a few interesting discoveries.

How Bad Math Becomes Gospel

The number floating around every trading desk is that Froneri earns 26%. Against Magnum’s mid-teens, so what is Magnum doing wrong? Add in the fact that Froneri owns Häagen-Dazs outright in North America, bought for cash a quarter of a century ago and royalty-free until 2110, which is a setup Magnum cannot replicate.

But Froneri’s audited accounts show 21.5%, and 21.6% the following year. The year before that, they earned 18.3% and actually lost money before tax. Even the rating agencies covering their debt only forecast a peak in the low twenties.

So where does the 26% come from? It is worth understanding, because this mechanism is incredibly common and it is exactly how garbage numbers travel around.

Somebody took a press-reported enterprise value, divided it by a press-reported transaction multiple to back out an EBITDA figure, and then divided that by revenue to get a margin. The multiple and the margin are the exact same data point and that phantom EBITDA sitting in the middle is a heavily adjusted figure prepared for a debt syndication pitch, running about a fifth higher than the audited number. The leverage statistic everyone blindly quotes for Froneri comes from that exact same denominator and understates the real debt load by more than a full turn.

The Math Works

So the open question I had in May has an answer and it is not the one implied. Magnum’s reachable range is 19% to 22% so the cap I could not rule out looks like it is not even actually there.

I want to be careful about what this does and does not settle. It does not make the cost-cutting program easy. Management is guiding to forty to sixty basis points a year, which is a glacial pace, and slow corporate turnarounds fail regularly. It also does not make Froneri a weak operator. Pulling 21.5% is a very strong margin in this category, and Magnum still has a steep hill to climb.

What it removes is the arithmetic impossibility of catching them.

Subscribe now

Taking the ‘L’ on the Overhang

There is one other thing I have changed my mind about, and this one cuts the other way.

In May, I described Unilever’s remaining stake as coming to market on an orderly, scheduled basis. That is not what the filings say.

Unilever still holds essentially all of it. There are absolutely no lock-up arrangements, and the whole position has been registered for an at-will sale since before I even bought the stock. We are talking about a block equivalent to roughly a quarter of the free float, just sitting there undated. The optimist stance is that Unilever trusts the business and wants to remain invested, the pessimist stance is that the can dump that stock at any time.

The Bottom Line

Despite the overhang, the position is up around 23% and I am keeping it at full size. It is still trading below the multiple its private peer commands so I will continue holding it as long as the margin continues to trend towards that number.

Globus Medical gets booted out from the medtech basket. The ETF remains, it is the simplest way to express this trade, this is only a small one-time adjustment to the basket.

The rationale here is painfully simple. Back in July 2024, the FDA sent Globus a warning letter regarding its flagship surgical robot. Fast forward two full years, and as of its most recent annual report, that letter was still open and I missed that. I could have left the basket as is but this is one of my highest conviction themes for the next 12 months so it deserves to be as “perfect” as I can make it.

To be fair, an FDA warning letter is not a product recall, nor is it a red flashing safety hazard. It is simply the regulator informing a company that its quality control system is lacking. Normally, management teams scramble and close these things out in a few months.

Globus has not. Worse, this lingering demerit sits squarely on the ExcelsiusGPS robot. That is the exact product carrying the company’s growth narrative, rather than some forgotten legacy side hustle. While that file stays open, management is busy negotiating with regulators instead of actually selling robots. You better believe any potential acquirer looking at the division is going to aggressively price in that exact headache.

I built this basket because medical device businesses are trading at irrationally cheap prices. That thesis is entirely intact but that argument relies on a rerating. A company nursing a two-year regulatory hangover on its flagship product does not get to be treated like a normal business until the FDA explicitly says so.

Then there is the sheer absurdity of the math. The robotics division accounts for less than 5% of their total sales. We are essentially dealing with a 24-month regulatory purgatory attached to a business line that was never going to radically move the needle anyway.

I apologize for missing that and I am correcting that omission by taking out the stock. On a positive note, you can witness that I continuously review existing holdings, so in theory, I hope to catch mistakes and adjust them before they cause too much damage.

I will be sharing a page with subscribers that lists all trades since inception (four months) and where they stand, I am just finalizing the layout.

In the meantime, here is the updated basket:

Four months ago we drew a map of the rates market. It held where it mattered and failed where we warned it might. Today we close the book: one large win, two small losses.

“The Wrong Map” went out in March on one admission, we did not know which tail would arrive, so we built three legs that would pay across different states of the world. Looking back, thank god we did not venture in predicting how the Hormuz crisis would evolve!

The scorecard

The tanker win dwarfs the two rates losses, and the premiums were small by design, so the book closes comfortably ahead.

Right direction, wrong vehicle

The risk we identified in that issue was a painful middle with oil high enough to keep core inflation sticky, not high enough to break growth. That is exactly what happened: the Fed held at 3.50 to 3.75%, the 10-year sits near 4.6%, and the disruption that froze our rates legs is the one that detonated the tankers.

Then the June reaction complicated the trade even more. US core inflation printed flat on the month, the friendliest bond number of the year and Treasuries shrugged. When the perfect input produces no move, it’s clear you are missing something: the private sector spending (i.e. AI infrastructure spending) is stronger than anticipated, and our read on the near-term path is unlikely to resolve in the timeframe of these options. Two multi-year theses, wrapped in early-2027 options, are not going to fly. So we close both and step back, rather than roll a decaying option into a view we can no longer fund with conviction.

We are not done with US rates

We will likely express a view on US rates in the coming quarters. The scenario we fear is an AI-capex unwind: risk assets down, credit wider, the Fed cutting hard, capital stampeding into long Treasuries. Against that, a long bond hedges the entire portfolio. But if the hedge has a negative carry you can’t risk being way too early. The tells are not on the tape, and a hedge bought early is theta paid to wait, the exact bill we are settling today.

The trade moves to Britain

The surviving view goes where the mispricing is cleaner and needs no American recession, the front of the UK curve. And the case is not just that the Bank is mispriced. The economy underneath it is weak enough to force the cut.

Britain is not America here. Its mortgages do not lock away for thirty years; they refix every two to five. So the rate the Bank set in 2023 is still biting, household by household, as cheap pandemic fixes roll off and reset several points higher, a slow tightening still under way, long after the last hike. To a British homeowner the policy rate is not a screen abstraction; it is a bigger number on next month’s direct debit. It shows. Households have pushed their saving rate to nearly 9% of income, against about 6% before the pandemic with a straight subtraction from demand. And the labour market is loosening the way that leads disinflation, not lags it: vacancies a tenth below pre-pandemic, and pay, which trails the labour market by about a year, already down to 2.9%, its weakest since 2020. The second-round wage spiral the Bank keeps invoking is nowhere in the wage data.

Yet the market is braced the other way. Bank Rate is 3.75%; in June two members voted to hike and none to cut; Britain now carries the highest 10-year yield in the G7, on the weakest growth in it.

Subscribe now

The vehicle, and the risk in it

We are long the front of the gilt curve through the short-gilt future. It fixes the flaw that just cost us, no premium bleeding while we wait and hands us another: a leveraged future can be stopped out before the thesis pays. So we are starting with a small position and a stop loss.

The stop is a weekly close above 4.577%, the 2-year’s 2026 high, the line “The Wrong Map” has fought since March, barely 10 basis points higher than now. That stop is not very far and that’s what we like it because the front end is not the safe end of the curve. In 2022 the 2-year sold off hardest of all, and the Bank’s rescue bought long gilts, not short. We own the 2-year because it is the cleanest expression of a cut, not because it protects us from a fiscal scare. That risk we manage with a small sized position and the stop loss.

We started at half size, and may add after the Bank meets on 30 July. Half, because the case is not clean yet with core still stuck at 2.6%, and July’s 13% jump in the energy cap will lift headline over the next prints.

The shape is convex. Base case, the 2-year drifts to 3.75–4% as the Bank cuts late. Growth scare, it rallies hard as the hawkish curve unwinds. Core stays hot, the stop caps it. Gilts and sterling break together, the co-selloff pulls us out. Downside known; upside not.

• Our call: the UK front end is priced for a hike an economy this weak cannot sustain; the next Bank move is a cut, and 2-year gilts rally.

• Wrong if: core and services stay hot through the energy-cap pass-through and the Bank holds or hikes; or gilts and sterling sell off together, the fiscal tail the front end is most exposed to.

• Status: long UK 2-year via the short-gilt future, half size from 23rd of July; hard stop on a weekly close above 4.577%; may add after the 30th of July meeting.

References

[1] UK June CPI — Office for National Statistics: headline 2.6%, services 3.6%, core 2.6%. Private-sector regular pay growth 2.9% (3 months to May, lowest since 2020, ONS AWE); job vacancies 707,000, 10.4% below the pre-pandemic Jan–Mar 2020 level (ONS Vacancy Survey); household saving ratio 8.9% in Q1 2026 versus ~6% in 2019 (ONS series DGD8). Ofgem default tariff cap +13% to £1,862 from 1 July (Ofgem). UK mortgages predominantly short fixed-rate terms that refix every 2–5 years, versus the US 30-year fix; sub-3% pandemic fixes are rolling onto materially higher rates through 2028 (Bank of England, July 2026 Financial Stability Report; UK Finance).

[2] Bank of England — Bank Rate 3.75%; June MPC vote 7-2, two members preferring a hike, none a cut (Monetary Policy Summary). UK 10-year yield the highest in the G7 (sovereign yields, market data). OIS-implied policy path is market pricing, not a Bank forecast.

[3] The 2022 UK gilt-market episode, in which the 2-year yield rose furthest across the curve and Bank of England intervention purchased long-dated gilts (Bank of England; market data). For context, UK whole-economy investment was the lowest in the G7 at 18.6% of GDP in Q3 2025 (ONS).

[4] US rates context — FOMC target 3.50–3.75%; US 10-year ~4.6% (Federal Reserve H.15); June US core inflation flat on the month (BLS). Position marks and the tanker close are our tracked book; the tanker basket closed at roughly +30% total return in Issue 17 (22 June).

[5] “The Wrong Map,” Stray Narratives Issue 03 (23 March 2026) — the painful-middle call and the argument that a Gulf disruption is borne by Europe and Asia, not the US.

We are opening a position in a Dutch company that builds the largest man-made objects that move, parks them in the deep ocean, and then does the single most boring, most profitable thing in energy: it collects the rent. The company is SBM Offshore. It trades in Amsterdam, most professional investors have never opened its filings, and it is the closest thing the oil industry has to a toll booth — sold to us, right now, at the price of a cyclical.

Let us start with the machine, because the machine is the whole story and hardly anyone outside the industry has seen one.

Imagine an oil tanker. Now imagine that instead of a hold full of crude, its deck carries a chemical plant — separators, compressors, gas-treatment trains, a power station, accommodation for a couple of hundred people — the entire apparatus of an onshore refinery, welded onto a hull the length of three football pitches and moored in two kilometres of water a hundred miles from land. Pipes run down to wells on the seabed. The raw mixture of oil, gas and water comes up, gets separated on deck, the oil is stored in the hull, and every week or so a shuttle tanker pulls alongside and carries it away. It is called an FPSO — Floating Production, Storage and Offloading vessel — and it is how the world now produces oil in places too deep for a pipeline to shore: offshore Brazil, offshore Guyana, off West Africa. One of these things can cost three to five billion dollars and will sit on station, pumping, for twenty years.

Source: SBM Offshore

https://media.sbmoffshore.com/en/asset-details-page/3328:image:3dc3e75dafc34c08b263d44f8251a982#/search-id=1784027290282_107140893/

Here is the part that matters. The oil major — Petrobras, ExxonMobil — does not, as a rule, want to own the machine. It wants the oil. Owning a floating factory is a headache: you have to build it, finance it, crew it, maintain it, keep it running through hurricanes and equipment failures for two decades. So the major does what any sensible company does with an expensive, specialised asset it needs but would rather not run. It rents.

That is SBM’s business. It builds the FPSO, and then it leases and operates it for the field’s life — ten, fifteen, twenty years — for a contracted day-rate, paid whether the oil price is forty dollars or a hundred. The major is contractually obliged to pay for the machine to be available, in the same way you pay your mortgage whether or not you enjoyed living in the house that month. SBM is not selling oil. It is selling uptime. It is a landlord.

By User:WikiDon - From en: Ther uploaded with GFDL- licence by User:WikiDon, CC BY-SA 3.0,

https://commons.wikimedia.org/w/index.php?curid=696897

The magic trick everyone watches instead

The reason nobody talks about the landlord is that offshore energy is a genuinely thrilling casino, and the landlord is the one person in the building not gambling.

When investors say “offshore is back” — and it is; deepwater now breaks even around forty-three dollars a barrel, which is marginally cheaper than American shale at forty-five, a quiet reversal of the last decade’s received wisdom [6] — they reach for the exciting names. They buy the drillers: Transocean, Valaris, the companies that own the rigs that drill the wells. These are magnificent, terrifying instruments. A drillship is rented by the day, and the day-rate swings from a hundred and fifty thousand dollars at the bottom of the cycle to six hundred thousand at the top and back again, tracking the oil price like a needle on a seismograph. Own one at the right moment and you triple your money; own one at the wrong moment and you go through Chapter 11, which most of them did in 2020. It is a wonderful trade and a terrible annuity.

Or they buy the plumbers: TechnipFMC, Subsea7, Saipem — the engineering firms that lay the pipe and install the hardware on the seabed. Better businesses, real backlogs, and right now the market’s darling: TechnipFMC trades at roughly fifteen times EBITDA, near an all-time high, because it is the quality name in a hot sector. There is even a merger to gossip about — Subsea7 and Saipem are combining into a contractor with a backlog north of forty billion euros, antitrust regulators permitting, which is its own soap opera.

So the whole room is watching the drillers gamble and the plumbers merge. And in the corner, unwatched, is the one company that already signed the twenty-year lease and is simply collecting.

A toll, not a bet

This is the distinction the market is not pricing, and it is the entire idea: a day-rate is a bet on the oil cycle; a twenty-year lease is a toll on production. The driller gets re-priced every few months by the crude curve. SBM’s tenant is locked in until the 2040s at a rate agreed before the first barrel flowed.

Put a number on it. SBM’s order backlog — the contracted, already-signed future revenue — is $31 billion, with cash-flow visibility running to 2050. Eighty per cent of that backlog, and roughly seventy per cent of the profit, is the lease-and-operate book: the rent. Not the oil price. The rent. The margin on that rent is fifty-four per cent, because once the machine is built and paid for, keeping it running is cheap and the cheque arrives every month regardless.

And the obvious objection — that a twenty-year fixed cheque is not an annuity but a bond, quietly halved by two decades of inflation — is one the company has already answered. The backlog is, in SBM’s own words, “backed by firm contracts from premium clients with inflation protection” [5]. The rent escalates. The landlord thought of it first.

We have spent a good deal of this publication's ink on the difference between businesses paid on volume and businesses paid on price — the toll booth versus the commodity. SBM is the cleanest toll in the entire energy complex, and it is hiding in the one corner of the market everyone has decided is a pure commodity play.

“But look at the debt”

Here is where the professional investor, glancing at the screen for the first time, recoils. SBM carries around eight billion dollars of net debt on the IFRS measure, against a market value near six. For a company that gambles, that is a death sentence. It is the reason the stock screens as a scary levered cyclical and gets a scary levered-cyclical multiple.

It is also the wrong number — and this is the crux of the trade.

There are two debt figures here, and the gap between them is the entire argument. The IFRS measure consolidates every financed vessel in full. The company’s own preferred measure, Directional, tracks what the parent is actually on the hook for, and it stands at US$3.2 billion. Most of the difference is non-recourse project finance. It does not sit at the parent. It sits inside each individual floating factory, secured against — and repaid by — that specific lease. The bank lent against the contracted cash flow of one machine and can only ever look to that machine. The debt is a mortgage on a building, and the building comes with a tenant who has signed for twenty years and cannot leave.

You do not have to take our word for it, because SBM ran the experiment for us. In February it sold one FPSO — a unit called ONE GUYANA — to ExxonMobil Guyana for about US$2.32 billion in cash, and used the proceeds to repay, in full, the US$1.74 billion of project financing secured against that vessel [2]. Read that again: one machine, sold for a third more than the debt tied to it. The mortgage was not a millstone. It was comfortably covered by the asset it was lent against. Directional net debt is now US$3.2 billion, down forty-three per cent year on year, primarily on that sale [3]. And the landlord did not even lose the tenant — SBM continues to operate and maintain the vessel until 2035 [2]. It sold the building and kept the management contract.

A levered cyclical cannot do that. A portfolio of financed annuities can, because the leverage was never fragility. It was the plumbing of the toll.

The price

So what does the market charge for the cleanest annuity in energy, run by a company that just proved its debt is a mortgage and not a millstone?

The whole enterprise — equity and debt together — costs about five times forward EBITDA: earnings before interest, tax, depreciation and amortisation. For context, the plumbers you are being encouraged to buy instead trade at six to fifteen times the same measure, with more oil-price exposure and more project risk. On the crude price-to-earnings line the stock looks like value-trap wallpaper — roughly seven times — but the sharper figure is the cash it hands back. SBM returned $440 million to shareholders last year, up fifty-seven per cent, a mix of dividend and buyback worth about seven per cent of the market value, and it has committed to a minimum of $2.1 billion of returns over the next six years. You are paid roughly seven per cent a year, in cash, to wait — while a book of twenty-year leases quietly amortises the debt for you and a deleveraging balance sheet does the re-rating in the background. The analysts who cover it — the few who do — carry an average target around thirty-five per cent above today’s price.

Where we could be wrong

We try never to fall in love, so here is the case against, stated as strongly as we can make it.

First, SBM is a hybrid, not a pure landlord — and the impure half is not a half. The Turnkey division, which builds the machines, booked US$2.9 billion of first-quarter revenue against Lease & Operate’s US$0.6 billion. Turnkey is most of the revenue and little of the profit; the rent is little of the revenue and most of the profit. That asymmetry is the company — and it means the construction arm can hurt you long before the annuity does.

Second, this is a concentrated toll, not a diversified perpetuity. Strip the backlog down and it is overwhelmingly two counterparties: Petrobras in Brazil, and ExxonMobil in Guyana. A Brazilian political lurch or a Guyanese disruption would hit the “annuity” harder than the word annuity implies. Two tenants is not a bond ladder.

Third, the rent roll does not grow in a straight line. The backlog actually fell by four billion dollars last year, because new mega-contracts are lumpy — a handful of awards drive the whole book — and the pace of new leases still rides the deepwater sanctioning cycle. If oil sags into the fifties and the majors defer their next round of projects, the existing rent keeps arriving on schedule, but the growth stalls and the stock waits longer for its re-rating.

Notice, though, what none of those objections touch: the cash flow already contracted to 2050. The bear case is an argument about SBM’s growth and its construction arm. It is not an argument against the rent. You are being asked to underwrite whether the landlord finds new tenants, not whether the current ones pay — and while you wait for the answer, you collect seven per cent.

Subscribe now

The trade

We are adding SBM Offshore to our list of positions, and we are underwriting it over three years.

The shape of the return is unusually legible for an energy name. In the sour case, growth stalls and the multiple sits where it is, and we are left clipping a seven-per-cent yield for something in the low teens a year — a poor outcome that is still a positive one. In the base case, the backlog replenishes, the balance sheet keeps shedding debt as the leases amortise, the stock drifts toward the analysts’ thirty-five-per-cent-higher target, and the yield rides on top, for high teens to low twenties a year. In the good case, the market finally notices that a twenty-five-year contracted cash-flow stream is an infrastructure asset and not an oilfield-services stock, re-rates it toward the multiple that pipelines and toll roads command rather than the one that drillers get, and the number has a three in front of it — a year.

The market is paying fifteen times for the company that installs the plumbing, and five times for the company that owns the factory and collects the rent until 2050 — and it calls the second one risky because it appears to have debt, when the debt is simply the mortgage on a building that someone else is contractually obliged to pay off.

We will happily be the landlord.

References

[1] SBM Offshore — Full-Year 2025 Earnings and Annual Report. Directional backlog US$31.1bn with cash-flow visibility to 2050; Lease & Operate backlog US$25.0bn of US$31.1bn (~80%); Lease & Operate Directional EBITDA US$1,235m of US$1,709m total (~72%) on revenue of US$2,295m (~54% margin); the ~US$4bn year-on-year backlog decline; IFRS net debt US$8.1bn; Directional net debt US$5,651m at 31 December 2025; capital returns US$440m (+57%) and the US$2.1bn six-year minimum distribution commitment.

[2] SBM Offshore — FPSO ONE GUYANA purchase by ExxonMobil Guyana completed (4 February 2026). Total cash consideration of c. US$2.32 billion; the net cash proceeds primarily used for the full repayment of the US$1.74 billion project financing; SBM Offshore continues to operate and maintain the FPSO up to 2035.

[3] SBM Offshore — First Quarter 2026 Trading Update. Directional net debt US$3.2 billion at 1Q 2026, a 43% decrease year on year, primarily reflecting the ONE GUYANA sale.

[4] TechnipFMC, Subsea7 and Saipem — company filings and the Subsea7/Saipem merger announcement. Comparator EV/EBITDA multiples (~6–15×) and the combined entity’s backlog.

[5] SBM Offshore — First Quarter 2025 Trading Update: “Our pro-forma Directional backlog … is backed by firm contracts from premium clients with inflation protection.”

[6] Rystad Energy — published upstream breakeven analysis: deepwater breakeven ~US$43/bbl against North American shale ~US$45/bbl.

[7] EODHD. Share prices, forward EV/EBITDA (~5×) and P/E (~7×), and the sell-side consensus target (€43.20, ~+35%).

The most important question for the next five years for your portfolio has absolutely nothing to do with whether artificial intelligence is “real.”

AI is real. So were the railroads, the telegraph, and the fibre-optic cable. Each of them was still real on the morning after its bubble burst.

The question that will actually dictate the next decade of returns is much drier: What holds the American economy up when the spending on AI slows down? And the slightly more awkward follow-up: Does anything?

The United States is sitting in the final stage of a forty-year disinflationary bull market. The technology currently credited with extending that bull market is being dangerously misread. AI is not a new paradigm. It is a late-cycle efficiency input arriving at the maturity phase of a long technology cycle, not the dawn of a fresh one.

Sameer Singh’s reference class is exactly right: AI is the new microprocessor.

Here is the problem with microprocessors. An efficiency input makes the entire economy more productive, while its own profit margins are ruthlessly beaten to death in a commodity price war. The economic surplus flows downstream to the users, not to the companies supplying the hardware.

The rent does not disappear so much as relocate. It gathers at the one or two chokepoints no one can design around, the instruction set and the single foundry that can etch it, and drains away from the vast, undifferentiated capacity the hyperscalers are buying and renting back out.

A two-sided platform (like an App Store or a search engine) does the exact opposite. It traps users, extracts rent, and keeps its pricing power for a generation.

Right now, the market is paying the platform’s price for the microprocessor’s economics.

Stray Narratives has so far published individual investment ideas and a few intellectual articles on how we are understanding AI. This piece opens a series that steps back to the picture they sit inside: where the market stands in its long cycle, what this AI boom actually is, how it rhymes with the booms before it and where it turns worse, what lies underneath it, and what to do about it. The pieces that follow take it up one question at a time.

The 200-Year Loop

Every infrastructure build-out in the history of capitalism runs the exact same loop. Capital violently crowds into the “new thing.” The few companies who can afford the scale overbuild it. The financing strains. The bubble bursts, leaving a pile of stranded assets when the assumed cash flows mysteriously fail to appear.

Canals, railroads, rural electrification, fibre optics, and shale oil each completed this exact, predictable circuit.

To be a structural bull today, you have to make a spectacular assumption. You have to look at two hundred years of financial history and say with a straight face: “Yes, but this time the guys burning the cash are really smart.” You must believe that this AI build-out will be the very first to magically escape the loop, and that the firms currently spending the money are acting rationally against revenues that literally none of them can forecast.

Hyperscaler capital expenditure is now running above 2% of GDP. This is among the largest peacetime build-outs on record. But this specific cycle is uniquely dangerous in three ways:

The Illusion of Earnings: Silicon Valley has leaned ever harder on the ancient, magical art of depreciation accounting. The excess is hiding in reported earnings rather than the multiple, as hyperscalers mysteriously decide their hardware will last much longer than anyone thought. Meanwhile the revenue, though real, still covers only a fraction of the cumulative bill, which the BIS now puts above a trillion dollars across 2025 and 2026, and is increasingly financed with debt, and free cash flow is evaporating: at Amazon alone from a filed $26 billion to barely $1 billion in a single year. It is a spectacular business model if you don’t like money.

The Duration Mismatch: We are currently funding thirty-year data-centre debt to produce a deflationary token whose cost, for any given level of capability, drops 70% to 90% a year. The entire structure is running on chips with the effective lifespan of a flagship phone (18 to 36 months). And just to keep things interesting, a state-backed Chinese competitor is actively trying to drive the marginal cost of that token to absolute zero.

The Stranded Asset Trap: When the railroad bubble burst, it politely left behind tracks that built the modern economy. When this bubble bursts, the buildings and the power lines will find new owners at a fraction of their financed cost, while the returns that were meant to justify them simply never appear, and the loss lands where the money did, on repriced debt syndicated through the insurers and pension funds of middle America.

Subscribe now

The Economy Beneath the Bubble

The American economy that will eventually have to absorb this bust is, to put it politely, exhausted.

Once you strip out the AI capital spending, the foundation looks incredibly weak. Capex categories supplied roughly three-quarters of Q1 GDP growth. Beneath that capex sits a frozen labour market, housing rolling over, and a personal savings rate slashed to 3%, nearing its pre-crisis lows.

Consumption is holding up in large part because the equity bubble itself is sustaining a wealth effect for the top decile of earners.

One asset price is currently working triple shifts: it is flattering the index’s earnings, supplying the marginal GDP growth, and propping up the consumption beneath both. Aswath Damodaran’s caution about pricing versus value applies to this entire house of cards, not merely the semiconductor leaders.

Subscribe now

The Asymmetry

This is the asymmetry of the current market, and it is the reason we are laying this groundwork now rather than writing a witty autopsy later.

The market looks undeniably safe right now precisely because the thing inflating it is so fragile. The wealth effect, the earnings, and the growth all rest on a single price, with nothing independent beneath any of them. John Templeton famously noted that bull markets die on euphoria. That euphoria is always strongest at the point of maximum dependence on one single input.

We are positioning our capital for that specific shape, rather than trying to guess the exact timing. Because if the history of these cycles teaches us one thing, it is that the timing of a break is the one secret the market flatly refuses to tell you in advance.

Next in the series: before you can say what this cycle will do, you have to know where you are standing in it. The next piece takes two hundred years to answer.

Every desk has a version of the complaint: right on the thesis, wrong on the P&L. Uranium is currently turning that complaint into an exhibit. In April we put the case for owning physical uranium through the Sprott Physical Uranium Trust and added it to our list of trades at C$27.83 [1][8]. We set out three conditions that would show whether the thesis was actually working — conditions, not vibes. At the end of June the third one fired: the long-term contract price reached US$95.50 a pound, its highest since 2008, having risen in five of the last six months and fallen in none [2].

And yet. The units mark around C$27 — a shade below our entry — and the trust has slid from net asset value to an 8 per cent discount [7][8]. The metal is making new highs while the cheapest way to own it has gone on sale. That is the paradox, and it has a perfectly good cause: the marginal financial buyer has wandered off at precisely the moment the world’s governments are taking the metal off the table. Which is why we hold the position with more conviction than we opened it.

The Invisible Bull Market (by design)

In Issue 08 we drew the line between utilities signing contracts in volume at $90 and the spot price merely drifting there because nobody is signing anything [1]. The term price is where that distinction lives. It is assessed off live offers and settled contracts rather than screen trades, which is why it moves at the pace of a barge. The market, meanwhile, spends its days glued to a spot ticker on which almost no real pound trades, while the contracts that actually keep reactors fuelled are struck quietly, off screen, by people with no interest in entertaining anyone.*

The monthly series since December reads [2]:

• December: $86.50

• January: $89.00

• February: $90.00

• March and April: $91.50

• May: $94.00

• June: $95.50 (spot, for comparison, sits at $85.00)

Five rises, one flat month, nothing down. A price built from real, signed contracts does not do that on noise. So where does that leave our three conditions? The second — the run-down of Kazatomprom’s inventory buffer — is still advancing. The first, and the one that actually drives price, is still missing: Western utilities contracting at or above the replacement rate of roughly 180 million pounds a year [1]. Two of three, with the decisive one outstanding. For now.

Utility fuel buyers are many things; a studio audience is not one of them.

Sharpening the Maths

We owe readers two corrections to the numbers we published in April.

First, Issue 08 put Kazakhstan at 45 per cent of global primary supply. The World Nuclear Association’s 2024 data has it at 23,270 tonnes of a 60,213-tonne global total, which is 39 per cent [3]. Our figure sat above the primary record; this publication will cite 39 per cent from here.

Second, we described Kazatomprom’s guidance cuts as evidence it could not rescue the market. The cuts are real but relative. Kazakhstan’s absolute output rose 11 per cent in 2025, to 25,839 tonnes, and the 2026 guidance of 27,500 to 29,000 tonnes is a cut only against the levels its subsoil permits allow — not against what it actually produced [3]. Readers should have had both halves.

Neither correction changes the trade. The inventory buffer that matters still sits at about four months, and rising output in Kazakhstan is worth very little to a Western utility that is not permitted to buy it.

The Iron Curtain

Since April the pool of uranium the West will eventually have to bid for has shrunk, and not gently. Follow the Kazakh pounds.

To the East. In February Kazatomprom signed a long-term supply agreement with India’s Department of Atomic Energy worth roughly US$4bn; stack that on its contracts with China’s CNNC group and more than half the company’s booked asset value is now spoken for before a Western utility gets a look [4].

To Russia. Budenovskoye — Kazakhstan’s largest growth project — has 100 per cent of its output through 2026 reserved for Russia’s civil nuclear industry [6].

Into law. A code signed in December now requires Kazatomprom to hold more than 75 per cent of any new uranium joint venture, and 90 per cent or more of licence extensions, unless the foreign partner transfers conversion or enrichment technology [5]. Western developers can read: Laramide Resources terminated its Chu-Sarysu Basin option in January, on the view that an economic case for foreign investment in Kazakh uranium exploration no longer exists [5]. When a junior walks away from cheap optionality, believe the junior.

The Financialised Blind Spot

In April the Sprott trust traded at net asset value. It now sits at a 6.06 per cent discount, holding 81.4 million pounds of uranium oxide against US$7.05bn of net assets [7].

The flywheel we described in Issue 08 is real — and capped. Sprott can buy at most 9 million pounds a year in the spot market, and it bought 8.57 million in 2025, hard up against the ceiling [7]. When the units trade below NAV the trust stops issuing new ones, and its spot buying stops with them. The one leg of the thesis that used to run on autopilot has, for now, switched itself off.

That is the bearish reading, and it is true. The other reading is equally true: a 6 per cent discount means the units cost less than the uranium sitting inside them. The asymmetry we bought in April is on offer again — wider than it was at entry.

The position

A rising term price on thin volume is exactly how this market behaves when a handful of utilities move early, and the documented failure mode of the sector is the rest of the tape mistaking those early movers for noise, getting bored, and selling the week before the herd shows up [1]. We know that script. We decline the part.

Read together, the evidence points one way. The term price is climbing. The inventory buffer is draining. The East has first call on the supply. And the premier physical trust can be bought for less than the metal it holds. The one thing missing is the decisive one — Western contracting in volume. What has moved since April confirms the structure without yet delivering the outcome [8].

We hold the position, with more conviction than we bought it. The metal has heard the news; the units have not. That gap is the trade.

References

[1] Stray Narratives Issue 08, Is Now the Time to Invest in Nuclear?, published 27 April 2026 (straynarratives.com). The three conditions; the ~$90 long-term price at publication; the Sprott mechanism description; the trust at a +0.56% premium to NAV per [8] therein.

[2] UxC and TradeTech long-term price indicators as published by Cameco (cameco.com/invest/markets/uranium-price): month-end long-term price US$86.50 (Dec 2025), US$89.00 (Jan 2026), $90.00 (Feb), $91.50 (Mar), $91.50 (Apr), $94.00 (May), $95.50 (Jun); spot $85.00 (Jun). June level the highest since 2008.

[3] World Nuclear Association, world uranium mining production data: Kazakhstan 23,270 tU of 60,213 tU global in 2024 (39%). Kazatomprom 4Q 2025 Operations and Trading Update, 2 February 2026 (kazatomprom.kz): 2025 output 25,839 tU, up 11%, within guidance of 25.0 to 26.5 ktU; 2026 guidance 27,500 to 29,000 tU.

[4] Kazatomprom Investor Relations (kazatomprom.kz): long-term U3O8 supply agreement with India’s Department of Atomic Energy (Directorate of Purchase & Stores), announced February 2026, value exceeding 50% of Kazatomprom’s booked asset value (no dollar value disclosed; press estimates near US$4bn), approved by shareholders April 2026; prior CNNC/CNUC agreements, which together with earlier transactions exceeded 50% of booked asset value by November 2024 (single Nov 2024 contract ~KZT1.25tn, ~US$2.5bn, per Interfax); 2026 production guidance and inventory disclosures per the 4Q 2025 operations update and FY25 results.

[5] Kazatomprom announcement on amendments to the Subsoil Use Code, signed into law 26 December 2025: new uranium exploration/mining joint ventures require Kazatomprom above 75%; subsoil-use agreement extensions or production/reserve increases require Kazatomprom at 90% or above, or transfer of conversion/enrichment technology by the foreign partner. Laramide Resources announcement, January 2026, on exiting its Kazakh project (Chu-Sarysu Basin option, terminated 20 January 2026).

[6] Kazatomprom 4Q 2025 Operations and Trading Update (2 February 2026): 100% of Budenovskoye output for 2024 to 2026 contracted to Russia’s civil nuclear industry at market-related terms; Rosatom-linked ownership of 49% per Interfax.

[7] Sprott Asset Management, Physical Uranium Trust (sprott.com), 3 July 2026: NAV US$20.47 per unit, market price US$19.23, discount to NAV 6.06%, net assets US$7.05bn, holdings 81,447,348 lbs U3O8. Spot-purchase cap of 9 million lbs per year per the trust’s base shelf prospectus undertakings and Annual Information Form; calendar-2025 purchases 8.57 million lbs.

[8] U-UN.TO daily closes, EODHD: entry C$27.83 (24 April 2026); close C$27.35 (3 July 2026). Move from entry: minus 1.7%. [Verify-before-publish: re-quote at the open on publish day.]

In May I published the case for Xunlei (NASDAQ: XNET) at $6.33: a 78% discount to net asset value, anchored by a 7.8% stake in Insta360, the Shanghai-listed camera maker, worth two and a half times Xunlei’s own market capitalisation, and by a hard date, the 11 June expiry of the lock-up on that stake. The title promised four weeks to close. The date has passed, the shares mark $5.64, the position is down 10.9%, and the discount has not closed [1][6]. The position stays on our list of trades, on humbler terms than the ones I published.

The discount itself behaved: 78% at publication, roughly 75% today. What fell was the net asset value. Insta360 has dropped 36% since the May mark, from 214 yuan to 135.93 [2], and Xunlei’s NAV per share has fallen by about a fifth [7]. The trade lost money the one way the piece spent least time pricing: the catalyst arrived and damaged its own collateral.

What happened

The 11 June expiry freed Xunlei to sell, and it freed everyone else: 56.5% of Insta360’s share register became tradable that day, roughly 37 billion yuan of stock across 28 pre-IPO investors, of which Xunlei’s holding is one line among many [3]. Insta360 was already de-rating from its September peak of 377.77 yuan, a price war with DJI gave the de-rating a fundamental engine through June [4], and the unlock supplied the paper. The stock set a 52-week low of 130.88 yuan on 29 June and closed at 135.93 on 3 July [2]. The May piece named 130 yuan as the level that changes the arithmetic. The mark sits two bad sessions above it.

Management, handed the moment the thesis required, chose patience. On the first-quarter call, two weeks before the unlock, there was no dividend, no timetable for the stake, and one load-bearing adverb: Xunlei would “gradually seek to adjust” its holdings toward a stated ceiling of investment securities at 45% of total assets, excluding cash and government securities [5]. No sale of Insta360 shares has been disclosed since the lock-up lifted. What has been disclosed is a repurchase: on 26 June the board approved a programme to buy back up to $20m of Xunlei’s own shares over the twelve months from 1 July, funded from the cash balance [8].

What I got wrong

The timing failed on its own stated deadline. The body of the piece described an 18-to-24-month monetisation window, but the title said four weeks, and the calendar was wrong: the discount-narrowing flow I expected into the date never appeared, and the shares closed as low as $4.66 before the unlock even arrived [6]. (The title, in hindsight, was the most falsifiable sentence in the piece.)

I priced what the unlock permitted our company to do, and missed what it permitted the other 27 unlocked holders to do. The event that made Xunlei’s exit possible cut the value of the thing being exited by a third, a scenario the published piece did not carry.

The catalyst rested on willingness. The Investment Company Act gives this management a direction and no timetable. The lesson filed in our mistake log: a catalyst date earns the name only when someone is obliged to act on it.

The published return arithmetic inherited the flaw. A base case of 55% to 75% a year was that willingness with a spreadsheet attached; the scenario that assumed the least of management, the bear case, is the one the market delivered.

The position, remarked

On the piece’s own methodology, at the 3 July close [7]: the Insta360 stake, 31.28m shares at 135.93 yuan and 6.77 to the dollar, is worth $628m, against $986m at publication. Net cash is roughly $234m, slightly higher than published. The operating business keeps its trough range of $410m to $745m, midpoint $580m: first-quarter revenue grew 54.1% and subscriptions 26.2%, though management guided to a modest slowdown ahead [5]. Net asset value is therefore about $1.44bn, or $22.65 per ADS, against $5.64 on the screen. The discount is 75%.

Two published facts survive the markdown. Net cash plus the trough value of the operating business comes to about $814m, 2.3 times the $359m market capitalisation, so the camera stake still comes free even marked a third below where I first called it free. And the regulatory pressure has not moved: at current marks investment securities sit near 74% of the relevant asset base against management’s stated 45% ceiling, and the sale required to reach that ceiling is roughly $446m, one and a quarter times the market value of the entire company [5][7].

Where I stand

The trade I published is dead, and pretending the horizon was always longer would be worse than the loss. What remains is a different position: a deep-discount holding with no clock, worth keeping only while that arithmetic holds, and re-examined in public whenever it is tested.

The case for closing is real: book NAV fell 25% in the two quarters to March, Insta360 is fighting a price war at its 52-week low, management’s pace is glacial, and wide discounts persist for years when nothing forces them shut. The case for holding is that at $5.64 the market pays less than the cash and the subscription business alone, the supply event has now happened rather than loomed, Insta360 sits at the May piece’s own bear-case mark, and the Investment Company Act still points in one direction. Selling at the bear-case mark would mean selling the cash and the subscription business below their own worth to be rid of a stake the price already writes off. The honest cost of staying is that the unlock supply has not cleared; it is now permanent.

I am holding, on those re-underwritten terms, with one mechanical exit: a next capital-allocation move of size that is an acquisition rather than a sale or a distribution. A decisive break below 130 yuan on Insta360 is a tripwire of a different kind. It sends the position back to the workbench for a fresh re-underwrite, published here, rather than forcing a sale on its own; a lower mark changes the arithmetic, and it is the arithmetic that decides. The first disclosed sale of Insta360 shares is what upgrades this from a discount we hope closes to one being closed. The watch items posted under the article in June stand, and the first has partially fired: the $20m repurchase is a real, if modest, step toward returning capital, while the pace of the 45% unwind and the next mark on the stake remain open.

References

[1] Stray Narratives Issue 11, An 80% Discount to NAV, Four Weeks to Close, published 14 May 2026 (straynarratives.com/p/an-80-discount-to-nav-four-weeks). Entry $6.33; NAV $28.57 per ADS on 62.86m ADS; net cash $230m; Insta360 stake $986m at the published 13 May mark of 214 yuan; operating business $410m to $745m at trough multiples, midpoint $580m.

[2] Arashi Vision Inc. (Insta360, SHA: 688775) closed at 135.93 yuan on 3 July 2026; 52-week low 130.88 yuan set 29 June 2026; 52-week high 377.77 yuan (September 2025). Stock Analysis and FT market data. Issue 11 published mark: 214 yuan, 13 May 2026, the basis on which the published $986m stake value was computed; the official 13 May close was 220.44 yuan.

[3] Pandaily, Insta360 at a Crossroads: 37 Billion Yuan Lock-Up Expiry Collides with DJI Rivalry, June 2026. 56.5% of shares, worth approximately 37 billion yuan, tradable from 11 June 2026; 28 pre-IPO financial investors; Xunlei at 7.84%.

[4] Caixin Global, DJI, Insta360 Lock Horns in Camera Pricing Standoff, 4 June 2026.

[5] Xunlei Limited, Unaudited Financial Results for the First Quarter Ended March 31, 2026, Form 6-K Exhibit 99.1, filed 28 May 2026, and the accompanying earnings call. Revenue +54.1% year on year to $98.6m; subscriptions +26.2%; cash, equivalents and short-term investments $303.6m; long-term investments $888.6m; total assets $1,415m. Management commentary on holding investment securities below 45% of total assets, excluding cash and government securities, and on adjusting holdings “gradually”; US Investment Company Act of 1940, s3(a)(1)(C), sets the statutory 40% test.

[6] XNET.US daily closes, EODHD: $6.33 on 13 May 2026 (entry); low close $4.66 on 5 June; $5.64 on 2 July 2026, the last session before the US holiday. Move from entry: minus 10.9%.

[7] Refreshed NAV on the Issue 11 methodology: 31.28m Arashi Vision shares × 135.93 yuan ÷ 6.7702 USDCNY = $628m. Net cash $234m ($303.6m cash, equivalents and short-term investments less $69.6m of bank borrowings, Q1 2026 balance sheet). Operating business at the published trough midpoint, $580m. NAV $1,442m ÷ 63.65m ADSs = $22.65 per ADS; market capitalisation $359m at $5.64 on 63.65m ADSs (318,268,921 shares outstanding at 31 March 2026; one ADS is five shares). Investment-securities ratio: $628m ÷ ($1,154m remarked total assets, the Q1 total of $1,415m with the $888.6m long-term investments line remarked to $628m, less $304m cash items) ≈ 74%; sale required to reach the 45% ceiling ≈ $446m.

[8] Xunlei Limited, 2026 Share Repurchase Program, Form 6-K filed 26 June 2026: up to US$20m of ADSs or common shares over the twelve months from 1 July 2026, funded from the cash balance.

If you prefer to listen rather than read, here is a podcast interview version of this same subject:

The American medical-device sector trades at its smallest premium to the wider market since the 2008 financial crisis, and on next year’s earnings it has slipped to an outright discount for the first time in living memory [1]. This is a group of companies that grew revenue from roughly seventy billion dollars to a hundred and seventy-six billion over the past decade, earns operating margins near twenty percent, and sells into the most dependable demand curve in the economy, which is the ageing of the rich world [8]. We are adding a basket of its cheapest quality names, equal-weighted, to the list of trades, because the market has priced a collapse that the businesses themselves refuse to deliver.

Two fears did the damage. The first is that weight-loss drugs will hollow out demand for everything from artificial knees to heart valves to glucose monitors. The second is that the sector’s profitability has quietly broken. Both are checkable against the operating numbers, and the numbers disagree with the price: the drugs are expanding several of these markets rather than shrinking them, and the margin dip everyone extrapolated was a post-pandemic inventory hangover that has already reversed.

How a good business got cheap

The de-rating is real and worth respecting. Large-cap medical devices spent most of the past decade carrying a four-to-six turn premium to the S&P 500 on forward earnings, the going rate for steady growth and recession resistance. That premium has compressed to roughly nothing, and the median name now trades near fifteen times forward earnings against an index above twenty-two [1]. The last time the sector was this cheap relative to the market, the world was busy recapitalising its banks.

Two things drove it. Through the pandemic, hospitals over-ordered devices and consumables, then spent 2022 and 2023 working that inventory back down, so reported revenue growth stalled to under two percent before recovering [2]. And from 2023 onward, every earnings call carried the same question, in one form or another: what does Ozempic do to your volumes. The market has spent two years concluding that weight-loss drugs will empty the operating theatres, and has priced the sector accordingly, which would be the correct response if the operating theatres were emptying.

The engine did not break

Cheap is only interesting if the business behind it is sound, and the ten-year record says it is. Set the pandemic distortions aside and the profitability picture is a cyclical dip that has already recovered: operating and free-cash-flow margins now sit at or above where they were a decade ago, EBITDA margins are at a ten-year high near twenty-six percent, cash conversion has improved, and research spending has held steady at about seven and a half percent of revenue the entire time [2]. The one genuine soft spot is gross margin, which has slipped two to three points over the decade on mix and input costs, an erosion the sector has more than recovered further down the income statement. A business that keeps investing the same share of revenue in its own future while its cash margins climb has not weakened. It has stopped commanding the market’s attention.

What the weight-loss trade got wrong

The GLP-1 fear deserves a real answer rather than a wave of the hand, because the logic is not absurd. Thinner, healthier people should, in time, need fewer of the interventions that obesity drives. The difficulty for the bears is that the operating data has begun to arrive, and for most of the sector it points the other way.

Sleep apnoea is the purest expression of the bear case, and its leading device-maker was punished hardest. The company ran its own analysis across roughly six hundred and sixty thousand patients and found that those prescribed a weight-loss drug were about eleven percentage points more likely to begin therapy, and its device volumes grew double digits straight through the scare [3]. The clinical trial that was supposed to prove the drugs cure apnoea instead found that most patients still had disease serious enough to treat [4]. The drugs are pulling people into the system, getting them diagnosed, and leaving the majority still needing the device.

The pattern repeats across the basket. Continuous glucose monitors were meant to lose their diabetic customers; instead more than two in five users of the leading monitor are now also taking a weight-loss drug and using the sensor to manage the transition [3]. Orthopaedics was meant to lose its knee and hip replacements; in practice surgeons routinely require patients to lose weight before they will operate, so a drug that reliably reduces body mass widens the pool of people eligible for surgery.

There is one place the bears are right. Bariatric surgery, the operations designed specifically to treat obesity, faces a genuine and management-acknowledged decline as the drugs substitute for the scalpel. For the diversified surgical-robotics franchise in our basket that is a small and recoverable share of a far larger procedure base, but it is the one unambiguous loss in the group, and pretending otherwise would be the sort of analysis that gets quoted back to you in eighteen months.

The asset nobody is pricing

There is a longer-dated reason to own these companies that the multiple ignores completely. The device-makers have spent two decades accumulating the one thing the artificial-intelligence boom has made suddenly precious: vast, proprietary, regulated clinical datasets that cannot be scraped from the open web or generated synthetically. The surgical-robotics leader holds the logged record of more than twenty million procedures. The connected-breathing-device maker has billions of nights of physiological data with the contractual right to use it. The diabetes and diagnostics franchises hold continuous, outcome-linked readings at a scale no model-builder can assemble, because a regulator will not accept invented data and a hospital cannot hand its patients’ records to a chatbot. Several of these firms already sell medical AI products the regulator has cleared, built on exactly this data [5]. A market paying extraordinary multiples for the companies that own AI models is paying nothing for the companies that own the one input those models cannot manufacture. We treat this as the free option on top of the value rather than the reason to buy.

The moat has one genuine challenger, and it is not a model-builder. Chinese device-makers are taking share in the high-volume, lower-margin corner of the industry, and they compete on price rather than on data. State procurement has cut the price of routine in-vitro diagnostic tests by as much as three-quarters, and China’s domestic champion in patient monitoring and bench diagnostics is moving upstream into the reagents and instruments Western firms long treated as captive; in gene sequencing a Chinese pair has already overtaken the former Western leader inside China [6]. The pressure falls hardest on the commodity-diagnostics and imaging lines, where the product is a consumable sold to a price-driven state buyer, and lightest on the franchises whose advantage is a regulated, outcome-linked dataset locked inside an installed base: a robotic-surgery record, years of physiological monitoring, a continuous glucose feed. The honest split is that the diagnostics and imaging names in the basket carry a structural margin headwind that quality does not fully offset, while the device-data franchises that anchor the thesis are the ones the substitution reaches last. It argues for leaning on the moat names inside the basket, not against owning it.

Why it re-rates

Cheap and sound can stay cheap for a long time, so the question is what closes the gap. The most concrete answer is that the buyers with the most information and the most capital are already taking these assets off the market outright. 2025 was the busiest year for medical-device dealmaking in more than a decade, including a private-equity take-private of a diagnostics company at roughly eighteen billion dollars and a twenty-one billion dollar acquisition by one of our own basket members [7]. When trade buyers and private equity pay cash for whole companies above where the public market marks them, they set a floor under the equity that sentiment struggles to break through. The slower catalyst is the ordinary one: as the inventory hangover fully laps and earnings revisions turn up, a sector at a discount to the market with a thirty-year demographic tailwind tends to be repriced by the same flows that left it. Those flows are only now beginning to reach devices: the group has turned up sharply off its June low in the past week, led by the diabetes and diagnostics names, though a few days off a six-month bottom is not yet a trend, and that hesitation is the honest timing risk in the trade.

Subscribe now

The position

We are adding an equal-weighted basket of ten names to the list of trades at the open: ResMed, Medtronic, Abbott, GE HealthCare, Boston Scientific, Stryker, Zimmer Biomet, Becton Dickinson, Globus Medical and Dexcom. Six are quality compounders growing earnings and cash flow into the demographic tailwind; three are deep-value or special-situation names where the cheapness and a catalyst do the work; one, Dexcom, is the clearest single embodiment of the thesis, a proprietary-data franchise that the weight-loss drugs turn out to help. Equal-weighted, the basket trades near sixteen times forward earnings, well below both the market and the cap-weighted sector index, which is dragged up by paying thirty to forty times for the two most expensive names in the group. We hold the data moat through its cheaper carriers rather than overpay for the marquee one.

For readers who would rather own the theme in a single line, the iShares U.S. Medical Devices ETF (ticker IHI) delivers the same exposure, with the caveat that its cap weighting makes you pay up for exactly the expensive names the basket is built to avoid. The basket is the better expression of the view, the ETF the more convenient one. A two-line case for each holding sits at the foot of this note.

Stray Narratives is free and carries no advertising, so it travels only as far as its readers send it. If this was worth your time, a like, a restack, or a forward to one colleague who should see it is the whole of our distribution, and what keeps it free.

The position implication

You are buying a sector at a discount to the market for the first time since 2008, with margins at decade highs, a demand curve that compounds for thirty years, acquirers setting a cash floor beneath it, and a proprietary-data moat the AI trade is pricing at zero. What that pays, on a three-year view and entered at the basket’s sixteen-times multiple, is a base case of roughly seventeen percent a year: about half the discount to the market closes, earnings compound near nine percent, and the dividend adds a point. If the discount shuts to parity and earnings run faster, the basket returns about twenty-six percent a year. If it never closes at all and earnings only grind, the basket still returns about six, because the growth and the yield carry it without the re-rating. The re-rating is the option you are handed for free.

That asymmetry is the reason to own it now: the discount exists because the market paid in advance for a catastrophe, and the catastrophe is being refuted in the operating numbers one quarter at a time. The repricing does not require the businesses to surprise anyone. It only requires the absence of the disaster that is already in the price.

References

[1] Forward and historical relative-valuation work on US large-cap medical devices (median next-twelve-month P/E versus the S&P 500 and the healthcare sector), from a sell-side equity research desk, May 2026; cross-checked against a major asset manager’s published note placing healthcare at its lowest relative valuation since 2008. Our own basket multiples are computed from company filings.

[2] Stray Narratives analysis of the ten largest US medical-device companies’ annual filings, 2015 to 2025 (revenue, gross/operating/EBITDA/free-cash-flow margins, research and development, return on equity), aggregated sum-of-parts.

[3] ResMed published analyses of GLP-1 prescription data and CPAP initiation/adherence (IQVIA claims dataset, ~660,000 patients), and company device-volume disclosures, 2025-2026; continuous-glucose-monitor co-prescription figures from the leading manufacturer’s disclosures.

[4] Eli Lilly SURMOUNT-OSA trial results and the FDA approval of tirzepatide for obstructive sleep apnoea, December 2024.

[5] FDA clearances of medical AI/machine-learning device software across the named manufacturers (FDA AI/ML-enabled device list, public).

[6] Chinese medtech competition and pricing: China’s centralised “volume-based procurement” of in-vitro diagnostics (provincial tender data via the National Healthcare Security Administration), the domestic patient-monitoring and IVD leader’s vertical-integration into reagents/raw materials, and the displacement of the incumbent gene-sequencing supplier inside China by a domestic pair (company filings and trade-press deal data; surfaced by a sell-side note, underlying primaries cited).

[7] Medical-device M&A activity, 2025: aggregate deal value and the take-private of a diagnostics company and a large strategic acquisition cited as illustrative (public deal announcements).

[8] US Census Bureau and UN Department of Economic and Social Affairs population projections (US 65+ population toward 71 million by 2030; global 65+ toward ~1.5 billion by 2050).

The basket, name by name

• ResMed (RMD). The cleanest rebuttal to the GLP-1 bear: its own claims data shows patients on weight-loss drugs are more likely to start CPAP, not less, while device volumes still grew double digits. At 16x forward earnings with operating margins above 30 percent, the market is paying late-cycle prices for a structural grower.

• Medtronic (MDT). The lowest-multiple way to own the data moat: several FDA-cleared AI products already run on its proprietary device data, and it converts roughly five billion dollars of free cash flow a year. At 13x forward earnings it is priced as a no-growth utility while revenue has returned to mid-single-digit growth.

• Abbott (ABT). The diversified ballast, spanning glucose monitors, diagnostics and devices, with free cash flow up roughly half over three years to more than seven billion dollars. Its Libre franchise is a GLP-1 beneficiary, as drug users adopt monitoring rather than abandon it.

• GE HealthCare (GEHC). The cheapest large-cap imaging franchise at 13x forward earnings, and the most prolific filer of FDA-cleared AI in the group. Growth is slow, so the case rests on the re-rating of a misjudged spin-off rather than on accelerating fundamentals.

• Boston Scientific (BSX). The Farapulse cardiac-ablation cycle is still driving low-double-digit revenue growth and steady share gains, with free cash flow roughly doubled over three years. At around 13x forward earnings it is the cheapest of the growth names, partly because the stock has sold off sharply in recent weeks on a flagged stagnation in its Watchman left-atrial-appendage line, a setback in a different franchise from the Farapulse ablation business that underpins the growth case.

• Stryker (SYK). Orthopaedic robotics with a genuine data moat: its Mako system has logged more than a million cases a rival cannot replicate, and weight-loss drugs help by clearing the BMI thresholds that gate joint replacement. The 21x multiple is the one premium we pay for quality and moat together.

• Zimmer Biomet (ZBH). The deepest value at under 11x forward earnings, a mean-reversion bet on an orthopaedics recovery rather than a growth story. The same GLP-1 dynamic that worries the bears makes more patients eligible for the hips and knees Zimmer sells.

• Becton Dickinson (BDX). The special situation: 11x forward earnings against a company pursuing a separation that could surface value the conglomerate structure hides. It also carries the most debt in the basket, so this is the position underwritten on the catalyst rather than the margins.

• Globus Medical (GMED). The quality mid-cap: a net-cash balance sheet, a double-digit return on equity, and operating margins climbing back toward the high teens as the NuVasive merger integrates. Small enough to stay a credible acquisition target in a consolidating sector.

• Dexcom (DXCM). The thesis in a single name: a proprietary glucose-data moat, a verified GLP-1 tailwind as drug users adopt monitoring, and free cash flow more than doubled to over a billion dollars. The 27x multiple is the basket’s richest, and the risk to watch is the price war with Abbott, not the drugs.

A crude tanker spends most of its working life earning a return that would embarrass a savings account. Then, for a few months every decade, a chokepoint makes the evening news, crude reroutes the long way round, and the holder books years of profits in a single quarter.

That window is the trade I put on in March. I called it the Wrong Map: a basket of three tanker equities, DHT Holdings, Frontline and International Seaways, equal-weighted, bought as a direct wager on Middle Eastern disruption. I closed it this week with a nice profit:

The reason to own these names was the disruption, and the disruption is resolving. The United States and Iran have signed a memorandum of understanding, the Strait of Hormuz is reopening, and a reopened strait is, with a lag, poor news for ton-miles: crude takes the direct route again, voyages shorten, idle capacity returns, and freight reverts toward breakeven. The war premium I was renting is being handed back.

The arithmetic

When conflict in the Gulf forces crude onto longer routes, ton-miles, the volume of oil multiplied by the distance it travels, rise, and tanker spot rates rise with them. At its early-March peak the Baltic benchmark for the Gulf-to-China route reached roughly $424,000 a day against an operating breakeven near $22,000, the spread that produces the best quarter of a shipping cycle and the record dividends that came with it. It had eased toward $100,000 by mid-June, and the congestion of the reopening keeps it bid for now, but the direction is no longer higher.

Why I am not holding out for the last dollar

Tanker equities have the awkward property of looking cheapest at the exact moment they are about to earn the least. They are priced off the prevailing spot rate, so at the crest of a disruption their trailing earnings look spectacular and the shares screen as the cheapest assets on the market, a multiple usually quoted without the word “trailing” attached. That multiple is a mirage built from a single, unrepeatable quarter, and when freight reverts the earnings behind it leave with the peace. I would rather hand the optically cheap multiple to whoever wants to own the normalisation than sell it to them a month later at a lower price.

There is a genuine floor under freight. Sanctioned Russian barrels, an expanding shadow fleet and structurally longer average hauls should keep ton-miles above the pre-war normal even after the queue clears, which is why the exit may be too early. But a floor a few thousand dollars above breakeven is a rounding error against a rate that spiked to nearly twenty times breakeven. It argues for owning these names at a normalised rate on a quiet day, not for riding them through the risk of a round trip.

Where I stand

I have now closed the Wrong Map tanker basket and all three legs finished in the black, including DHT, which spent most of the trade as the laggard before the reopening congestion firmed rates into June and pulled it level with the others. The other leg of the Wrong Map, the position in rates, runs on a different mechanism and a much longer clock, and it stays on; I will however restructure that trade in the coming weeks to make sure it embeds enough time horizon to materialise.

Another free article, so please like & share if you enjoy it!

In the Central Valley of California, farmers are ripping more than 210,000 acres of almond trees out of the dirt.

They aren’t doing this to rotate crops. They are doing it because the maths of the last decade has completely broken. The University of California, Davis recently ran the numbers on planting a new almond orchard: factoring in land, trees, water, ten years of capex, and debt service, a debt-financed grower now needs roughly twenty years to break even at today’s price deck near US$2.70 a pound, and about twelve years even with no land debt to service [1].

Nobody invests with a twenty-year breakeven. Twenty years is not an investment horizon; it is a sentence. The margin of safety is gone, marginal growers are walking away, and the supply response that historically kills almond bull markets is dead.

An almond tree bears its first crop in year three, reaches full yield around year seven, and produces for roughly twenty-five years before the orchard is replaced. Supply is set the better part of a decade in advance, and it cuts both ways: the planting that should feed 2030 is not going into the ground, and the trees being pulled now will not come back at these economics.

But while California capitulates, the cycle break has opened an asymmetric mispricing 12,000 kilometres away.

I am adding Select Harvests (ASX:SHV) as a trade at the open, at roughly A$3.93. The trade is the global almond cycle break, expressed through the lowest-cost producer outside California, with a revenue tailwind from US-China tariff arbitrage and a share price trading at a steep discount to its hard assets.

Here is why the cycle is broken — and why the downside on this trade is bound by physics rather than by market sentiment.

The Ultimate Capitulation Signal

When an industry quietly stops publishing the data that historically hurt its prices, that is not a methodology change, it is a confession. It is the agricultural equivalent of a struggling tech company quietly retiring quarterly guidance.

In December 2025, the Almond Board of California voted to stop funding the USDA Objective Measurement Report — the rigorous July estimate, built on physical orchard sampling, that historically corrected the looser May grower survey [1]. The timing gives away the motive. The vote came months after the July 2025 estimate landed high and triggered the single largest one-day almond price drop in modern memory [1]. Confronted with the one number that reliably delivered downside surprises, the industry chose to stop publishing it rather than argue with it. Officially, the reason was accuracy. Read it however you like — but an industry quietly euthanising the data that keeps embarrassing its prices is not the behaviour of people who think supply is about to flood.

The data backs the tell. California controls 80% of global almond production, so what happens in the Central Valley dictates the cycle [2]. Three things are happening at once:

• Removals are unprecedented. Land IQ data show more than 210,000 acres — roughly 15% of bearing acreage — pulled since 2022, including about 67,000 in 2024 alone. In 2026, California’s bearing almond acreage fell for the first time since 1995 [2].

• The pipeline has collapsed. Non-bearing acreage, the leading indicator for production three-to-seven years out, fell 26% year-on-year. The supply that should feed the 2028–2032 window simply will not exist [2].

• The marginal acre is uneconomic. Westlands Water District, historically the largest destination for new plantings, planted essentially zero new acres in 2023 and 2024, as water costs sat structurally elevated under California’s Sustainable Groundwater Management Act (SGMA) [2].

Combine that with Australian output running broadly flat toward 2030 as growers replant rather than expand, and structural drought across Spain’s growing regions, and the global supply curve is locked. Meanwhile demand keeps climbing — on tariff-driven Chinese buying, Indian consumption, and the developed world’s determination to put almonds in everything that once contained milk.

Capital Allocation: Utility Over Signalling

The market is pricing SHV on cyclical fear, assuming we are near a top. Management is quietly arbitraging that behavioural error with a textbook capital-return playbook [3].

After a strong 1H FY2026 result, the board announced a 10% on-market buyback — around 14.2 million shares, 10% of the issued capital — and resumed the interim dividend. It also secured a new A$60 million debt facility on favourable terms, not for distressed working capital, but to shrink the float at a depressed valuation [3].

This is the unglamorous version of smart capital allocation: financial utility over market signalling. Buying a dollar of productive almond orchard for eighty cents is not a sophisticated manoeuvre — it is just an unfashionable one. And the mechanics are cycle-proof: a 10% reduction in the share count lifts earnings per share by roughly 11% on a flat earnings base, entirely independent of where almond prices go next.

The Asset-Backing Floor

Replacement-cost analysis is the discipline that stops a cyclical equity from becoming a cyclical trade. On a sum-of-the-parts replacement basis [3][4]:

That is an enterprise value of A$557–803 million. Strip out net debt of roughly A$80 million — down 51% to A$79 million at the last full year — and, across 142 million shares, the implied equity sits at A$3.35–5.10 per share [3][4].

At around A$4.00, the market is paying for the assets at below the midpoint of their own replacement value, with zero operating premium. You get the largest Australian almond producer — A$118–130 million of FY26 base-case EBITDA — thrown in for free on top of the dirt [3].

Unless the orchards literally burn down, the hard assets bind the downside around A$3.30–3.50. Below that, the assets are worth more than the equity, which makes a private buyout increasingly plausible — the kind of valuation that tends to attract people with cheque books and patience. Precedent: Costa Group went private at A$3.20 in 2024 on exactly this sub-replacement logic [4].

Reality Check: The Bear Cases

A thesis is only as strong as its weakest link. Here is what breaks this trade, in order of probability.

1. A Trump-China tariff thaw (~12–15%). Australian almonds have been winning share in China on the US-China tariff differential. A partial agricultural deal in late 2026 or 2027 narrows that gap and reroutes Chinese buyers back toward US origin. Impact: a 15–20% cut to the realised-price assumption.

2. A Murray-Darling water shock (~8–12%). This is the hidden leverage in the model. SHV owns the permanent rights to only about 18% of the water it needs and sources the rest from the allocation market — which is to say, from the spot market and the goodwill of the sky. Under El Niño conditions, Australian allocation prices spike viciously (A$300+/ML) [4]. In a concurrent drought, global almond prices rise but SHV’s unit margins compress as spot water costs bite. Impact:15–25% off EBITDA in the affected year — though the hard-asset floor still protects the equity base.

3. Tail shocks (~3–5%). August-September frost, or a resurgence of the varroa mite that hit Australian hives in 2022. Impact: a 20–30% drawdown.

The unifying point: in every realistic bear case, the asset floor binds. Outside the tail scenario, SHV does not go meaningfully below ~A$3.30 without the assets being worth more than the equity. The downside is capped by physics, not by sentiment.

Position and Expected Return

Adding SHV to the list at the open, ~A$3.93, for a 12-to-18-month window — with optionality to extend into the 2028–2030 production gap if the supply-destruction story fully materialises.

*Fair values recomputed on the corrected 142.1m share count (the 1H result confirmed 142.1m, not the 152m the original model used). Upside rows hold a ~7× EV/EBITDA backed out of the base case; bear and tail rows are anchored to the ~A$3.30–3.50 asset floor. These are total returns over the 12–18 month horizon, not annualised. Probability-weighted, the trade carries roughly +23% to +33%, with the downside floored by the assets — a sharper asymmetry than the original model showed, because fewer shares lift every per-share figure and the buyback is retiring stock below replacement value.*

What would make me add: California’s 2026 crop confirmed at ≤2.65 billion pounds; Murray-Darling allocation easing below A$120/ML; FY26 EBITDA tracking above A$130 million. What would make me trim: a US-China deal that narrows the tariff differential; spot water sustained above A$250/ML for two quarters; FY26 EBITDA tracking below A$95 million.

Bottom line. The cycle is not at its peak. The cycle has broken. New orchards do not get planted at twenty-year breakevens, the pipeline has collapsed, and the industry has stopped publishing the number that used to keep it honest. Select Harvests is the cleanest way to own that break: capped downside bound by physical assets, asymmetric upside, and a management team compounding capital while everyone else is reading the cycle backwards. There is, occasionally, a free lunch in markets. It is usually buried under something nobody wants to look at — in this case, more than 210,000 acres of dead trees.

Subscribe now

References

[1] University of California, Davis — 2024 Almond Cost Study (Sacramento Valley & Southern San Joaquin editions), for the new-orchard breakeven horizon; Almond Board of California position reports, including the December 2025 vote to discontinue funding the USDA Objective Measurement Report.

[2] Land IQ Standing Almond Acreage estimates (commissioned by the Almond Board of California; the USDA NASS standalone California almond acreage survey was discontinued), 2022–2026 — bearing removals (more than 210,000 acres / ~15% since 2022; first bearing-acreage decline since 1995, 1,401,097 → 1,385,870), the ~26% drop in non-bearing acreage to the 2025 final (142,306 → 104,900), and Westlands Water District / Land IQ planting data; California ≈ 80% of global production (International Nut & Dried Fruit Council). (SGMA — California’s Sustainable Groundwater Management Act — is the regime behind the elevated marginal-acre water costs.)

[3] Select Harvests 1H FY2026 results announcement and presentation, 28 May 2026 — FY26 EBITDA guidance, the 10% on-market buyback, A$60m debt facility, resumed interim dividend, and orchard / processing / water-entitlement asset disclosures.

[4] Asset-floor anchors — Australian agricultural land and water transaction benchmarks (incl. Murray-Darling Basin entitlement and allocation pricing) and the Costa Group / Paine Schwartz Partners take-private at A$3.20 (completed early 2024; asset-backed precedent).

When a technology concentrates wealth this completely in the hands of capital owners, the political system reacts, as it has at every comparable point in modern history. AI does not ask for stock options, does not unionise, and does not complain about the return-to-office mandate, which is precisely why the reaction will not come from inside the companies. The previous issues in this series traced where economic value migrates as the cost of cognition collapses: toward what AI cannot make — the human element, physical limits, and verification capacity.

The central investable question of the next decade is what the political system will do to the winners.

Morrow’s Diagnosis

In the years around the Great Financial Crisis, Robertson Morrow, a macro thinker and former CIO at Peter Thiel’s Clarium Capital, wrote a paper titled “The Bull Market in Politics” — the paper this issue takes its name from [2]. His argument was simple: going forward, government influence over trade, immigration, social policy, and the decisions of war and peace would matter more to investors than business fundamentals ever would.

He was exactly right, and his thesis perfectly describes the post-AI economy. When the economic engine stops working for the median voter, policy swings take over the wheel.

What is currently breaking that engine? We are watching the bimodal compression of cognitive labour begin [3]. It is driven by two opposing forces acting on the market simultaneously:

• The Pull of Commodification: Capital pushes anything codifiable toward zero cost. If a task can be modelled, AI does it for fractions of a cent. The cost curve for that work collapses.

• The Pull of Irreducibility: What AI cannot do — cumulative judgment, signature liability, primary-source relationships, taste, and verification authority — becomes aggressively more scarce and exponentially more expensive.

The middle will get eaten from both sides. The routinely competent professional — the junior associate, the mid-tier analyst, the junior radiologist — is structurally exposed: their commodifiable output is absorbable by the model, while they lack the accumulated reputation and liability-bearing capacity that defines the top tier. As AI capability matures, the two ends of the economic spectrum will pull violently apart, and the middle will vanish.

Roughly 13% of US cognitive jobs, over 21 million workers, sit directly in the path of that compression [4]. Wage data reveals polarisation of this kind only in retrospect, on a lag of years, and the deployment wave is barely three years old. What shows up so far is concentrated in the occupations where adoption ran first. But the real political danger isn’t happening at the bottom of the income ladder. It is happening in the upper-middle-class waiting room that this compression is emptying.

Elite Overproduction: The Holding Pen is on Fire

This brings us to Elite Overproduction, a concept pioneered by historical macro-sociologist Peter Turchin [5].

Turchin studies history through the lens of population dynamics, and his thesis is dangerously relevant today: late-cycle societies produce far more credentialed “elite aspirants” than there are elite slots to absorb them. The American numbers tell the story plainly. The number of US lawyers tripled from roughly 400,000 in the mid-1970s to 1.2 million by 2011, while the population grew only 45 percent. MBAs over the same window grew six-fold [6]. The supply of credentialed aspirants for elite positions has run miles ahead of the fixed supply of positions themselves.

For decades, the economy kept these surplus elites pacified with high-paying, middle-management cognitive jobs. It was a massive holding pen for ambition. AI is the match held to it.

Nothing radicalises a population faster than a newly minted professional who realises their expensive master’s degree merely qualifies them to supervise a chatbot. When you have too many ambitious, highly educated people fighting over a rapidly shrinking pool of status-bearing jobs, they don’t form a support group. They fracture the political system.

Historically — whether you are looking at the late Roman Republic, pre-revolution France, or the late Russian Empire — intra-elite competition destroys institutional trust. The frustrated elite aspirants turn populist, harnessing the anger of the broader working class to tear down the very structures that locked them out.

The Fujiwara Effect: When Labour Throws a Brick

If Elite Overproduction explains the anger of the professional class, the Fujiwara Effect explains the mechanics of how that anger rewrites the rules of the market.

Borrowed from meteorology, where it describes two or more cyclones merging into a single larger system, the Fujiwara Effect has been made the organising metaphor of Viktor Shvets’s recent work as Global Strategist at Macquarie Capital, most fully developed in The Twilight Before the Storm [7]. In Shvets’s reading, three cyclones are converging: three decades of neoliberal policy, deep financialisation, and the disruptive Information Age now turbo-charged by AI. The central observation is grim: as these forces compound, the “winner-takes-all” dynamic becomes absolute, and the marginal economic utility of human labour collapses. First they came for the blue-collar assembly lines; now they have come for the white-collar spreadsheets.

But Shvets makes a lethal political observation: the changing value of votes.

When a median worker realises their labour no longer gives them any leverage over capital, they remember they still possess one asset the market hasn’t fully depreciated: their vote. And they tend to throw it like a brick.

The democratic system’s traditional mechanism for absorbing inequality is to translate economic discontent into gradual policy shifts. But when labour’s leverage falls this fast, gradualism dies. Silicon Valley billionaires suddenly discovering a deep, philosophical interest in Universal Basic Income isn’t altruism [8]. It is fire insurance. The serious versions are already being drafted, from corporate “displacement taxes” to an AI dividend fund modelled on Alaska’s [9]. They can read the room. They understand that a society where capital owns everything AI cannot make, and labour fights over the shrinking pie of what it can, is a society begging for a wealth tax.

The Catalyst: The Sovereign Accounting Trap

The sociological anger is the fuel, but the spark that ignites the political snapback is pure accounting. Modern Western governments fund themselves primarily through taxing labour (income and payroll taxes). Capital is highly mobile and taxed at much lower effective rates.

Every point of GDP that migrates from salaries to capital income leaves the payroll-tax net with it. If the bimodal compression runs its course, the sovereign tax base hollows out at precisely the moment the political system requires trillions of dollars to pacify a fractured, jobless middle class (via UBI, bailouts, or aggressive stimulus).

How does a sovereign fund a massive redistribution with a shrinking tax base? It cannot do it through gradual policy. It is forced into draconian asset taxes, financial repression, and aggressive currency debasement. This is the mechanical proof for why the coming crisis resolves via inflation.

The Inevitable Snapback: History’s Rhythm

No serious analytical framework predicts that accelerating disparity at record levels resolves itself gracefully. The tension always breaks, and it breaks via state intervention.

Look at the resolution of the three prior US secular bear markets:

• The 1930s (Deflationary): Resolved through the New Deal. A massive, state-executed restructuring of financial regulation, forced by the undeniable political pressure of mass unemployment.

• The 1970s (Inflationary): Resolved through Volcker’s extreme disinflation, after the failure of Nixon’s wage-price controls [10].

• The 2000s (Deflationary): Resolved through Quantitative Easing and financial repression [11]. This was a massive wealth transfer to capital owners to solve a banking crisis. It saved the system, but the bill for that disparity is now coming due.

A forced redistribution is coming. The only question for an investor is how to position capital before the politicians finish drafting the legislation.

The Investable Playbook: What This Means for Capital

The political question is not if the state will intervene to redistribute wealth, but where it will aim when it does.

• Overweight the Diffuse Scarcities (Real Assets & The Human Element): Capital owners of what AI cannot make — relational-sector luxury brands and physical-input chokepoints — hold pricing power that is durable to inflation. More importantly, they possess political camouflage. A politician is highly unlikely to nationalise an ultra-luxury handbag boutique, and they cannot easily expropriate a fragmented network of prime agricultural land. The populist case against them is weak.

  *The Macro Caveat:* Do not mistake political camouflage for invincibility. During an inflationary bust, the state will not outright steal the copper mine, but it will absolutely levy windfall taxes on the excess margin or impose price caps. You own these assets to survive the currency debasement, but you must actively manage the margin risk.

• Overweight Verification Infrastructure: If AI drops the cost of cognitive output to near-zero, the economic value migrates entirely to the entity that stamps it as true. Think of the global testing, inspection, and certification giants — Swiss-based SGS, French-based Bureau Veritas, or British-based Intertek. An AI can optimise a global supply chain flawlessly, but it cannot legally certify that a shipment of medical supplies meets EU safety standards. The liability, and therefore the economic moat, remains human.

• Underweight the Coordination Layer (The Expropriation Target): This refers to the digital middlemen. These are the platforms that don’t actually produce the goods or provide the service, but extract rent by controlling the matchmaking — think Amazon’s marketplace, Uber, App Stores, and the digital advertising duopolies. They are massive, highly visible, centralised, and deeply unpopular. When an angry political system needs a scapegoat to appease a fracturing society, Big Tech has a giant bullseye painted on its back. Margin-crushing regulation or antitrust dismantling of these platforms is a populist dream.

• Hedge with Store-of-Value (With Cyclical Caution): When the political resolution inevitably demands wealth redistribution, the fiat currency bears the cost. Gold and structural currency diversification compound significantly when governments inflate their way out of a social crisis. However, this is not a one-way street; precious metals typically suffer brutal liquidity drawdowns during initial market shocks before ripping higher in the policy response.

  ---

  What This Is Not

  To ensure we are trading on market realities and not political emotion, we must define what this framework does not claim:

  • It is not a prediction of exact timing. These are secular trends. In terms of sequencing, there are actually growing arguments — which I will cover in a coming issue — that we will likely see a cyclical, deflationary recession first. That recession will act as the catalyst, pushing the political system to panic and finally triggering the inflationary policy bust.

  • It is not a partisan political warning. State intervention historically comes from both sides of the aisle. Nixon (a Republican) imposed wage-price controls; FDR (a Democrat) built the New Deal. The math forces the state’s hand regardless of who occupies the office.

  • It is not a doomsday argument. The historical pattern is resolution, not total collapse. If you are familiar with Neil Howe and William Strauss’s The Fourth Turning [10], you recognise this rhythm. Late-cycle secular bear markets force an institutional reset, producing new social contracts and redirecting capital flows. Properly positioned capital does not just survive these transitions — historically, it multiplies.

    Subscribe now

    ---

    The Bottom Line

    We are living in a late-cycle society where the marginal worker’s economic leverage is falling vastly faster than the political system can comfortably absorb. Bimodal compression and Elite Overproduction are the disease; populist redistribution is the historical cure.

    The most dangerous assumption an investor can make today is that the benign regulatory environment of the 2010s will survive the brutal economics of the late 2020s. Asset prices reprice the exact moment a political mechanism becomes visible — not when the bill is signed into law.

    The people who get this right will hold the assets the resolution rewards: the diffuse, inflation-resistant owners of the human element, physical limits, and verification moats. The people who get it wrong will hold the massive, highly visible coordination-layer platforms, watching their margins compress against regulatory caps while the wealth they thought they had concentrated quietly migrates away.

    Thanks for reading! Subscribe for free to receive new posts and support my work.

    ---

    References

    [1] Chief Justice of India Surya Kant’s remarks in open court, 15 May 2026, and the rise of the satirical Cockroach Janta Party (founded 16 May 2026 by Abhijeet Dipke; roughly 19–22 million Instagram followers within a week, surpassing the ruling BJP’s official account): BBC News, CNN, The Guardian and Al Jazeera, 19 May – 8 June 2026. Graduate unemployment: nearly 40% of Indian graduates under 25 are jobless, per Azim Premji University research cited by CNN, 22 May 2026.

    [2] Robertson Morrow, “The Bull Market in Politics” (Clarium Capital, circulated mid-to-late 2000s). Morrow was CIO of Clarium Capital. Primary attestation in George Packer, “No Death, No Taxes: The Libertarian Futurism of a Silicon Valley Billionaire,” The New Yorker, 28 November 2011, which discusses the paper and the thesis within a profile of Peter Thiel. Morrow’s later by-line at The American Conservative (2002–2005) provides corroborating biographical detail.

    [3] Stray Narratives, Issue 12: “Who Gets Paid After AI,” May 2026. The four-mechanism framework and the bimodal compression that runs through each.

    [4] Stray Narratives, Issue 06: “The Noise Economy,” April 2026. The 13% / 77% workforce split.

    [5] Peter Turchin, End Times: Elites, Counter-Elites, and the Path of Political Disintegration, Penguin Press, 2023. The elite-overproduction framework is developed across the book; earlier statements appear in Ages of Discord (2016).

    [6] American lawyer counts ~400,000 (mid-1970s) to 1.2 million (2011), per the American Bar Association, cited in Peter Turchin, Ages of Discord: A Structural-Demographic Analysis of American History (Beresta Books, 2016) and End Times (Penguin Press, 2023). MBA six-fold expansion over the same window per the same Turchin sources. Secondary summary in Noah Smith, “Blame Rich, Overeducated Elites as Our Society Frays,” Bloomberg Opinion, 20 November 2013.

    [7] Viktor Shvets, The Twilight Before the Storm: From the Fractured 1930s to Today’s Crisis Culture (Boyle & Dalton, August 2024). Shvets is a global strategist at Macquarie Capital; the book is cited here as a public work. The Fujiwhara Effect itself is a 1921 meteorological term (named after Sakuhei Fujiwhara) describing the interaction of two or more cyclones; Shvets adopted it as the organising metaphor of his macro thesis, deployed across both The Twilight Before the Storm and his earlier The Great Rupture (2020). His three-cyclone reading (neoliberalism + financialisation + Information Age) is laid out at viktorshvets.com/the-fujiwara-effect and discussed with Alan Kohler in The New Daily, “The Fujiwara Effect — three cyclones of modern world,” 26 August 2024.

    [8] Silicon Valley figures advocating Universal Basic Income, including Sam Altman’s Y Combinator Research UBI experiment and public commentary, 2016 onward.

    [9] Alap Shah, “The Global Intelligence Crisis, Part Three: The Path Forward” (alapshah1.substack.com, March 2026) — the “American Prosperity Compact”: a corporate displacement tax funding income security and an American AI Dividend Fund modelled on the Alaska Permanent Fund.

    [10] Paul Volcker’s Federal Reserve tightening cycle, 1979–1982, taming inflation following the failure of Nixon’s 1971 wage-price controls. Federal Reserve archive and historical record.

    [11] Federal Reserve Quantitative Easing programmes commencing November 2008; the ensuing decade of financial repression as the macro-policy response to the Global Financial Crisis. Federal Reserve communications and successor literature.

    [12] William Strauss and Neil Howe, The Fourth Turning: An American Prophecy, Broadway Books, 1997. The generational-cycle framework.