The AI chain under stress

Research note · 2026-11-01 · Yichen Zhan · Prices at the 2026-10-30 close

Who in the AI build-out pays for whom, who would run short of cash first if new money stopped, how locked-in each layer's customers are, and what past manias suggest about the aftermath. The chain is read from the user end to the machine end: model labs and apps, clouds, chips, foundry and equipment.

Key findings1 · Valuation and cash2 · Stickiness3 · Past manias4 · Recent bustsNotes

Key findings

1. The cash flows the opposite way to the hype. The upstream suppliers (NVIDIA, Broadcom, Micron, TSMC, ASML, AMD) fund themselves even in our stress case. Microsoft, Alphabet and Meta fund themselves on trailing numbers but not if operating cash fell 30% while spending held at its latest run-rate. The firms that already need outside money are CoreWeave, Oracle, SpaceX (which now contains xAI), Amazon on a trailing basis, and OpenAI.
2. Who could run dry within 12 to 24 months if funding stopped. CoreWeave about 3 to 14 months, depending on whether its $10bn of undrawn facilities can be drawn; Oracle about 9 to 17; OpenAI about 16 on one press-based burn path; SpaceX/xAI about 16 to 27; Amazon about 15 in the stress case only. Anthropic generates cash today; its risk is contracted compute arriving faster than revenue, which exhausts its $132bn cash pile in about 24 to 31 months if revenue stopped growing and no money came in.
3. The chain is sticky in the middle and fragile where the balance sheets are thinnest. Clouds, memory and custom-chip suppliers hold years of contracted backlog, and CUDA and EUV lock in the ends. The weak layer is the model labs: short-dated, pay-as-you-go revenue set against years of take-or-pay compute, part of it financed by the suppliers themselves. In the scenario sketched below, a slowdown would show up first as credit stress rather than cancelled orders.
4. By size, this is bigger than the 2000 telecom boom and smaller than railways or housing. Five firms' trailing capex was about 1.8% of US GDP (an upper bound), against 1.2% for telecom in 2000 and 5 to 7% for railways and for housing at their peaks. Unlike the railway, telecom and housing manias, most of it is paid out of operating cash flow; since 2025 a growing share is debt and off-balance-sheet vehicles.
5. History is unkind to the boom sector, kinder to the economy. In the episodes we measured, the broad US market regained its old peak in real total-return terms within 10 years in 4 of 6 cases; the boom sectors did so in 0 of 10. The physical assets of railways, power and fibre were later used at lower prices; GPUs, on five- to six-year lives, cannot wait that long. Buildings and power hookups can.
6. Recent busts all left the technology standing. What decided the losses was who paid and how long the assets lasted. Incumbent-funded busts (the metaverse, self-driving) hurt the funders and stopped there; retail- and leverage-funded ones (crypto, SPAC electric vehicles) wiped out the holders. Humanoid robotics today looks most like self-driving in 2017-19: about $116bn of headline valuations on under $1bn of revenue, waiting on a capability that has not arrived.

1 · Valuation, cash generation and runway

Listed companies, downstream to upstream

Trailing four quarters. FCF = operating cash flow − purchases of property and equipment. Cash includes marketable securities; debt includes finance leases, excludes operating leases. Revenue growth is trailing four quarters on the four before (SpaceX: nine months on nine months). TSMC and ASML caps are ADR-based approximations. Fiscal windows end between 31 May 2026 (Oracle) and 30 September 2026 (SpaceX, TSMC, ASML). *Alphabet's and Amazon's earnings are inflated by revaluations of private stakes: Alphabet booked $99.0bn of equity-security gains in Q2 2026 alone, which its 10-Q attributes primarily to SpaceX and a private company; Amazon booked a $50.5bn upward adjustment on its Anthropic stake. Taking non-operating items out gives P/Es of roughly 25 to 28 for Alphabet and 29 to 37 for Amazon, depending on whether all non-operating income or only this year's stake gains are removed, and on the tax treatment (est.). NVIDIA's earnings include $23.7bn of equity gains in the first half of its fiscal year, about 17% of pre-tax income.

Will the cash run out within 12 or 24 months?

Three mechanical cases. (a) Trailing free cash flow continues. (b) Stress: operating cash flow falls 30% while capex stays at its latest quarterly run-rate (or its trailing level, if higher). (c) The stress case with no new debt or equity, and debt falling due repaid from cash. Maturing debt is otherwise assumed to roll over. Guidance was not used, to keep every company on one footing.

"Self-funding" means cash never falls below zero over 24 months. CoreWeave counts the $10.0bn of undrawn facilities stated in its Q2 10-Q in cases (a) and (b) but not in (c), which allows no new borrowing; the company says its liquidity covers the next 12 months, and these months measure its dependence on new financing, not a default call. The sealed cloud names' capex is a quarter behind the others while capex was rising fast, so their stress burn is probably understated. Their filings carry no debt-maturity schedule we could use, so (c) equals (b) for them. Capex here excludes finance-lease additions and assets built through partners and special-purpose vehicles, which understates the cloud names' true spending.

The private labs

OpenAI figures are unaudited press reports. Lab runways are straight-line illustrations, not forecasts. Whether SpaceX's largest customer is Anthropic is our inference; the 10-Q does not name it.

2 · Customer stickiness, layer by layer

How stress would travel

1. Apps (weeks): usage-based revenue slows; enterprises route tokens to cheaper or open models.
2. Lab funding (1-2 quarters): growth conditions in supplier note agreements stop being met; private valuations reset; suppliers' revaluation gains on lab stakes reverse, hitting their earnings before their revenue.
3. Neoclouds and Oracle (2-4 quarters): the most concentrated and leveraged lessors renegotiate or slow capacity acceptance. The first cancellable contract in the chain, Anthropic's up-to-$120.5bn deal with SpaceX, can be ended on 90 days' notice from early 2027.
4. Cloud capex (next budget cycle): contracted revenue keeps flowing, but growth capex and not-yet-commenced leases are deferred.
5. Chips: NVIDIA and AMD, selling on purchase orders with little backlog, feel it first; Broadcom and Micron have contractual cushions but carry credit exposure through backstops and notes.
6. Foundry, then equipment: TSMC utilisation follows a quarter or two later; ASML comes last and swings most, because a slowdown in capex growth becomes an outright fall in tool orders.

The pattern that makes demand look stickier than it is: Amazon, Alphabet, NVIDIA, AMD, Broadcom and Microsoft all invest in, lend to or guarantee the labs and clouds whose purchases fill their own backlogs. Contracts hold on paper until a counterparty cannot pay. On this reading, the labs' credit says more about the chain's fragility than the vendors' backlog headlines do.

3 · Against past manias and crises

US index figures are computed from Robert Shiller's monthly data; sector figures from Kenneth French's industry portfolios; "lasting" means the market stayed above the old peak afterwards. AI: trailing capex of Microsoft, Alphabet, Amazon, Meta and Oracle to 2026Q2, $566bn, over US GDP; the capex includes non-US and non-AI spending. On the government's own lines, investment in IT equipment and software is 5.05% of GDP, against 4.46% at the end of 2000. The BFI is this site's index; its level at past peaks varied too much to mark a peak.

Base rates for recovery (descriptive, a handful of episodes, not a forecast)

• Broad US market, real total return, first time back at the peak: under 5 years in 2 of 6 episodes, under 10 in 4 of 6, under 25 in 5 of 6. Counting only lasting recoveries: 1, 3 and 5 of 6.
• The boom sector itself: 0 of 10 regained its peak in real terms within 10 years; all 10 within 25. In every pair we could measure, the boom sector took about 1.5 to 3.4 times as long as the market. Single leaders did worse or never returned: Cisco regained its 2000 price high only in December 2025, after 25.7 years; WorldCom, Nortel, Lehman, Insull and Jay Cooke failed.
• Who carried the leverage. Japan, where banks lent against land, had by far the slowest recovery; 1929 ran on margin debt and holding-company pyramids; 2007-09, carried by households and banks, recovered in about 5 years but with state support. In the AI build-out the leverage sits mainly with corporations and private-credit funds, but it is growing: debt including finance leases rose by an amount equal to 47% of five-firm capex in the year to mid-2026, against roughly zero in 2024.
• Was the asset used later? Railways, power plants and fibre were, after their first owners were wiped out: US track grew 48% between 1893 and 1907, and fibre laid in 1996-2000 was lit years later. AI's data-centre shells, power and cooling have that kind of life. Its accelerators, depreciated over 5 to 6 years and possibly obsolete sooner, do not.

4 · Against recent technology busts

Risk profile (our judgment on a 1-5 scale, 5 = most risky; not a measurement)

How AI differs. Its revenue is real and growing fast: NVIDIA's data-centre revenue was $89bn in one quarter, and the two largest labs run at roughly $115bn a year combined (partly press-reported). But the four largest clouds' capex, at about $660-690bn a year on the latest quarter, is still about six times that lab revenue. Its funding is better than crypto's or the SPACs', because it comes mostly from the richest balance sheets, though not as clean as the metaverse or self-driving, which their sponsors paid for outright; in 2026 the margin moved to bonds, leases and joint ventures. It is the only case here large enough to matter for the whole economy, and its main asset loses value faster than any of the others. Power, turbines and transformers slow the build; that delays returns but also caps how far it can overshoot.

How much might be recovered. As a scenario: if capex overshoots, and if buildings, power and cooling are 40-50% of it, roughly 40-60% of the spending would stay productive. That is the share in buildings, power and cooling, assumed to be 40-50% of capex with 20-plus-year lives. The chips would not keep their value. Under the same assumptions, losses would be concentrated in leveraged builders and their lenders (the neoclouds, Oracle, the special-purpose vehicles) rather than in the four richest clouds. In a humanoid stall, perhaps 10-30% of headline valuations would survive, with profitable hardware makers like Unitree as the floor. All of these shares are assumptions, stated to make the reasoning checkable.

Notes and limits

• Sources. SEC filings (10-K, 10-Q, 20-F, 6-K, S-1) for listed companies and Anthropic; company releases for TSMC and ASML; named press for OpenAI, private robotics firms and the deal headlines; Shiller, French, BEA and FRED for history. Every figure was traced to a source and checked by an independent audit pass before publication; the working files are kept with the research archive.
• Selection. The listed companies are today's AI leaders, chosen with hindsight. Failed and private firms are mostly missing, so the chain's total funding need is understated.
• Mixed dates. Fiscal years differ, and the four sealed clouds and Oracle lag the rest by about a quarter.
• Mechanical cases. The stress and runway cases ignore working capital, buybacks, dividends, interest and new contracts. They show dependence on outside money, not probabilities of failure.
• Judgments. The stickiness grades, the 1-5 ratings and the recovery shares are reasoned judgments, not measurements.
• History. The recovery base rates come from a few overlapping, US-heavy episodes chosen because they were famous busts. They describe the past.

Research, not investment advice. Every reader gets the same page; nothing here is a recommendation to buy, sell or hold anything. The author may own securities discussed here; personal holdings are not published. Numbers can be wrong or later revised by their publishers. anti-fragility.app · Letters