Is AI a bubble, and should we cut risk into this sell-off?

The question behind the sell-off

In the week to October 15, hedge funds sold technology stocks faster than at any time in at least a decade. ¹ CoreWeave's bonds trade like distressed debt, and insuring Oracle's debt now costs what the market charges junk-rated companies. ^2,3 Yet the people running the build say they cannot build fast enough. Amazon's Andy Jassy says that even after about $220B of capex this year, his capacity is sold out for 2026, largely sold out for 2027 and partly sold for 2028. ⁴ "This is not like fiber, it's not dark," says Google's Sundar Pichai. "We are using it." ⁵

Both sides are right, and that is the puzzle. The demand is real. What may not be real is the money at the edge of the build: the borrowers who built ahead of demand, and OpenAI's promise to pay for compute it has already ordered. The bubble debate runs three questions together, and we take them apart. Can the builders pay their bills? Will the build earn its return? Will the stocks pay from today's prices? The cash-rich core can pay. Whether the build earns its return is not yet settled. The debt-funded edge has already broken. For a portfolio, that means keeping market exposure but cutting how much of it rides on AI.

The bears have serious company. Meta has told investors it may need to "raise hundreds of billions of dollars". ⁷ The most-read bear piece of the month, from TS Lombard, imagines the AI boom ending in early 2028 with the S&P 500 down 30%. ⁸ So the question for an investor is not whether AI is real; it plainly is. The question is whether the money behind the build runs out before the revenue arrives, and how much of that risk the chip sell-off has already priced.

We answer in three steps. First, is it a bubble? We set out the bear and bull cases as their best advocates make them and test both against the money. We then follow the money to its weakest link, OpenAI, and check whether the biggest buyers will keep spending in 2027. Second, how far could prices fall from here, with chip stocks already deep in a sell-off? Third, what should an investor do?

The bear case

The bears are right about the mechanics. Vendor financing, debt kept off balance sheets and two labs carrying most of the revenue are all real and measurable, and at the debt-funded edge the damage has started. Where the numbers do not back them is the timing: chip rents are rising, not falling.

Dario Perkins: a bubble that ends in a balance-sheet bust

TS Lombard's "Perkini Report" is written as a post-mortem from 2030. ⁸ Perkins calls it a scenario, not a prediction. Its five fault lines are the most complete version of the bear case we have read:

• Capex recycling. One company's capex is another's earnings, so the more the hyperscalers (Microsoft, Alphabet, Amazon and Meta, the biggest buyers of AI capacity) spend, the better AI revenue looks. Revenue inside the loop proves nothing; what matters is whether people outside it pay. That is the test we apply in the section on outside money.
• Vendor financing. Suppliers invest in customers who then buy from them. Money leaves as equity and comes back as orders, "with NVIDIA at the centre of everything".
• Off-balance-sheet leverage. Data centres are increasingly built by special-purpose vehicles and leased back, so the debt sits outside the hyperscalers' balance sheets.
• The labs' numbers. OpenAI's and Anthropic's figures arrive through leaks, are unaudited, and lean on annualised run-rates that flatter them.
• Concentration. About $200B of current revenue supports more than $1T of commitments from two companies.

In his story the turn comes when OpenAI's IPO shows audited numbers that do not match the pitch. In January 2028 the hyperscalers cut their 2028 capex guidance by 45%, to about $600B. By April the S&P 500 is down about 30%, AI stocks 40% and the rest of the market 20%. Neoclouds trip their loan covenants and go bankrupt, and a large leveraged cloud provider is cut to junk. The hyperscalers' own credit holds. The recession is mild, more 2001 than 2008, and the Fed cuts hard. His advice is to hold a normal amount of risk with less of the AI complex than the index gives you, add bonds as the cracks show, and not try to be early.

What the numbers say. Most of his fault lines are real, and several can be measured. Take vendor financing first. NVIDIA has guaranteed $108.5B of its customers' obligations, a sum equal to 36% of a year's revenue. Lucent, the telecom equipment maker whose loans to its customers became the textbook case of vendor financing gone wrong, was at 24% in 2000. By size NVIDIA looks worse, though the section on outside money shows its guarantee is narrower than Lucent's loans were. ^9,10 Off-balance-sheet leverage is growing too. Meta's signed leases on data centres that have not yet opened, which do not appear as debt, rose from $103.8B to $355.7B in nine months. ¹¹ His concentration figure holds up as well: his $200B of revenue is close to our own count of about $160B a year including Anthropic. On the labs' numbers he has a point, but a smaller one than it first appears. OpenAI says its revenue is running at about $50B a year. The revenue it actually booked in the second quarter, $6.7B, works out to only $27B a year. The section on OpenAI explains why most of that gap is fast growth rather than flattering arithmetic. ^12,13

What the numbers do not support is his timing and depth. His turn is more than a year away, and a lot can change before then. His story also needs chip rents to collapse, and today they are rising. That does not refute him. High rents now are exactly what draws in the supply that would land in 2028, and a glut then is what his story needs.

Ed Zitron and Craig Moffett: too little revenue, at falling prices

Ed Zitron, a newsletter writer and the most prominent AI bear, argues that almost all AI revenue is two labs, OpenAI and Anthropic, paying the clouds for computing power. If he is right, the clouds' AI growth is not evidence of broad demand. It is money from the labs' investors passing through. He says brokers put the labs at 73% of Amazon's AI revenue and 70% or more of Microsoft's. We could not find those figures in the brokers' own reports, so we treat them as his. He also adds up more than $1.35T of lease and purchase commitments at four hyperscalers that do not appear on their balance sheets. ^14,15 Craig Moffett, interviewed on Stratechery, adds a second problem: deflation. Most work does not need the best model, so customers move to cheaper ones, and the price of AI falls faster than its use grows. In that world revenue can shrink even while usage booms. ¹⁶

What the numbers say. For Microsoft, the labs' weight is confirmed by its own figures: it booked $24.1B of revenue from OpenAI in fiscal 2026. ¹⁷ For Amazon, the labs' share is not disclosed, and we could not confirm the estimates he relays. Either way, counting the labs' payments to the clouds as demand would count the labs twice. So when we count revenue below, we include only cloud AI revenue that does not come from the labs. Moffett is right that the price of a unit of AI is collapsing. Epoch AI, a research group, estimates that the cost of reaching a given level of capability falls about 13-fold each year. Token prices, the price per unit of text a model reads or writes, fell 43% in ten weeks this summer. ^18,19 Usage is rising fast too: the number of tokens Google processes each month rose 6.7-fold in a year. ²⁰ These figures cannot tell us whether revenue is outgrowing the price fall, because they cover different products and periods. But one test does not need that comparison. In an overbuilt market, the rent for computing capacity should fall. Instead, the price of one-year rental contracts for NVIDIA's H100 chip rose 38% from October 2025 to March 2026. ²¹

Chamath Palihapitiya: the payoff is slow and lopsided

In September, Palihapitiya argued that AI pays off only for firms that first redesign how their work is done, and that most firms instead bolt it onto old processes and see little gain. ²² The evidence he leans on holds up. In a field trial with 515 start-ups, the firms that were shown how others had reorganised around AI went on to earn 1.9 times the revenue of firms that were not. Almost all of that gain came from the top tenth of firms. ²³ The economy-wide numbers show no boom yet either. US output per hour worked has grown 2.06% a year since 2019, against 3.30% a year in 1995–2003, when the last technology wave was paying off. ²⁴ And Ramp, which tracks business card spending, finds that 55.9% of businesses paid for AI in September, but that share is rising more slowly each month. ²⁵

This is the bear point we weigh most, because it is about timing. In 1987 the economist Robert Solow complained that you could see the computer age everywhere except in the productivity statistics. Productivity did pick up, but only about eight years later. If AI runs on a similar lag, buyers' gains arrive after the bill for the build falls due, and whoever financed the build has to carry it in the meantime.

Lenders put CoreWeave's default near a coin flip

Lenders sit closest to the edge, and this month they were the loudest bears. On the All-In podcast, the hosts looked at CoreWeave's credit default swaps, the price of insuring its debt against default. At about 953bp, or 9.53% of the debt's value a year, they read that price as roughly a 55% chance of default within five years: in their words, "a coin flip". The reading assumes lenders get back the standard 40% in a default, and it is what the insurance market charges, not a forecast. They also noted that three customers make up about 75% of CoreWeave's revenue, and that it needs to raise roughly as much debt again as the $35B it already owes. ²⁶

On the Monetary Matters podcast, its host Jack Farley estimated that about $500B of next year's AI capex will have to be borrowed, through bonds, private credit, special-purpose vehicles and debt backed by the chips themselves. Because the same build-out stands behind all of it, he argued, defaults could become "extremely correlated" within 18 to 36 months. Jim Bianco of Bianco Research, a bond-market strategist, agreed that was the risk. ²⁷ And CoreWeave's ordinary bonds rank behind the private-credit vehicles that own its GPUs (the chips that run AI models), so in a default its bondholders would be paid last. ²⁸

The same week brought evidence for Moffett's deflation point from the buyers themselves. Meta told investors it expects its AI to become ten times more cost-efficient within a year. If that happens, Meta can do the same work with far less compute. ²⁹ On the 20VC podcast, Jason Lemkin, founder of the software community SaaStr, noted that Replit, a coding start-up, had switched its default model to OpenAI's because it gave similar quality at 30–40% lower cost. His warning was that there is "only so much R in the ARR": annual recurring revenue is only as recurring as customers' loyalty, and customers switch on price. ³⁰ Anthropic's quarterly revenue growth slowed from 147% in Q2 to 45% in Q3, on the reported figures. ²⁸ On his Plain English podcast, the writer Derek Thompson pointed out that outside the chip and memory makers there is little visible return yet, and that only about 3% of people pay for AI themselves. ³¹ Even Satya Nadella describes Microsoft's strategy as building to meet current demand and deliberately spending less than it could at the frontier. ³²

Ken Griffin and Freda Duan: the depreciation bill and the airline problem

Two of the most-shared bear arguments on X this week were about cost. ³³ Freda Duan of Altimeter Capital put 2027 AI infrastructure spending at roughly $2.5–3.0T for about 50 GW of capacity. Chips and buildings wear out, and their cost is charged against profits over their working lives as depreciation. On her numbers, that spending implies $400–480B a year of depreciation. Against it, rumours put the labs' revenue near $200B a year by the end of 2026. She thinks the labs need something closer to $400B by the end of 2027 for the sums to work. Asked whether the return was ever a slam dunk, her answer was "Never, really." ³⁴ Ken Griffin, who runs Citadel, compared GPUs to airline seats. Once the plane is flying, an extra passenger costs almost nothing to carry, so when there are more seats than travellers, fares collapse. Fibre-optic bandwidth went the same way after 2000. He also questioned whether GPUs stay useful as long as their owners' depreciation schedules assume. If they wear out sooner, today's profits are overstated. ³⁵ And Brad Gerstner of Altimeter, a prominent AI bull, sold $650M of CoreWeave shares a month ago, reportedly because lenders will no longer make cheap loans secured only on GPUs. ³⁶

The market is already treating capex as a cost rather than a promise of future profit. Meta beat analysts' estimates last week and raised the bottom of its 2026 capex range (the section on the biggest buyers has the numbers). Its shares fell 9.4% on the week. Investors marked the stock down for spending more. ^37,38

The builders are already breaking

At the debt-funded edge, the bear case is already happening. Oracle is the clearest case. Its five-year credit default swap, the yearly price of insuring its debt, was about 295bp on October 19 after touching 300bp, a level usually seen on junk-rated companies. It sold $20B of new stock last quarter to help pay for the build. Its net debt is 3.2 times its EBITDA (earnings before interest, tax, depreciation and amortisation, a rough measure of cash profit), and its free cash flow, the cash left after capex, is negative. ^39,40,41 CoreWeave is further along. Its bonds due in 2031 and 2032 yield about 15%, the return investors demand when they doubt they will be repaid in full. Its $35.1B of debt equals 34% of all the revenue it has under contract. ^42,43 The stress is also spreading beyond AI. The extra yield on the lowest-rated corporate bonds, those rated CCC, has crossed 13 percentage points over Treasuries, a line we set in advance as a warning. ⁴⁴

The bull case

The people signing the purchase orders say demand exceeds supply, and for 2027 the evidence backs them. Rents are rising, most of 2027's capacity is already under construction or financed, and demand reaches well beyond the two labs. What the bulls concede is that a few firms, many of them funded by the same boom, do most of the paying.

The builders say demand exceeds supply

The bear case is made mostly by commentators and lenders. The bull case is made by the people who sign the purchase orders, with their own money on the line. Jassy says Amazon could spend a good deal more and it would still be a good use of capital. He puts the payback on a server at about three years, against a working life of five to six years and contracts of about five. He also thinks inference, the work of running models for users, is still roughly ten times too expensive for pervasive use. On that view, cheaper tokens bring more demand, not less revenue. ⁴ Pichai says everyone in the industry is short of compute and wishes they had twice as much. His confidence rests on signed contracts and backlogs: "it's not just hope". He concedes the industry can overbuild, and that Alphabet's free cash flow turned negative this summer for the first time since its IPO. ^5,45

The outside forecasters agree on scarcity, for now. SemiAnalysis forecasts about 50 GW of data-centre capacity added in 2027, roughly double the consensus, and says megawatts are now harder to get than GPUs. It names a data-centre oversupply in 2028 as the main risk. ⁴⁶ On All-In, David Sacks noted that even old H100 and A100 chips rent at a premium, and Ben Horowitz drew the contrast with 1999–2001, when bandwidth fell from about $1,500 to $50 per Mbit: today token prices are falling, but GPU prices are not. ²⁶ And Jim Bianco argued that the CCC blowout is a few problem borrowers rather than a warning for the whole market. In his words, "it's 1997", not the spring of 2000. ²⁷

The physical build backs that up. SemiAnalysis counts more than 24 GW of data centres in an advanced stage of construction across just 40 campuses, most of it due in 2027, and over $1T of binding, non-cancellable leases signed so far in 2026. The leases matter because developers usually spend only 5–10% of a project's capital before a tenant signs; the lease is what lets them borrow the rest. ⁴⁷ The political backlash is real but small in megawatts. Of more than 300 local moratoriums, SemiAnalysis finds that about 20 GW of planned capacity sits inside a restricted area, but only about 1.5 GW is actually delayed, or about 2.3 GW counting New York's state order. That is against its forecast of 38 GW added in the US in 2027. ⁴⁸ Most of 2027's capacity is already under construction or financed. If a bust comes, it is more likely a 2028 story.

Not everyone thinks the capacity will arrive that fast; some expect power shortages to push part of the 2027 build into 2028. SemiAnalysis says the power problem has been "broadly solved" by on-site generation. Both views point the same way for 2027: too little capacity rather than too much. They differ on how much of the delayed supply then lands in 2028, which is exactly where the oversupply risk sits.

Demand reaches well beyond the labs

The second leg of the bull case is that demand is broad, not just two labs, and the companies' own reports bear that out. Microsoft has 100,000 customers building on its Foundry AI platform and more than 30M paid seats for Copilot, its AI assistant. Nearly 500 Google Cloud customers each processed more than a trillion tokens last year, and Google Cloud's revenue grew 82%. AWS grew 37%, and its AI business is running above $25B a year. Palantir's US commercial revenue grew 149%, and the coding tool Cursor passed $4B of annualised revenue in June. ^{50,51,52,53,54,55,56} Advertising, which has nothing to do with the labs, tells the same story. Meta's ad revenue grew 26% in Q3, and Meta credits its AI ranking models with 15.7% more conversions on Facebook: advertisers get more sales for the same spend, and pay Meta for it. ^38,57 One caution applies. The cloud growth figures include what the labs pay, so they overstate demand from everyone else. Meta's advertising figures do not have that problem.

Palihapitiya again: value collects at the chokepoints

In May, Palihapitiya laid out the AI stack in six layers: infrastructure, chips, data, models, execution and applications. ⁵⁸ His argument is that each layer has a few assets everything above it must pass through, and those owners keep the value. The cost of running a model, he notes, has fallen about 1,500× in six years. So the models themselves commoditise, and the money goes to the chokepoints: lithography machines, power, chips and the software that runs them. That is a bull case for the suppliers, and it cuts against our view that they are hit first. We take it seriously, and come back to it in the verdict.

Rents for available chips are rising, though not everywhere

The cleanest test of overbuilding is price. When a market has built too much capacity, the price of renting it falls, and in the telecom bust it fell fast. Internet transit prices, what networks paid to carry each other's traffic, fell 16% in 2000, 40% in 2001 and 50% in 2002. ⁵⁹ Today the price neoclouds charge to rent an AI chip is rising. Silicon Data's index of their hourly H100 prices rose from $2.00 to $2.88 between December and October. ⁶⁰ The big clouds' list prices tell a different story: Silicon Data finds their H100 rates fell 26–54% by region between early 2025 and mid-2026, though it notes that its figures cannot separate price cuts from a change in the mix of contracts. ⁶¹ The market for future rent has also stopped pricing further rises for older chips. Carmen Li, Silicon Data's founder, says its forward curves sloped upward earlier this year, but those for the A100 and H100 have recently flattened, while the newer B200's still slopes up (Pitch The PM podcast, October 16). One trade report goes further and says the curves now slope down. ⁶² Oracle reports that 97.9% of its GPUs are in use, and that customers renewing contracts on older GPUs pay a 20% premium. ⁶³ Capacity that can be had now is scarce. That is the clearest difference from 2000, though rents differ by provider and region, and a fully rented chip can still serve a customer who loses money. Part of the scarcity may come from shortages of memory and power rather than from demand alone. Either way, the capacity that exists is being used.

Valuations are high, but not 2000-high

The ten largest companies make up 40.6% of the S&P 500's value, but on every measure in the chart they stand on firmer ground than the top ten of March 2000, and they earn far more. Their earnings are somewhat flattered by gains on stakes in AI labs, such as Alphabet's gain on Anthropic in Q2. ⁶⁴

The chart here is now drawn in the data section. It showed: Top ten S&P 500 companies: 26.5 times earnings today against 53.6 in March 2000; profit margin 33% against 9.1%; 36.1% of index profits against about 15%.

The biggest spenders also carry little debt. Their combined net debt is about 0.3 times EBITDA as reported, and 0.6–0.8 times if leases and special-purpose vehicles are counted. Oracle, at more than 3 times, is the outlier. ⁶⁵ Those figures leave out leases signed for buildings not yet finished, which reach the balance sheet only when a site opens. Reuters put them at about $1.09T across Microsoft, Meta, Oracle, Amazon and Alphabet in August; Microsoft's alone reached $329B in June, up from $197B in its previous filing. ⁶⁶ Paid over the usual 15 to 20 years, that is roughly $55–75B a year, against about $700B of yearly operating cash flow at the five. It is a fixed cost that grows with every lease signed, and it cannot be cut the way capex can, but it is not a debt problem today.

The size of the build is less reassuring. Counting only the buildings, data-centre capex was about 0.8% of US GDP in Q1 2026. ⁶⁷ Counting the computers as well, Cape Fear Advisors puts AI investment at 1.3% of GDP over the year to June, or 1.5% at the second quarter's pace. That already matches telecom's peak of 1.4% in 2000, though it is below the railways' 3.1% in 1893. ⁶⁸ So the build is no longer small for the economy. It is about the size of the last bubble's, and still growing. The difference is who pays: mostly companies with far more profit than the telecom builders had.

At today's rents, the build pays

The room to spare is about $3B a gigawatt a year: rents can fall 15% before a cash-funded build stops earning its return. Matt Dratch, replying to Duan on X, showed this with a break-even calculation built from her own numbers. A gigawatt of AI capacity costs about $55B to build. About 30% of that goes on buildings and power equipment that last 25–30 years, so only the rest wears out quickly. The chart sets what a gigawatt brings in today, on Duan's own figures, against what it needs to earn its owners 15% a year. ⁶⁹

The chart here is now drawn in the data section. It showed: Revenue per gigawatt today is $20B, against break-evens of $17B paid in cash and $13-14B if 80% borrowed; rents a third lower give $13B.

If rents fall by a third, the room is gone and Griffin's airline problem arrives. That is why we watch GPU rents more closely than any other single number. We expect rents to fall as each new chip generation's supply lands, even with the chips still fully used. A sustained fall of more than 15% would use up Dratch's margin. The build does not need rents to stay where they are, but it cannot afford a fall much larger than that.

Anthropic's filing backs its numbers

Anthropic told investors it had positive free cash flow in Q3 and an adjusted operating margin above 11%. ^70,71 Its S-1, filed October 26, backs the direction: first-half revenue of $16.3B, an operating profit of $974M in Q2, operating cash flow of +$3.9B and more than 1,900 customers paying over $1M a year. ¹⁸⁴ The half-year figures are unaudited interim statements, and this letter was prepared with Anthropic's model (see the disclosure), so we count them only where the filing states them. So far, Perkins's IPO moment is working the other way for this lab.

What the bull case concedes. Three facts cut against the breadth argument. First, Microsoft's backlog of contracted future revenue grew 84% in total but only 25% excluding OpenAI, so most of the growth came from one customer. ⁵⁰ Second, on Ramp's card data, 80% of what businesses spend with OpenAI and Anthropic comes from 1% of firms. Many firms use AI, but a few do most of the paying. ⁷³ And those few are often funded by the same boom: on Ramp's data, as cited by Tomasz Tunguz, 80% of the largest AI spenders are venture-backed, against 13.1% of the median spender. US venture investment reached $516B in the first nine months of 2026, with 82.7% of it going to AI. ⁷⁴ Part of what we count as outside money is investors' money passing through start-ups to the labs. Third, Microsoft has not updated its $37B AI revenue run-rate since April. A company with good news usually shares it, so we read the silence as a mild negative. ⁷⁵

Does outside money pay for the build?

Not yet in full. Outside revenue comes to 69% of the yearly capital cost of what is already built without Anthropic, and 118% with it, before running costs. Paying for the build planned to 2030 needs that revenue to grow 52–73% a year. We put about 45% low confidence, 30–60% on it getting there on time counting Anthropic, and 20% low confidence, 10–30% without. This sizes the gap. It does not measure profit.

Companies are not waiting for the return. In a Gartner survey published October 13, 85% of business leaders expected to raise AI spending, while 13% of CIOs reported impact beyond productivity gains. ⁷⁶ That is the pattern this section measures: spending runs ahead of the revenue that would pay for it.

The two cases agree on more than they admit. Both say demand is real. They disagree on whether money from outside the AI industry will grow fast enough to pay for what is being built. Moffett put the test well: are "the revenues coming into the system greater than the costs of building the whole value chain"? ¹⁶ As one investor summarised his interview on X, "the technology being real & the capex earning its return are 2 different questions." ⁷⁷ We count that money, line by line, and set it against the cost.

The money loop is real, but smaller than it looks

Suppliers funding their own customers is the heart of the bear case, so it now has its own page: The AI Money Loop, with a section on what each of Anthropic's vendors puts in, risks and gets back. The short version follows.

NVIDIA's guarantees come to $108.5B, or 36% of a year's revenue, against 24% for Lucent in 2000. But $105B of that is one capped guarantee on OpenAI's Ohio leases. It covers only a shortfall, starts from 2028, and OpenAI must pay back anything NVIDIA pays out. Leave it out and NVIDIA's guarantees are 1.2% of revenue, and it has paid out nothing so far. Lucent had already lent cash equal to 6% of a year's revenue to carriers that then failed. ^9,10,78

Anthropic's S-1 puts the vendor money in writing. Google, Amazon, Broadcom and AMD have put in or promised about $103B, or $118B if Microsoft's and NVIDIA's reported stakes are added, which is 12 to 15% of a compute bill of roughly $810B. About $77B of it is not yet drawn. The vendors are already well ahead on paper: Amazon's $6.2B of notes are carried at $108.6B. The risk is that everything moves together. If Anthropic's growth slips, the vendors' stake, their revenue from Anthropic and the collateral behind their loans all fall at once. ¹⁸⁴ The other big loop is Amazon's $50B in OpenAI's March round, all of it now paid in. ^80,81 In its October stability report, as press coverage summarises it, the IMF counted 14 such financings since 2023 and named an AI correction as one of three near-term triggers for market stress. ⁸²

Outside revenue covers two-thirds to all of the capital cost, before running costs

The table counts outside revenue at third-quarter 2026 annual rates. Each input has its own row, so you can swap in your own estimates.

Against that, Microsoft, Alphabet, Amazon, Meta and Oracle spent $1,134.9B of cash capex from January 2023 through June 2026. ⁴¹ We turn that into a yearly capital charge: the level yearly payment that would repay the capital over its life and pay a return on it at the cost of capital. Chips are paid off over their working life and buildings over 20 years, each at the cost of capital.

The chart here is now drawn in the data section. It showed: Coverage of the yearly cost by outside money. Bull: 82 percent without Anthropic, 141 percent with. Base: 69 and 118 percent. Bear: 41 and 71 percent..

Each column uses the base outside-money total. The cost leaves out leases, power, staff and the labs' own running costs, so coverage is overstated; Meta alone has $355.7B of signed leases that have not yet started. ¹¹ The capex figures are compiled from filings by a market-data vendor.

So outside revenue covers between 69% and 118% of the yearly capital charge on the build, depending on Anthropic. Neither end is a measure of profit, for three reasons.

• Revenue is not cash left over. The charge already includes the cost of capital, but revenue must also pay for power, staff and research. As an illustration only, if 60% of the $160B were left after those costs, it would cover about 71% of the charge.
• Some of the payoff has no AI revenue line. Lower costs, better ad conversion and customers kept inside the hyperscalers' own businesses are real returns this count misses, beyond the ad share in the table.
• The cost side is rough. It applies one AI share to five companies' total capex and treats everything spent since 2023 as working capacity, including buildings still under construction and spending on ordinary cloud and warehouses.

We show the count without Anthropic because this letter was written with Anthropic's model and its figures are unaudited, not because its revenue should be ignored. A consistent count would include it, net of what it pays the clouds, and that lands between the two ends.

Each year's spending needs new revenue of about 12% of its cost, every year

The count above takes the whole build at once. The same arithmetic can be run on each year's spending. On the table's base terms, each $100B of capex needs about $12B of revenue a year to repay it and pay the cost of capital. So the test for each year is whether outside revenue grew by that much.

Capex for 2023–2025 is cash capex from filings, compiled by a market-data vendor; Oracle's years are its fiscal years. 2026 uses each company's guidance, which for Microsoft and Meta includes finance leases. 2027 spans MarketScreener's consensus for the four largest and the totals Goldman Sachs and Morgan Stanley published, as reported. The charge uses the base column of the table above. ^41,83

Over the past year, the two labs alone added revenue of about twice the 2025 charge. How much customers outside the labs added is not disclosed. OpenAI's run-rate rose about $38B, from $12B in July 2025 to about $50B in September 2026. Anthropic's rose about $60B, from about $7B in October 2025 to $67B, on its own gross basis. ^84,12,85,70 So the newest finished year of spending is paying its way on the margin. The test is the next two. The 2026 and 2027 spending need outside revenue to grow by about $100B next year and $120–140B the year after, on top of closing the existing gap. That means the labs repeating their best year and then beating it, or customers outside the labs starting to carry far more of the load.

The 2030 bill needs growth almost no company has managed

That is the build already finished. The build still planned is far larger. J.P. Morgan puts AI capex at $5.5T through 2030. ⁸⁶ On the same terms as the table, that build would carry a yearly cost of about $950B. For outside revenue to cover it by 2030, 4¼ years from now, the revenue would have to grow 73% a year from today's $93B. Even counting Anthropic's revenue, which starts the count at $160B, it would need to grow 52% a year.

The labs grow that fast today. OpenAI went from about $12B annualised in July 2025 to close to $50B in September 2026, and Anthropic's quarterly revenue rose 45% from Q2 to Q3. ^84,12,70 But the labs are only part of the total, and the advertising and cloud lines in the table grow far more slowly. Growth this fast is also rare for companies this large. Of US-listed companies with $20–80B of yearly revenue at some point between 2004 and 2026, only 1.5% went on to grow 30% a year for the next four years. Only one grew 50% a year: NVIDIA, the seller in this same boom. The sample includes only companies still listed today, so it misses the ones that failed. Treat it as a rough frequency, not a probability.

Our judgment is that outside money covers the 2030 bill on time with about 45% odds counting Anthropic's revenue (40% before its S-1), and 20% without it. The likelier outcome is that it gets there late. That is why we expect the strain to show first at the edge, where the bills are fixed and the cash is borrowed.

One company's plan carries most of that gap. As the FT reported it in September, OpenAI's revenue rises from $36B in 2026 to $350B in 2030, a 77% yearly pace. Compute costs $856B over the period against $840B of revenue, and free cash flow is −$278B. The $122B it raised in March runs out by 2028. ^87,81 That is where we look next.

Can OpenAI pay for what it has ordered?

Through 2027, yes. After that, not on its own. Cash in hand and the rest of the March round carry OpenAI through 2027, and its revenue is on plan for now. For 2028 it needs an IPO, a new round or new terms with its suppliers. Its executives say the constraint is computing capacity, not customers. So we put only about 25% low confidence, 15–40% on it cutting or pushing back at least $100B of contracts by the end of 2027 because it cannot pay, and 10% or less on a failure to pay. Oracle takes the first hit if it does.

Every path to paying for the build runs through the labs' revenue, and the largest single bet is OpenAI's. It is the biggest customer in Oracle's backlog, a large share of Microsoft's and Amazon's, and the tenant behind NVIDIA's largest guarantee. If one company turns strain at the edge into losses for the core's suppliers, it is OpenAI. So we take its position apart: what it has signed, how it pays, how solid its revenue is, and what companies in its position have done before.

$665B of commitments, almost none on its balance sheet

Start with what OpenAI has promised to buy. Its purchase commitments, almost all of them for computing capacity, were about $665B at March 31, according to The Information's review of its draft IPO financial statements. Against that it had no debt and less than $750M of lease liabilities. ⁸⁸ A promise to buy compute in future years is not debt under accounting rules, so anyone who reads only OpenAI's balance sheet misses it. It shows up instead as backlog on its suppliers' books, where it counts as their future revenue. Since March OpenAI has added a 20-year lease on a 4.25 GW campus in Ohio, whose value has not been disclosed. ⁹ The table lists the main contracts.

OpenAI has also signed headline deals with chip makers: about 12 GW with NVIDIA, 6 GW with AMD and 10 GW with Broadcom. These are mostly the same capacity seen from the other side. The chips go into the cloud data centres and leases in the table, so adding the chip deals to the cloud contracts would count the same compute twice. Some of the chip deals are also softer than their headlines. NVIDIA's $100B letter of intent was never binding and became an equity stake of about $30B. ⁹⁵ AMD's deal gives OpenAI a warrant, a right to buy AMD shares, that vests only as the capacity is actually deployed. ⁹⁶ Added together, the cloud contracts come to about $750B. OpenAI's revenue is running at about $50B a year, so it has signed for roughly 15 years of today's revenue.

It has redrawn the plan four times in a year

OpenAI's compute plan was $1.4T in October 2025, $600B in February 2026, about $750B in July and $856B in September. ^97,98,87 Those are plans, not contracts, but they show a company that resizes often and in both directions. The one cut to a signed project, Abilene, was absorbed without a loss: the site went to another tenant. ⁹¹ In the last three months we found no reported cut or delay to an OpenAI contract.

Funded through 2027, not 2028

Next, how OpenAI pays. It had $73B of cash at the end of March, after burning $3.7B in the first quarter. ⁹⁹ Its $122B funding round in March brought in Amazon for $50B, and Amazon's filings show all of it has now been paid in. ¹⁰⁰ SoftBank, the round's other large investor, paid the last of its $30B on October 1 and has repaid the bridge loan it first used, refinancing it with bonds. S&P rates SoftBank BB+, one notch below investment grade. ¹⁰¹ So the money already raised is in hand. What OpenAI still needs is money for 2028 and after. As for spending, OpenAI's February plan burns $25B of cash in 2026, $57B in 2027 and $85B in 2028. The plan the FT reported in September burns $278B in total through 2030. ^97,102

The table sets OpenAI's cash against its own plans, year by year. Cash rose from $40B at the end of 2025 to $73B at the end of March while OpenAI burned $3.7B, so about $37B of the March round had already arrived. That fits Amazon's first $15B and $10B each from SoftBank and NVIDIA. The rest arrived later: Amazon's other $35B, and SoftBank's and NVIDIA's last $20B each. ^{99,100,101,103} The other investors' money, about $12B, is left out, which makes the gap a little larger than it may be.

February plan as reported by The Information via The Decoder; September plan as reported by the FT. Both are OpenAI's projections, not audited figures. ^97,102

So the money in hand carries OpenAI through 2027 with about $70B to spare, and it needs $66–126B more by 2030, depending on which of its own plans holds. The round now being raised, $30B or more, would cover 2028 on the February plan and about a quarter of the gap on the September one. For the rest it needs one of three things: the IPO, a new private round, or new terms with its suppliers. ¹⁰⁴ That timing shapes our call. A company funded through 2027 does not have to cut anything before then. But before its IPO it does have to show investors a plan that pays for 2028, and the cheapest way to do that is to push some of 2028's bills into later years. So if a restructuring comes, the likeliest moment is the second half of 2027, around the IPO.

The new round was reported in late September as $30B or more at a $1.4T valuation. On October 24 Michael Parekh, a former Goldman Sachs internet analyst who writes the newsletter AI: Reset to Zero, citing CNBC and the FT, reported it trimmed to a $1.25T valuation, with commitments due October 30 and demand from wealthy private clients softer than pitched. We have not seen the original reports, so we treat the cut as unconfirmed. If it holds, it is the first sign that private money for OpenAI is getting harder to raise. ¹⁰⁵

The revenue is on plan, for now

OpenAI's revenue figures look inconsistent at first, and the bear case uses that against it. Two numbers get quoted. The first is its run-rate: the revenue of its latest month multiplied by twelve, which was about $50B a year at the end of September. The second is the revenue it actually booked in its accounts. That was $6.7B in the second quarter, which times four is only $27B a year. A reader could take the gap as a sign that the run-rate is padded. It mostly is not. The run-rate grew more than 70% during the third quarter, so in early July it was about $28–29B, roughly the same pace as the second quarter's booked revenue. ^106,13 The two figures differ because one is measured at the end of a quarter of very fast growth and the other is an average over the quarter before it.

A third number adds to the confusion. In late September Axios reported OpenAI's run-rate as nearly $70B. The $20B difference from $50B is a matter of definition. When OpenAI's models are sold through a partner, such as Microsoft's Azure OpenAI service, OpenAI counts only the share of the sale it receives. Anthropic counts the full price the customer pays. Axios's figure counted OpenAI's partner sales the way Anthropic does. So OpenAI's $50B and Anthropic's figures are not measured the same way, and Anthropic's look larger by comparison. In short, all three numbers are right. They measure different things. ¹⁰⁷

On those figures OpenAI is roughly on track for the $36B of revenue its plan sets for 2026. The plan implies about $85B in 2027, which is within reach if OpenAI ends this year with a run-rate of $70B, as it is targeting. The hard years are 2028–30, when revenue has to climb from about $85B to $350B. Two warnings sit against the growth story. ChatGPT's weekly users have stayed near 900M since February, on secondary reports, so new revenue has to come from existing users paying more and from businesses. And the WSJ reported in April that OpenAI had missed its internal targets for users and revenue early in the year, a report the company called "ridiculous". ¹⁰⁸ Anthropic has also filed to go public, so the two will be competing for the same investors' money at the same time.

Anthropic carries a bill almost as large

OpenAI is also not the only lab with a bill this size. Anthropic's public S-1 commits it to spend at least $114.5B with Google, $110.0B with Amazon and $44.9B with Microsoft whatever it uses, plus about $161.2B of rent on Broadcom TPU leases, up to $120.5B with SpaceXAI and other deals, roughly $800B in all. The press had reported $518B or more from the draft. ^184,109,110 So the weak-link question has two names. What separates them is how much revenue each earns from its compute. On Parekh's arithmetic from the filing, Anthropic now earns $1.81 of revenue for each dollar of compute, up from $0.63 a year earlier. ¹⁰⁵ SemiAnalysis adds a reason OpenAI's ratio is likely worse. Flat-rate subscriptions are only about 10% of Anthropic's revenue but use more than 40% of its inference compute, and they are a larger share of OpenAI's revenue. ¹¹¹ A lab that earns more of its money from subscriptions needs more compute for each dollar it takes in.

The two labs also compete for the same customers, so their plans should not simply be added together. Each lab's plan assumes it wins. If the suppliers' backlogs carry both plans in full, they carry more demand than the two can jointly deliver, and the loser's contracts are the ones that get renegotiated.

Overcommitted buyers usually renegotiate before they default

Large buyers whose commitments ran far ahead of their revenue have usually renegotiated long before they defaulted. Whether it ended in a deferral or a default turned on one thing: whether the supplier wanted the customer to survive.

The table points to one lesson. The telecom start-ups defaulted because their suppliers had no reason to keep them alive and capital markets had shut, so nobody was willing to carry them. OpenAI's case looks more like the airlines' after 9/11. Its suppliers are also its shareholders: about two-thirds of the money it raised in 2026 came from Amazon, Microsoft, NVIDIA and Google, all of which sell it compute. A supplier that owns part of its customer would rather push a delivery back than force a default and book a loss on both the contract and the stake. It can often rent the capacity to someone else in the meantime.

Oracle takes the first hit

If OpenAI does cut or delay $100B of contracts, who loses? Spread across the cloud contracts in proportion to their size, the cut would take about $40B from Oracle's backlog, $37B from Microsoft's, $18B from AWS's and $3B from CoreWeave's. None of that would be fatal to any of them. The risk is that a cut would not be spread evenly. Oracle's Stargate contract is the newest, the least built and the most expensive to finance, so it is the obvious place to cut first. A cut of $60–100B there would be 9–15% of Oracle's $664B backlog. That matters more for Oracle than for the others because its credit is already weak. Oracle is rated BBB-, the lowest investment-grade rating, and its five-year CDS trades near 300bp, a level typical of junk-rated companies. ^119,120 Oracle has also already borrowed for the sites due in 2028: about $38B for Texas and Wisconsin and $18B for New Mexico. ¹²¹ So a cut would leave Oracle with debt raised against revenue that no longer arrives on time. SoftBank is the next most exposed, through its stake in OpenAI, the bonds it sold to pay for it, and its own Stargate sites. NVIDIA has nothing at risk on OpenAI before 2028. Its guarantee starts in its fiscal 2029, and it pays only if the Ohio buildings fetch less than their guaranteed value. ⁹

NVIDIA's exposure to the edge as a whole is wider than that one guarantee. SemiAnalysis, reading NVIDIA's latest quarterly filing, counts $530B of obligations that sit off its balance sheet, up from $184B. They are $279B of supply commitments, mostly for memory and 96% due by fiscal 2029; $108.5B of guarantees, almost all of it the Ohio site; $36B of agreements to rent cloud capacity back from its customers; and about $20B of leases it expects to pass on to others. The rental agreements set minimum payments of about $2.35 an hour per GB300 chip, against a market rate of $4.50–4.60, and SemiAnalysis reports the programme is now paused. ¹²² The guarantees, rental agreements and leases add up to about $164.5B. With about $7B of equity in neoclouds, that matches the "more than $170B" of support for neoclouds that The Information reported. ¹⁰⁵ NVIDIA is not only a supplier to the edge. It is partly its guarantor, so if the edge breaks, the losses come back to NVIDIA through these commitments as well as through lost orders.

Why OpenAI may pay in full

The case that OpenAI pays its bills is stronger than its critics allow. Its revenue is on plan through 2027. Capital is still arriving: $122B in March, a bridge round of $30B or more closing now even if at a lower valuation, and a 2027 IPO that could be the largest ever, which would also release Amazon's remaining $35B. Its counterparties own equity in it and would rather extend terms than write anything off. Compute can be re-let if OpenAI does not need it: Abilene's expansion went to another tenant, and Anthropic is signing for capacity wherever it can find it. Some of the obligations already flex by design. AMD's warrant vests only as capacity is deployed. Oracle has moved $75B of its AI backlog to capacity that is prepaid or uses customer-supplied chips, so it carries less of the cost. NVIDIA's guarantee does not start until 2028–29. The weakness of this case is that it rests on investors funding about $278B of cash burn through 2030, at a time when OpenAI's closest rival is asking the same investors for money.

How likely is a renegotiation?

The evidence so far leaves a narrower question than "will OpenAI pay?": will it change the terms first? No model answers that, so what follows is our judgment, and we show how we reach it. We start from the past cases in the table above. Buyers that had ordered far more than their revenue could carry usually renegotiated long before they defaulted, especially when their suppliers wanted them to survive, as OpenAI's do. Then we weigh what is particular to OpenAI.

First, the event, defined so it can be scored later. By December 31, 2027, at least $100B of OpenAI's contracted compute is cancelled or pushed back six months or more, at OpenAI's request, because it cannot fund it on the original terms. Delays caused only by a supplier's construction do not count. Nor does OpenAI moving work onto its own chip or between clouds: that changes who it buys from, not whether it can pay, and counting it would overstate the strain.

What OpenAI's executives said in interviews this autumn moves the odds down. ¹²³

• They are short of capacity, not customers. The CFO, Sarah Friar, says OpenAI under-invested and with hindsight would have bought more compute last year (Big Tech Show, 0:22). The president, Greg Brockman, says "compute is revenue"; before its last launch OpenAI ranked which features to cut to free capacity (Odd Lots, 0:37). A buyer rationing capacity does not ask to push deliveries back.
• The plan has grown lately. It went from $600B in February to about $750B in July and $856B in September.
• The funding plan is equity and then debt, not deferral. Friar describes raising in public markets after an IPO, then moving to convertible bonds and plain debt (Business Post, 0:29).
• Revenue is on plan and the money lasts through 2027. Friar confirmed the $50B run-rate and 77% growth in the third quarter (Big Tech Show, 0:20).

Against that, three things keep the odds well above zero. The executives are describing their own company, unaudited. Friar did not answer when asked about the FT's estimate of about $280B of cash burn over five years (Big Tech Show, 0:19). And the bridge round may have been cut to a $1.25T valuation.

Weighing those, we put a renegotiation forced by OpenAI's finances at about 25%, within a range of 15–40%. We do not put a number on OpenAI moving work to other suppliers, because nothing public gives the size or timing of its own chip. A failure to pay stays at 10% or less, because the cash lasts past the deadline and the suppliers who would have to force a default are also shareholders who would rather wait.

If the edge breaks, this is the shape we expect: a renegotiation, not a default. Its cost lands first on Oracle's growth and credit, and on the chip orders for 2028 behind them, before it reaches the hyperscalers. That is a painful repricing for the edge and for the suppliers' later years. On its own, it is not a crisis for the core.

What would move the odds

These are the readings that would move our 25%, and when each arrives.

Will the biggest buyers keep spending?

Into 2027, yes. Meta has already signed most of its 2027 spending, and SpaceX has become a large new buyer. The strain is in the cash: Meta's free cash flow turned negative last quarter, and SpaceX is financed more like a neocloud than a hyperscaler.

The edge decides who takes the losses. The core decides the chip cycle. If the cash-rich buyers keep raising capex in 2027, the cycle holds even while the edge strains. Two of the largest buyers reported last week. Neither gave a 2027 number, but both said more than a number would.

Meta has signed for heavy 2027 spending, and its cash flow has turned negative

In its third-quarter report, Meta narrowed its 2026 capex guidance to $135–145B, from $130–145B, after spending $34.9B in the third quarter, a figure that includes payments on finance leases. Its free cash flow for the quarter was −$2.1B: it spent more than it earned. Its cash of $82.4B is now below its debt of $83.7B. ¹²⁴ Meta gives its first 2027 guidance on the fourth-quarter call in late January. For now, the CFO said, Meta is "quite early in developing our budget and capacity plans". An analyst on the call recalled management's earlier plan of "maximizing compute in 2026 and 2027" while keeping more room to adjust in 2028. ¹²⁵

Much of Meta's 2027 spending is already contracted. Its quarterly filing lists $454.5B of commitments it cannot cancel, $121.4B of them due in 2027. Most are for rented cloud capacity, servers and network equipment. Some are for energy and Reality Labs hardware, so the figure is not a floor under capex alone. On top of that sit $355.7B of data-centre and energy leases that have been signed but have not yet started. ¹²⁶ Cutting 2027 spending sharply would mean breaking contracts, so we do not expect a sharp cut.

The return on that spending is not yet visible in Meta's profits. Third-quarter operating income fell to $13.0B from $20.5B a year earlier, but that includes a one-off $9.89B legal settlement. Without it, operating income rose 11% to $22.9B. ¹²⁴ So the quarter does not show the spending failing. It shows that the spending is running ahead of the cash it brings in: free cash flow was negative.

For the bull case, this is the firmest support that 2027 capex will grow: the money is already contracted. For the bear case, it is the clearest sign that spending is running ahead of its payback. The test comes in January, and we expect a raise. A raise on its own is not the warning. The warning would be a first 2027 guide with a midpoint above about $198B, roughly 40% above this year's midpoint, without a matching rise in Meta's revenue outlook. That would tell us Meta is spending past the point where the money earns its cost.

SpaceX: a giant new buyer, financed like a neocloud

Meta supports the case for continued spending. SpaceX adds a large new buyer, but one financed by borrowing, so it adds to the financing risk as well. It went public in June, and besides Starlink and its launch business it now runs an AI computing business that rents capacity to others. It guided fourth-quarter capex to $35–40B, against $19.6B in the third quarter. ¹²⁸ That is a pace of $140–160B a year, as much as Meta spends. Its computing capacity will reach 2.5–3 GW by the end of the year, and Elon Musk said he is "highly confident" of adding 10 GW in 2027. ¹²⁹ At our estimate of about $37B per gigawatt, 10 GW would cost about $370B. For scale, SpaceX spent $48.1B on capital in the first nine months of 2026.

The demand behind it is real on today's terms. SpaceX's AI computing revenue rose from $0.8B in the first quarter to $7.6B in the third, and it guides total revenue of $33–37B for the first quarter of 2027. ^128,129 The CFO says new capacity rents for "slightly above the high end of the $30 to $50 per watt range" a year, which would pay back its cost in under a year. That is management's claim, not something we can check. ¹²⁹

SpaceX's customers can leave within about six months; its debt lasts for years. Its computing contracts run for 90 days, and customers can exit with 90 days' notice. In nine months it spent $48.1B on capital against $13.1B of cash from operations, and closed the gap with its $85.7B IPO and $25B of bonds. ^128,129 That is how the edge funds itself, not the core.

The revenue also shows capacity not yet earning. On the arithmetic of Tomasz Tunguz, a venture investor at Theory Ventures, from SpaceX's filings, its contracts price a GPU at about $5.27 an hour, but its AI revenue divided across its installed capacity comes to only $2.44. ¹³⁰ Leaf Ye, a CFA who writes the GrowthStock newsletter, estimates that about 75% of the segment's revenue is short-term rental to a few labs at spot prices, possibly half of it from Anthropic. ¹³¹ If so, part of SpaceX's demand is Anthropic's demand resold, not a separate source.

What this does to the capex question. Chip demand in 2027 no longer rests only on the four hyperscalers. If SpaceX simply holds its fourth-quarter pace through 2027, it adds about $50–75B of spending over 2026's roughly $86B. That makes total 2027 capex more likely to grow, which supports our second call (in the calls table below): 85% that none of the big four plans lower capex for 2027 than for 2026. But it also moves more of the risk to the edge. The buyer adding the most is borrowing to do it, and its customers can leave at short notice. If rents fall, it is the one that cuts first. The scorecard under the verdict tracks both.

The verdict: the core can pay; whether it earns a return is open

Can the builders pay their bills? The cash-rich core can. The debt-funded edge is struggling, and its stocks already trade as if it has broken. Will the build earn its return? That is not settled for anyone yet. Will the stocks pay from today's prices? For the edge, the bad news is largely priced. For the core, the risk is returns that disappoint while AI keeps growing.

A disappointing return turns into a crisis fastest at the edge, because the edge borrowed to build. Its stocks already trade as if that has happened. Oracle, CoreWeave and the other neoclouds borrowed to build capacity that they lease to labs, and the labs' revenue has to grow 50% or more a year to pay for it. OpenAI has signed for more compute than its own plan can fund. The market has seen this. Oracle closed October 23 at $126.16, 62% below its peak close of $328.33 on September 10, 2025. CoreWeave closed at $54.24, 70% below its peak of $183.58 on June 20, 2025. ¹³² So the bubble at the edge is not a forecast. It is deflating now. Two questions remain: whether it has further to fall, and whether it spreads to the core. The edge is tied to the core. NVIDIA has agreed to buy capacity that AI clouds fail to sell and has guaranteed data-centre leases tied to OpenAI. Its filing warns that it may not recover what OpenAI has agreed to repay, promptly or in full. ¹³³

Lenders draw the same line between core and edge. Apollo's president, Jim Zelter, expects debt for AI spending to stay plentiful at least until 2028. ¹³⁴ But Blue Owl, a large data-centre landlord, says it will not lend on a lab's own credit. It needs a guarantor so that it is paid "come hell or high water" (Compound Interest podcast, October 8). Apollo's fixed-income partners also name the point where strain would spread: spreads on the best borrowers widening to 1.5–2 percentage points over Treasuries. Oracle is already at about 2. ¹³⁵ So we expect money to stay cheap for buildings that an investment-grade cloud has leased, and to get dearer first for the labs, the neoclouds and Oracle.

Does the edge have further to fall? We think the price is not yet the bottom: CoreWeave's bonds still yield about 15%, and Oracle's credit is priced like junk. Does the strain spread to the core? History says the edge breaks first, and well before the spending stops. In the telecom bust, defaults started rising in 1999. The Nasdaq telecom index peaked in March 2000. High-yield telecom bond spreads reached 1,200bp in November 2000. Telecom capex fell only in 2001, two years after the first defaults. ¹³⁶ Today's stress at the bottom of the credit market looks more like 1999's rising defaults than November 2000's blowout: the warning stage, not the crash.

We agree with Perkins about the mechanism and about what to do with a portfolio. We disagree about the odds. He writes as if the crash is certain and the only question is when. The section on outside money puts the odds of outside money covering the 2030 bill on time at about 45% counting Anthropic, and 20% without it. The section on OpenAI puts a renegotiation forced by OpenAI's finances at about 25%, and a failure to pay at 10% or less. That is real pain at the edge, but it is not, on its own, a further 40% fall in chip stocks from today's prices. For that we put the odds at about 17% over two years. The section on how far chips could fall shows how we get there.

On Palihapitiya's chokepoints: we agree the suppliers own the scarcest assets, and that is their strength. But chokepoints protect margins, not volumes. If the hyperscalers slow 2027 spending, chip orders fall first and fastest, as they did in every past semiconductor cycle. So on a real cut in investment, the suppliers are hit first.

What would prove us wrong

Both cases make claims that the next few quarters will test. These are the readings that would settle them, with where each stands today. The thresholds are our judgment, set before the October earnings.

Most of these read for the bull case or sit between the two today. Only the credit readings, Oracle's CDS and CoreWeave's bonds, lean to the bear case, and that fits our view that the edge breaks before the core.

One bear reading is likely to trip this week. Our first call puts 70% on Alphabet or Amazon spending more than its operating cash flow in Q3; with Meta already at 103%, that would make two of the four. The funding half of that test is partly met already: Alphabet announced $84.75B of new stock in June, and on October 26 Meta announced a sale of about $35B, with up to $95B planned. ^143,186 We would read that result as the spenders' shareholders taking on more of the risk, which is the risk we think prices carry least, not as a sign the build has stopped paying.

What is the market already pricing?

Most of what the bear case predicts is already in prices. The debt-funded builders are down 62–70%, and chip stocks trade as if this year's earnings boom is the peak: NVIDIA sits at 13 times next year's expected earnings, against 19 for the S&P 500. The broad index prices little fear. The S&P 500 is 2.5% off its record and above its 200-day average, while chip options price swings nearly three times the index's. We put 26.3% on the S&P 500 and the Nasdaq both falling a further 15% within a year, against a 19.5% long-run rate, which is close to the price. The one place we would look for a real gap is NVIDIA, but we have not built the model that would settle it.

Paper gains on AI stakes make Alphabet and Amazon look cheaper than they are

Alphabet and Amazon own stakes in the labs and in SpaceX, and when those companies raise money or go public at higher prices, the gain counts as profit. In the four quarters to June, those gains were about half of Alphabet's pre-tax profit and 46% of Amazon's. In the second quarter alone Alphabet booked $99.0B of gains on equity stakes, mostly SpaceX and an unnamed private company, which added $77.1B to its net income. Amazon booked $53.4B, mostly from its Anthropic shares. ^137,100 These are marks on private or newly listed shares, not cash. For the labs, they come from the same money loop: the labs raise money, spend it on the clouds, and the clouds' stakes in the labs rise in value. Valued on operating profit instead, Alphabet and Amazon trade at twice the multiple their reported earnings suggest.

Last four reported quarters: to June 2026 for Alphabet, Amazon, Microsoft and Meta; to July for NVIDIA; to August for Oracle. Meta's third quarter, which carries a one-off $9.89B legal charge, is not yet included. Free cash flow is operating cash flow less cash capex, which leaves out finance-lease payments. Next year is calendar 2027, except Microsoft (year to June 2027), NVIDIA (year to January 2028) and Oracle (year to May 2027). Figures from filings and consensus estimates compiled by a market-data vendor; the 18% tax rate is our assumption. ¹³⁸

Analysts already strip most of the gains out: their 2027 forecasts are below 2026's for Alphabet, $15.34 a share against $20.58, and for Amazon, $10.67 against $12.86, because they do not forecast new gains. So the next-year column is the fair one. On it, the four big buyers trade at 19–27 times next year's earnings, while their free cash flow, after the build, yields 2.5% or less, half the 5.0% on a ten-year Treasury. The buyers are priced for the build to pay off. Their shareholders, not their lenders, carry the risk that it pays off late.

Where we differ from the price

A view is only worth acting on where it differs from what prices already assume. The bear case in this letter is not news to the market: it is the reason Oracle and CoreWeave have fallen by two-thirds and chip stocks by a quarter. So before saying what to do, we set each of our views against what the price implies. Prices are October 23 closes; the earnings figures are analysts' consensus estimates. ^132,138

NVIDIA's price already assumes a sharp fall in earnings. Analysts expect its earnings to rise 70% next year, yet the stock trades at 13 times them, against 19 for the S&P 500. At the market's multiple, the price fits earnings of about $11.10 a share, 30% below next year's estimate. That multiple is our assumption, not a measurement: a stock at the top of a cycle usually trades below the market's multiple, so part of the discount is normal.

A cut of 30% could come from several places: slower spending in 2028, lower prices and margins as supply catches up, share lost to custom chips, or losses on the support NVIDIA gives its customers. On share, custom chips are gaining: OpenAI's first chip, Jalapeño, was built with Broadcom and is due from late 2026, and Google now sells its TPUs to outside customers such as Anthropic. ^139,140 On spending, the 2027 guides are still rising: Microsoft says its capex will grow in the year to June 2027, and Meta has signed for heavy 2027 spending. ¹⁴¹ NVIDIA's filing shows it has committed to buy $87B of supply in its fiscal 2028 and $88B in fiscal 2029. ¹³³ Those commitments say NVIDIA expects demand to hold, but a company that has signed for supply can be left holding it.

So NVIDIA's price already allows for a large fall in earnings. Whether it allows enough needs an earnings model through a full cycle, which we have not built. Until then we make no call on it.

The edge is largely priced. On a 15 times exit multiple and a 12% yearly return, Oracle's price fits earnings in 2030 about 36% below analysts' estimates. Both inputs are our assumptions. Our odds on OpenAI say how likely a cut is, not how deep, so they cannot tell us whether 36% is too much. Oracle is also funding the build partly by selling stock, and the estimates do not count the shares still to come. So we make no call on Oracle.

If the build keeps growing, the risk moves to the spenders' shareholders. If the October and January guides show 2027 capex rising 20–35%, the chip makers book the revenue while the spenders' cash flow gets worse, and the market has already shown it punishes that: Meta fell 9.4% in the week it raised its range. Whether the suppliers' stocks do better from here depends on how much of that revenue they keep as margin, which we have not modelled. Amazon is the steadiest of the spenders: Andy Jassy puts the payback on a server at about three years, against contracts of about five.

The broad market prices far less

Being right about the bubble does not tell you how far prices fall. So we turn to the broad market: what it already prices, and the paths the next year could take.

The question we test is whether the S&P 500 and the Nasdaq both fall 15% or more at some point in the next 12 months. For a base rate, we took each starting date in the long history of both indices and asked how often such a fall followed within a year. It followed about 19.5% of the time, with a 90% interval of 15.9–24.0%. We then tested 21 rules that might say today is riskier or safer than usual: trend, valuation, credit, volatility, oil, yields and the election cycle. Each rule was fixed before it was run on history, so it could not be tuned to fit. None passed. So history gives no tested reason to move far from 19.5% today. Our own judgment, built from the four scenarios below, is 26.3%. That number is not tested, and we show the arithmetic so a reader can change the weights.

The odds of a large fall are one question; what to expect is a wider one. After the 120 pullbacks like today's since 1928, the S&P 500's median return over the next year was +7.5%, about the same as any year. Semiconductors are where history splits: after past breaks like this one, they kept falling in some years and recovered fast in others. The full distributions are in What usually follows a market like this one.

The chart here is now drawn in the data section. It showed: Probability comparison: tested base rate 19.5 percent, judgment 26.3 percent with a 25 to 30 percent band, sensitivity case 32 percent.

The trend has not broken. The S&P 500 closed at 7,623.68 on October 23, 4.6% above its 200-day average of 7,291.31, the line many investors use to tell a rising market from a falling one. ¹⁴² For a bear market to start, investors need a reason to sell that lasts longer than the dip-buyers' appetite. Usually that is falling earnings, a rising discount rate (the return investors demand, which rises with interest rates) or forced sellers, such as funds meeting margin calls. Here is what we see on each front, in order of how much it matters.

Four ways the next 12 months could go

These are judgment weights, held with low confidence: each could be 10 points higher or lower. The last column names the readings that would move each one.

0.48×10% + 0.22×50% + 0.15×70% + 0.15×0% = 26.3% (27.5% before Anthropic's S-1, with weights of 45% and 25%). If the AI path's weight rose to 40%, taken from churn, the sum would be 33.5%. If the AI path's chance of a fall were 40% rather than 50%, it would be 24.1%.

The popular sell signals fail on history

A widely shared daily signal report called the market "bearish into next week". We tested its six signals back to 1955 where data allow: QQQ and SPY below their 50-day averages, small caps below their 200-day average, consumer staples outperforming, credit spreads widening, a split between stocks making new highs and new lows, and a rising VIX (the market's fear gauge). None passed. A few raised the odds of a fall on paper: staples leading by 1.76 times the normal rate, a rising VIX by 1.23 times. But the effect varied too much from one era to the next to tell from chance. An oil shock came closest to passing, at 1.57 times, but that signal is not on today. The one rule that nearly passed points the other way. After a midterm-election October like this one, a 15% fall followed only 10% of the time, about half the usual rate.

How far could chips fall?

Five measures, from chip-cycle history to option prices, point the same way. We put about 40% medium confidence, 33–50% on chips reaching 40% below the June peak, an ordinary chip-cycle bottom, and about 17% low confidence, 8–25% on a bust that takes them a further 40% lower from here, which in the past has needed a recession.

Two numbers matter for an investor: how likely a big fall is, and how deep it would go. No single method answers either well, so we use five and see where they agree. All are measured from the October 23 close, and chip stocks start this exercise well into a sell-off. The broad market has barely moved. ¹⁷²

So "a 40% fall in chip stocks" can mean two very different things, and we give odds for both.

The chart here is now drawn in the data section. It showed: SOXX ruler: peak $655, October 23 close $478, an ordinary bottom 40% below the peak at $393, a bust a further 40% from today at $287.

Chip history: half the falls this deep reached 40% below the peak

The PHLX Semiconductor Index (SOX) gives us the longest history; SOXX tracked it until 2021 and moves closely with it. It has fallen 20% or more below its two-year high 13 times since 1994, not counting this one. ¹⁷²

The chart here is now drawn in the data section. It showed: Depth of each fall of 20% or more in the PHLX Semiconductor Index since 1994, measured from its two-year high. The recession downturns of 2000 to 2002 fell 74%, 2007 to 2009 66% and 2020 34%. The others ranged from 21% to 49%. The current fall is 25% so far..

• Ten of the 13 got as deep as today's 25%. Half of those ten went on to 40% or more. The median final depth was 42%.
• Chips can fall 45–50% without a recession. The 1995–96 and 1997–98 inventory gluts and the 2022 rate shock all did. That is the chip cycle doing what it does.
• Only the two recessions went past 49%: 2000–02 fell 74% and 2007–09 fell 66%. The bust our 17% describes, 56% below the peak, is deeper than every other fall and shallower than those two. Two cases cannot show that a recession is required, and an investment bust can itself cause one.

Broad bear markets: the deepest came with recessions

Since 1971 the Nasdaq Composite has fallen 20% or more below its two-year high 13 times. The median fall was 30%. Three went past 50%: 1973–74, 2000–02 and 2007–09, and each came with a recession. The S&P 500 has had 11 bear markets since 1950. The seven that came with a recession had a median fall of 36%; the four without one, in 1962, 1966, 1987 and 2022, had a median of 27%. ¹⁷²

Past booms: outside 1929, a doubling did not raise the odds of a crash

The third measure looks at what happened after past booms in any industry, not just chips. Ken French's data split US stocks into 49 industries going back to 1926. We looked for industries whose stocks had doubled in two years, as AI hardware did into this year, and asked how often they fell a further 40% within the next two years. It happened 17% of the time, and 22% for the fastest, most volatile run-ups like today's. But most of that signal comes from the 1929 crash. Leave out 1928–37 and the rate drops to 8%, below the 9.8% chance of such a fall for any industry in any month. Outside the Depression, a doubling has not made a crash more likely.

Option prices: about a one-in-five chance of a chip bust

The fourth measure reads the odds that option prices imply. Put options pay out if a fund falls below a set price, so their prices can be turned into a market-implied chance of a fall. Puts on SMH, a chip fund, price a 22–23% chance that it ends December 2028 at least 40% below today, and 33% for a fall of 30% or more. For QQQ, the Nasdaq-100 fund, the figures are 11–12% and 16–17%. These odds are biased in both directions. Buyers pay extra for protection, so option prices overstate real-world odds. But they price only where the fund ends on that date, not its low along the way, so they understate the chance of touching a level.

Short-dated options tell the same story. On October 23, one-month options on SMH and SOXX priced moves of ±8.0% and ±8.8%, about 2.6 and 2.9 times SPY's ±3.1%. ¹⁶⁸ In all four funds (SPY, QQQ, SMH and SOXX), implied volatility sat 6–7 points above the swings the funds actually made over the past month, and puts cost more than calls. For SMH, SOXX and QQQ, one-month volatility is above three-month volatility. That shape, called backwardation, means buyers fear the next few weeks more than the months after. SPY shows the normal, calmer shape, with volatility low for its past year.

Part of the chip funds' fear is about NVIDIA's results, expected around November 18, near the end of the one-month window. Broadcom and TSMC, which do not report in that window, show the calm shape, with volatility near the bottom of their one-year range. ¹⁶⁸ So the market is paying for protection on chip funds and on tech as a group, not on the broad index or on the largest suppliers one by one.

Our scenarios: about 41% on an ordinary chip bottom

The fifth measure starts from scenarios rather than history. The section on what the market prices weighs four paths for the next year. In the AI financing squeeze, which we weight at 22%, AI stocks fall 30–50% and the index 12–20%. Perkins's scenario plays out a year later: a mild recession, AI stocks down 40% and the S&P 500 down 30% by April 2028. ⁸ Neither path reaches a 2000-style bust in chips. For each path we estimate the chance that chips fall to 40% below their June peak: 90% in the squeeze, 60% if oil and the Fed cause a slowdown, 25% in churn and none in a relief rally. Weighted by how likely we think each path is, that comes to about 41% (43% before the S-1 moved weight from the squeeze to churn).

Combined: 40% on an ordinary bottom, 17% on a bust

The 17% is our judgment, not an average of the five measures. Their figures run from 8% (industry run-ups, leaving out 1929) to 23% (SMH options), and the middle of the run-up, option and chip-history figures is about 20%. We sit below that middle for one main reason. Every past chip fall this deep came with a long recession, and we expect 2027 capex to keep growing (call 2, at 85%), which makes a recession set off by AI less likely. They also measure different things: options price where a fund ends on one date, while history measures the lowest point along the way. We also sit below the chip history on the 40% for an ordinary bottom, for the same reason.

Two cautions apply. Thirteen chip downturns and 11 bear markets are small samples, and the busts cluster in two episodes, 2000–02 and 2007–09. And only the option prices reflect this year's stock sales and credit stress; the historical measures cannot, and both push the odds up.

What this says. The likeliest outcome is that the chip sell-off ends somewhere between today's level and 40% below the peak: a normal, painful chip-cycle bottom, with the broad market down far less. The busts that went that deep have come with a recession. That is why the recession signals we watch, jobless claims and credit spreads, matter as much as the AI news.

It also sets the bar for the portfolio section. After a fall like today's 27%, chip stocks rebounded whenever the cycle held, as in 2018 and 2020. A capex warning matters because it can mark a turn in spending. A falling share price, on its own, does not.

What it means for a portfolio

Keep broad equity exposure, but reduce how much of it rides on AI. The layers of the AI trade do not carry equal risk, so the order in which they break matters more than the size of any one bet. Turn bearish on the market if the S&P 500 closes below its 200-day average, high-yield credit breaks, or the big four cut planned 2027 investment.

Broad exposure. The tested odds of a bear market are close to their long-run rate. Our 26.3% is about seven points above it, a small edge on an untested judgment. In the two paths that carry 63% of our weight, churn and a relief rally, the index ends the year flat or higher, and a seller would have to buy back higher. Perkins makes the same point from the other side: those who shorted the boom too early lost money for years.

Concentration, not exposure. In the AI financing squeeze, AI stocks fall 30–50% while the index falls 12–20%. A portfolio that is mostly AI takes that squeeze at close to full size, so holding the market is not the same as holding an AI-heavy portfolio. As a rough guide, on our arithmetic, a portfolio with about 60% in AI stocks would fall roughly 20–30% in the squeeze, and more with leverage. Its owner should set that against the largest fall they will accept.

• The edge breaks first. Oracle, CoreWeave, the other neoclouds and the private labs carry the most risk, and their prices already show much of it.
• Suppliers are hit first on a real cut. If the first of the big four (Microsoft, Alphabet, Amazon and Meta) cuts its planned 2027 investment, which we put at about 15%, chip orders fall first and fastest. In the likelier case, where the big four guide 2027 capex up 20–35%, the suppliers book that spending as revenue while the spenders' cash flow gets worse.
• For the hyperscalers, the line is rents. The danger is capex staying above operating cash flow while GPU rents turn down: the core borrowing to build capacity that is losing value. The first half of that test is already met: Alphabet and Amazon spent 115% and 119% of their operating cash flow on capex in Q2, and Meta 103% in Q3. So in practice the line is the one in the scorecard: the H100 rental index down 20% over six months.
• The end of the chip sell-off. We would call the chip sell-off over if SOXX closes back above $590, its early-October high.

The hyperscalers earn on compute whichever lab wins, and even in Perkins's crash their credit holds. The suppliers sit one step closer to the edge.

Turn bearish on the market if the S&P 500 closes below its 200-day average; if high-yield spreads close a week above 350bp while investment-grade spreads are above 100bp; or if the big four cut planned 2027 investment. We read a guide as calendar-year investment including leases, compared with what was expected and checked against orders at the suppliers, not as one reported capex line. A cut would also lift the AI path to 40% and our market odds to 33.5%.

Five calls to score us on

These are our judgments, not back-tested rates. Each states what would prove it wrong. Confidence says how firmly the evidence pins the number down, and the range is where we think the true odds could sit. High: reported figures carry most of it. Medium: the evidence points one way, but the outcome turns on decisions not yet made. Low: few precedents, or the measures disagree.

The market lines we watch

The market and economy lines we watch, each set before this week. The AI lines are in the scorecard under the verdict.

Credit and rate readings are FRED closes from October 22–23. ^{151,152,44,155,176}

Earnings week: what each number would tell us

The first test comes this week. For each reading, the table gives what the company guided or analysts expect, and the result that would move us. We read capex against expectations and with leases added back, because three of the four now build partly outside the capex line. Microsoft moved some data-centre leases out of capex in July, which cut its reported 2026 figure from about $190B to $175B with no change in what it builds. Alphabet is renting outside capacity as a bridge, which counts as an operating cost. Meta's Louisiana campus sits in a joint venture that Meta leases. ^177,178

Guidance from each company's last results call; consensus from market data, Bloomberg via moomoo, MarketScreener and broker notes as reported. Consensus is press-reported where marked. ^{177,179,180,181}

What we change. We count a capex reading as a cut if a company lowers a guide it has given, or points to 2027 spending more than about 15% below consensus once leases are added back; the 15% is our judgment. A cut at any of the three triggers the October 28–29 row in the table below: market odds from 26.3% to 33.5% and chip-bust odds from 17% to about 25%. Cloud growth on the bearish side at two of the three would cut our odds that outside revenue covers the 2030 bill on time from 45% to 35%, because it would mean customers outside the labs are not yet taking up the load. Bullish readings at all three leave the call unchanged until Meta's 2027 guide in late January.

What would change our views, by date

Each row is a dated reading, what it would take to move us, and how far. Where the move is in our market odds, it comes from re-weighting the four paths in the section on what the market prices: a 26.3% chance of a further 15% fall in the S&P 500 and the Nasdaq within a year. The other moves are our judgment, set before the readings arrive.

Dates marked "not announced" are estimates from market data and a market-data vendor, read October 25. Earnings dates are from company investor sites where announced. ¹⁸³

Conflict, limits and dates

• Dates. Written October 25, 2026, updated October 26 for Anthropic's S-1 and Meta's stock sale, and rebuilt chart-first the same day. Market data runs to the October 23 close.
• Conflict of interest. This letter was prepared with Claude, an AI model made by Anthropic, which is one of the companies discussed. Anthropic's figures appear in their own row, and every coverage figure is given both with and without them.
• Probabilities. Figures we call tested come from rules fixed before they were run on history; the code and data behind them are not published with this letter. The tested measures estimate different events over different horizons, so they bracket a risk rather than agree on one number. Every other probability here is our judgment, with the reasons stated; none has a track record yet. Each headline judgment carries a confidence label and the range we think it could sit in; the range is itself a judgment, not a measured interval.
• Sources. Anthropic's figures now come from its S-1, where the half-year statements are unaudited interim figures; OpenAI's are still press reports of company statements, not audited. Several sources are paywalled; where so, the list names the readable copy we used. Most broker figures reached us second-hand. Podcast remarks are taken from automatic captions and paraphrased.
• What the coverage count leaves out. On the cost side, power, staff and the labs' own running costs, which revenue must also pay; on the benefit side, savings and better products inside the hyperscalers' own businesses. The count sizes the gap; it is not a measure of profit.
• Not advice. This is research, not a recommendation to buy or sell any security, and it does not account for any reader's circumstances.

Change log

• Oct 25, 2026. First published, as a long-form letter. Market data to the Oct 23 close.
• Oct 26, Anthropic's S-1. Outside money covering the 2030 bill, counting Anthropic, 40% → 45%. The financing-squeeze path 25% → 22% and churn 45% → 48%, so market odds 27.5% → 26.3%. Our paths' odds of an ordinary chip bottom 43% → 41%; the combined 40% is unchanged. The update has the detail.
• Oct 26, Meta's stock sale. No odds changed. A two-week test was set both ways. The update has the detail.
• Oct 26, rebuilt chart-first. The facts and our models are now 26 charts drawn from one data file, with our reading in a separate section and the full argument, tables and sources in this appendix. The two dated updates moved into this log, and the opening summary was folded into the headlines and the reading. Portfolio actions on named companies (trim, no new money, keep) were taken out, because shared notes carry no per-stock ratings; the risk by layer and every fact, odds and source are kept.
• Oct 26, before public release. Four points from a subscription research service whose terms allow internal use only were taken out: its 2027 power-limited supply estimate, its expected rent declines per chip generation (now our own inference, without its figures), its reading of Microsoft's latest backlog growth, and its reports on how the two labs' growth alternates and on Meta's shelved plan to rent out GPUs. No odds changed.

Update, October 26: what Anthropic's S-1 changed

Anthropic's public filing backs its revenue and shows it now pays its way, but its compute bill is larger than the press reported and part of it is financed by its own suppliers. We raise our odds that outside money covers the 2030 bill from 40% to 45%, trim the financing-squeeze path from 25% to 22%, and lower our market odds from 27.5% to 26.3%. Our call stays the same: keep broad market exposure, cut AI concentration.

What the filing shows

• Revenue is real and growing fast. $16.3B in the first half of 2026, against $1.2B a year earlier. Q2 alone was $11.6B, up from $4.7B in Q1, and Q2 made an operating profit of $974M, the first in the filing. ¹⁸⁴ The half-year figures are interim and unaudited, but they sit in an SEC filing, so misstating them carries legal risk that a leak does not.
• It pays its way. Operating cash flow was +$3.9B in the half, the adjusted operating margin went from −159% to +10%, and it held $131.7B of cash and short-term investments at June 30. Compute and infrastructure cost $8.9B, or 55% of revenue, down from about twice revenue a year earlier. We could not find a gross margin line. ¹⁸⁴
• The compute bill is bigger than the draft suggested. Anthropic must spend at least $114.5B with Google (2026 to 2033), $110.0B with Amazon (2026 to 2036) and $44.9B with Microsoft (2026 to 2033), "regardless of actual usage". That is $269.4B owed to the three clouds. On top sit about $161.2B of rent on TPU equipment leased from Broadcom over five years, up to $120.5B with SpaceXAI (cancellable on 90 days' notice from early 2027), about $86B with Nscale and Lambda (paid as capacity arrives), more than $20B with AMD, and an estimated $125–150B of extra TPU capacity for 2028–31. ¹⁸⁴ Added up loosely, that is roughly $800B of commitments of very different strength.
• Suppliers are funding part of it. Google can lend up to $20B more against revenue targets, Amazon has a $15B facility it settles in stock as capacity is delivered, Broadcom will lend up to $42B to finance the rent on its own leases, and AMD can buy up to $5B of stock. ¹⁸⁴ This is the vendor financing the bear case describes, now in writing. The AI Money Loop sets out what each vendor puts in and gets back.
• Some revenue is flattered. 48% of first-half revenue came through the clouds' marketplaces and is booked at the full price the customer pays. ¹⁸⁴

Bull case, check and counter-case

The bull case. The lab with the most aggressive compute plan now covers its running costs from customers, not investors. A quarterly run-rate of $46B in Q2, and $67B in Q3 if the reported $16.8B holds, means the clouds' Anthropic backlog has a payer with cash behind it. The $114.5B Google and $110.0B Amazon minimums are the demand behind this week's capex guides. On X the morning of the filing, finance accounts led with the growth and treated the $80.7B net loss, $72.4B of which is a fair-value mark on convertible notes, as accounting noise. ¹⁸⁵

The check. The revenue and profit claims hold in the filing. The net loss is mostly that non-cash mark. The customer base is broad: no customer above 10% of revenue, and 79% of Q3 revenue came from outside the top five. ¹⁸⁴

The counter-case. On our rough arithmetic, the cloud minimums, the Broadcom rent and SpaceXAI at full rate add up to more than $100B a year by 2027, roughly twice the Q2 pace. Anthropic has to keep growing at a pace few companies have sustained. More than 80% of Q3 revenue came from models released in the past six months, so a stumble on the next model would show up fast. The suppliers financing the build are the same companies booking its orders. And a growing share runs through leased chips in dedicated sites, which is spend that may not flow through cloud backlogs at all.

Net view. Demand risk is lower than we thought on October 25. Concentration and vendor-financing risk are higher. The money now depends more on one listing going well.

How our odds move

Still to come. An amended filing should add the price range, the run-rate as of October 31 and a total for the commitments. The roadshow is reported for the week of November 9. An offering cut or priced below its range would take outside-money odds from 45% to 35% and the squeeze path from 22% to 30%. Strong demand for the shares would take the squeeze path to 20%.

Update, October 26: Meta sells stock to fund the build

Meta is selling about $35B of new stock and convertible preferred, about $40B if the underwriters take their options, and plans a further program of about $60B from early 2027. NVIDIA, AMD and Broadcom have said they may buy up to $5B of it. The biggest single buyer of AI capacity is now funding the build partly with outside equity, which is the financing stage the bear case describes. We change no odds yet. The test is how the stock is received over the next two weeks.

What Meta is selling

• $12.5B of mandatory convertible preferred stock, sold as 250 million depositary shares, each a twentieth of a $1,000 preferred share. It turns into common stock around August 15, 2029. The dividend rate is set at pricing. ¹⁸⁶
• $22.5B of Class A common stock, sold forward. Morgan Stanley and Goldman Sachs sell the shares now, and Meta receives the cash when it settles the forward contracts, which it can do as late as November 19, 2027. So Meta can draw the money as its 2027 bills come due. ¹⁸⁶
• Options and a later program. The underwriters can buy up to $1.875B more preferred and $3.375B more common. Meta also plans an at-the-market program of about $60B in its new Class C shares, not yet signed and not expected to start before the first quarter of 2027. Press reports tally the whole plan at up to $95B. ¹⁸⁶
• Chip suppliers in the book. NVIDIA, AMD and Broadcom have indicated interest in buying up to $5.0B of the common stock, on the same terms as other buyers. The indications are not binding. ¹⁸⁶

What it means for this letter

It confirms the financing stage. Our scorecard's bear reading on capex is two or more of the big four spending above their operating cash flow, with debt or stock sales funding the gap. Meta now meets both halves on its own: capex was 103% of operating cash flow in Q3, and it is now selling stock. If Alphabet or Amazon also spends more than its operating cash flow in Q3, which our first call puts at 70%, the reading trips this week. Alphabet's $84.75B stock package in June already covers the funding half for it. ¹⁴³

It is vendor financing again, at small scale. The chip makers that sell to Meta may now own part of it, as Broadcom and AMD already do at Anthropic. $5B is small against Meta's spending, but it is the same loop the bear case worries about.

It points to more spending, not less. A forward sale that can settle through late 2027 is money raised for next year's bills. That supports our second call, 85% that none of the big four plans lower capex for 2027, and our expectation that Meta raises its 2027 guide in January.

It is a live test of the squeeze path. That path names "equity raises land badly" as one of its triggers. The first reaction is calm. Meta opened at $637.89 on October 26, dipped to $629.02 in the first hour and was back at $643.75 by late morning, against its $642.62 close on October 23. Press coverage reads the choice of equity over debt as prudent. ¹³² One day does not settle it, so we set a two-week test both ways, below.

What it costs Meta's shareholders. On our arithmetic, the $35B now adds about 2% to Meta's share count once the forward sale settles and the preferred converts, about 2.5% with the options, and about 6% with the 2027 program. In return Meta's balance sheet stays at about $84B of debt instead of climbing toward $230B, and its Aa3/AA− ratings are safer. So the raise lowers the risk that Meta cannot fund the build and caps how much each share gains if the build pays.

How our odds move

The test, both ways. If the offering is cut, or Meta closes 10% or more below $642.62 within two weeks of pricing, the squeeze path goes from 22% to 25% and market odds from 26.3% to 27.5%. If it prices at or near the reference price with the options taken and the chipmakers' interest confirmed, and Meta holds within 5% of $642.62 for two weeks, the squeeze path goes from 22% to 21% and market odds from 26.3% to 25.9%. A clean sale shows equity is still open to the biggest buyer, which is the opposite of the squeeze.

Net view. The call is unchanged. Meta's shareholders, not its lenders, are taking on more of the build's risk, which is the risk we think prices carry least.

Sources

1. Goldman Sachs prime brokerage via Bloomberg, Oct 19 (paywalled). bloomberg.com; readable: 24/7 Wall St
2. CNBC, Oct 14, 2026, CoreWeave bonds (did not load). cnbc.com; readable: investingLive
3. Oracle CDS at 295bp, Oct 19, citing CNBC (secondary). primexbt.com
4. The a16z Show, "Amazon CEO Andy Jassy on the $220 Billion AI Buildout", with Ben Horowitz and Erik Torenberg, Oct 23, 2026. Amazon holds stakes in Anthropic and OpenAI. a16z.simplecast.com
5. Big Technology Podcast, Sundar Pichai with Alex Kantrowitz, Oct 20, 2026. Google is an investor in Anthropic. bigtechnology.com
6. US Bureau of Labor Statistics, Employment Situation for September 2026 (released Oct 2), as summarised by First Trust. ftportfolios.com
7. Meta PRE 14C, Oct 21, 2026. sec.gov
8. Dario Perkins, "The Perkini Report", TS Lombard, October 2026. A scenario written as a post-mortem from 2030. hub.tslombard.com
9. NVIDIA 8-K, Aug 17, 2026 (residual value guaranties up to $105B, PORTS campus, Pike County, Ohio). sec.gov
10. Lucent Technologies 10-K405, FY ended Sep 30, 2000 (credit commitments, guarantees, revenue). sec.gov
11. Meta 10-Q, Q3 2026, commitments note (leases not yet commenced $355.7B). sec.gov
12. Financial Times, Oct 8, 2026, OpenAI run-rate close to $50B, and why the $70B differed (paywalled). ft.com; readable: TechCrunch
13. WSJ, Aug 18, 2026, OpenAI Q2 revenue $6.7B (paywalled). wsj.com; readable: SiliconANGLE
14. Ed Zitron, "The AI Demand Bubble", Aug 4, 2026 (relays Barclays, UBS, Wells Fargo and BofA notes). wheresyoured.at
15. Ed Zitron, "The More You Buy, The More You Lose" (citing Nikkei), and Benzinga on the $1.65T figure. wheresyoured.at, benzinga.com
16. Stratechery, "An Interview with Craig Moffett About SpaceX and Bubbles", Oct 22, 2026 (paywalled, not read). stratechery.com; quotes from Whitney Tilson's Daily. The Oct 24 X summary by @TMTMoats could not be retrieved.
17. Microsoft, Form 10-K for fiscal 2026 (OpenAI related-party revenue), and the April 2026 partnership amendment as reported by Redmond Magazine. sec.gov
18. Epoch AI, "The plunging price of thought", Sep 22, 2026 (about 47% a quarter, 13× a year). epoch.ai
19. Silicon UK, Aug 12, 2026, Silicon Data token prices $2.04 to $1.16–1.18 per million tokens. Secondary; the Jefferies note is not public. silicon.co.uk
20. Google I/O 2026 keynote (about 480T to over 3.2 quadrillion tokens a month). blog.google
21. SemiAnalysis, "The Great GPU Shortage", Apr 2, 2026 (one-year H100 contracts $1.70 to $2.35). semianalysis.com
22. Chamath Palihapitiya, "Deep Dive: Why AI is booming, but productivity isn't", Social Capital, Sept 25, 2026 (public part; guide paywalled). research.socialcapital.com
23. Kim, Kim and Koning, "Mapping AI into Production: A Field Experiment on Firm Performance", SSRN, Mar 30, 2026. ssrn.com
24. BLS nonfarm business output per hour (OPHNFB), FRED; growth rates are our calculation from Q4 levels. fred.stlouisfed.org
25. Ramp AI Index, overall adoption (RAMPAIALL), FRED. fred.stlouisfed.org
26. All-In, "CoreWeave crash, AI trade cracks, Anthropic IPO red flags", Oct 17, 2026. The hosts are venture investors; Ben Horowitz said a16z holds OpenAI and Anthropic. libsyn.com
27. Monetary Matters with Jack Farley, "Jim Bianco: Why I'm Finally Buying Bonds After Six Years", Oct 18, 2026. show page
28. Monetary Matters with Jack Farley, "Is the AI Trade Over?", Oct 20, 2026. show page
29. Meta Platforms Q3 2026 earnings call, Oct 21, 2026 (from automatic captions; check figures against the 10-Q). feed
30. 20VC, "Trillion or Bust", with Rory O'Driscoll and Jason Lemkin, Oct 22, 2026. libsyn.com
31. Plain English, "AI is smarter than ever, so where's the revolution?", Derek Thompson with Ethan Mollick, Oct 20, 2026. theringer.com
32. Semafor, Satya Nadella with Reed Albergotti, Oct 21, 2026. semafor.com
33. "AI Bubble: what the X bookmarks say", our digest of AI-bubble posts, Oct 24, 2026. notion
34. Freda Duan (@FredaDuan) on X, Oct 17, 2026. x.com
35. Ken Griffin, remarks of Oct 21, 2026, as relayed by @mollysoshea and @deitaone on X. x.com, x.com
36. @realnickmugalli on X, relaying Brad Gerstner on his CoreWeave sale. x.com
37. X signal daily, Oct 23, 2026, our digest of the day's market posts. notion
38. Meta Q3 2026 release (revenue $64.6B; ads +26%; 2026 capex $135–145B). sec.gov
39. Oracle CDS at 295bp, Oct 19, 2026, citing CNBC. Secondary; the dealer quote itself was not seen. primexbt.com
40. Oracle FQ1 FY27 release ($20B of stock sold through an ATM). sec.gov
41. Financial statements via a market-data vendor, compiled from filings: capex, operating cash flow, revenue, Oracle net debt to EBITDA, and NVIDIA's quarterly revenue. Filings behind them: MSFT 10-K FY26, GOOGL 10-Q Q2, AMZN 10-Q Q2, ORCL 10-Q FQ1 FY27.
42. CNBC, Oct 14, 2026, CoreWeave bond yields (did not load). cnbc.com; readable: investingLive
43. CoreWeave Q2 2026 release (revenue $2.58B; backlog about $104B; debt lines summing to $35.1B). sec.gov
44. ICE BofA CCC & Lower OAS (BAMLH0A3HYC), FRED. fred.stlouisfed.org
45. Sources, "Sundar Pichai on Gemini 4 and the AI compute crunch", with Alex Heath, Oct 21, 2026. sources.news
46. SemiAnalysis, Ep. 039, "Our numbers are insane, here's the proof", Oct 20, 2026. spotify
47. SemiAnalysis, "Datacenters 2027: The Year of Acceleration, +50GW in a Single Year", Oct 18, 2026 (free part read; paid). newsletter.semianalysis.com
48. SemiAnalysis, "Everyone Says Datacenter Moratoriums Are Killing the US Buildout. We disagree", Sep 15, 2026 (free part read; paid). newsletter.semianalysis.com
49. Microsoft FY26 Q4 earnings call ("RPO increased 25% when excluding OpenAI"; 100,000 Foundry customers). microsoft.com
50. Microsoft FY26 Q4 release (commercial RPO $678B, +84%; over 30M paid Copilot seats). sec.gov
51. Sundar Pichai, Q2 2026 remarks (Cloud backlog $514B; nearly 500 customers over 1T tokens). blog.google
52. Alphabet Q2 2026 release (Google Cloud $24.8B, +82%; revenue $119.8B). sec.gov
53. Amazon Q2 2026 release (AWS +36.7%; AI run-rate over $25B). sec.gov
54. Palantir Q2 2026 release (US commercial +149%). sec.gov
55. Dealroom, Jun 9, 2026 (Cursor over $4B annualised). dealroom.co
56. Meta Q2 2026 call transcript (15.7% conversion uplift on Facebook). investor.atmeta.com PDF
57. Chamath Palihapitiya, "Deep Dive: Where Value Accrues in the AI Stack", Social Capital, May 1, 2026 (public part read; the 138-page report is paywalled). research.socialcapital.com
58. DrPeering, internet transit pricing, historical (informal series). drpeering.net
59. Silicon Data H100 rental index (live $2.88 on Oct 24; $2.00 on Dec 9, 2025; covers neoclouds only). silicondata.com, baseline
60. Silicon Data, H100 rental prices by region (Q3 2026), July 31, 2026. silicondata.com
61. Crypto Briefing, Oct 15, 2026, CME Group brings forward-curve pricing to Silicon Data's GPU rental indices; both curves reported in backwardation. cryptobriefing.com
62. DatacenterDynamics, Sep 11, 2026, on Oracle's call (97.9% utilisation; 20% premium). Secondary; Oracle's own transcript could not be opened. datacenterdynamics.com
63. @para_capital on X, Oct 18, 2026, top-ten valuations now and in March 2000. x.com
64. @unclealpha007 on X, Oct 24, 2026, net debt to EBITDA of the largest spenders. x.com
65. Reuters via The Star, AI data-centre race builds $1 trillion lease burden for Big Tech, Aug 5, 2026; Microsoft figure via Bloomberg Government. thestar.com.my news.bgov.com
66. Aakash Gupta (@aakashgupta) on X, Apr 19, 2026, capex against GDP. x.com
67. Cape Fear Advisors, "Adding It Up: The Asset Stays, the Owner Changes", Oct 13, 2026. capefearadvisors.substack.com
68. Matt Dratch (@DratchCap) on X, Oct 18, 2026, replying to Freda Duan on revenue needed per gigawatt. x.com
69. Bloomberg, Oct 15, 2026, Anthropic Q3 revenue $16.8B; and Axios. Bloomberg Government, axios.com
70. Anthropic, Google and Broadcom compute post, Apr 6, 2026 (over 1,000 customers at $1M+). anthropic.com
71. Ramp AI Index, October 2026 (adoption 55.9%; 80% of lab spend from 1% of firms). ramp.com
72. Tomasz Tunguz, "How Large is the AI Circle?", Oct 23, 2026, citing Ramp data via Apollo and PitchBook-NVCA. tomtunguz.com
73. Microsoft FY26 Q3 release and call (AI run-rate over $37B, +123%). sec.gov, call
74. Gartner, research abstract, Oct 13, 2026 (full report paywalled). gartner.com
75. Conch Shell Capital (@TMTMoats) on X, Oct 24, 2026, summarising Craig Moffett on Stratechery. x.com
76. NVIDIA 10-Q, Q2 FY27 (guarantees $3,529M; equity securities $51.2B non-marketable, $42.8B marketable). sec.gov
77. NVIDIA 10-Q, Q2 FY27, commitments and contingencies note ($36B AI-cloud capacity, $20B leases to reassign, $25B equity commitments, $279B supply). sec.gov R17
78. Bloomberg Law, Apr 1, 2026 (Amazon's $50B, $35B contingent). bloomberglaw.com
79. OpenAI, "Accelerating the next phase of AI", Mar 31, 2026 ($122B at $852B post-money). openai.com
80. IMF, Global Financial Stability Report, October 13, 2026, as reported by Telangana Today and Securities.io; we have not read the report itself. telanganatoday.com securities.io
81. 2027 capex estimates: MarketScreener consensus pages for Alphabet, Amazon, Meta and Microsoft (Oct 23, 2026); Morgan Stanley and S&P totals via market data, Oct 22, 2026. marketscreener.com, the vendor's site
82. The Information via Reuters, Jul 30, 2025 (OpenAI about $12B annualised). Market data
83. Reuters, Oct 2025, Anthropic's run-rate nearing $7B; via market data. the vendor's site
84. J.P. Morgan Global Research, AI capex ($5.5T through 2030). jpmorgan.com
85. Financial Times, Sep 18, 2026, OpenAI projections (paywalled). ft.com; readable: Reuters via Investing.com
86. The Information, review of OpenAI's draft IPO financials: about $665B of purchase commitments, no debt, under $750M of lease liabilities at Mar 31, 2026 (paywalled); readable: The Next Web. thenextweb.com
87. WSJ, Sep 10, 2025, OpenAI–Oracle $300B cloud contract (paywalled); readable: SiliconANGLE. siliconangle.com
88. Bloomberg, Dec 12, 2025, Oracle delays some OpenAI data-centre projects to 2028 (Oracle denied it); copy via Bloomberg Law. bloomberglaw.com
89. Bloomberg, Mar 2026, Oracle and OpenAI scrap the 600MW Abilene expansion (paywalled); readable: Tom's Hardware. tomshardware.com
90. Amazon, Form 10-Q for the quarter to Jun 30, 2026 (OpenAI commitments of $138B). sec.gov
91. CoreWeave, Form 10-K for 2025 (OpenAI contracts). sec.gov
92. Axios, Apr 20, 2026, on Cerebras's S-1 (OpenAI agreement and $1B loan). axios.com
93. WSJ, NVIDIA halts plan to invest $100B in OpenAI (paywalled); copy via Bloomberg Law. bloomberglaw.com
94. AMD, Form 10-Q for the quarter to Jun 27, 2026 (OpenAI warrant, none vested). sec.gov
95. The Decoder, OpenAI's February 2026 cash-burn forecast by year. the-decoder.com
96. OpenAI lifts planned compute spending, market data, 2026. the vendor's site
97. The Information, OpenAI's cash and first-quarter burn (paywalled); readable: PYMNTS. pymnts.com
98. Bloomberg, OpenAI valued at $852B after closing a $122B round; copy via Bloomberg Law. bloomberglaw.com. Amazon, Form 10-Q for the quarter to June 30, 2026: $13.7B of the $35B commitment invested in Q2 and the remaining $21.3B after the quarter. sec.gov
99. SoftBank Group, Execution of follow-on investment (third tranche) in OpenAI, October 1, 2026: $30B completed, funded with senior notes; bridge facility repaid and cancelled. group.softbank
100. Financial Times, Sep 2026, OpenAI forecasts cash burn near $280B by 2030 (paywalled); readable: The Star/Reuters. thestar.com.my
101. Dealroom, Oct 1, 2026, NVIDIA and SoftBank make the final $10B payments on their $30B OpenAI pledges. dealroom.co
102. Bloomberg, Oct 2026, OpenAI raise talks at $1.4T (paywalled); readable: Taipei Times. taipeitimes.com
103. Michael Parekh, "AI: Anthropic's $750B Compute Bill, OpenAI's Trim, Nvidia's Safety Net & More", AI: Reset to Zero #1249, Oct 24, 2026, citing CNBC, the FT and The Information; we could not find the underlying reports. michaelparekh.substack.com
104. Reuters, Oct 2026, OpenAI's annual revenue pace and its Q3 growth; market data. the vendor's site
105. TechCrunch, Oct 8, 2026, why OpenAI's $50B run-rate differs from the $70B reported earlier (net versus gross). techcrunch.com
106. WSJ, Apr 27, 2026, OpenAI falls short of internal revenue and user targets (paywalled); readable: market data. the vendor's site
107. The Next Web, Oct 2026, on the WSJ's report of Anthropic's draft S-1 ($750B of possible compute, about $450B mostly non-cancellable); Reuters reported at least $518B. thenextweb.com
108. IBTimes UK, Sep 29, 2026, relaying Reuters' review of Anthropic's confidential draft prospectus: at least $518B of commitments, about 80% non-cancellable or payable regardless of use. ibtimes.co.uk
109. SemiAnalysis, "Anthropic Subscriptions Offer 5x+ More Value Than OpenAI", Oct 5, 2026 (free part read; paid). newsletter.semianalysis.com
110. CNN Money, Jan 17, 2002, Delta defers aircraft deliveries. money.cnn.com
111. Oil & Gas Journal, settlements reduce take-or-pay exposure. ogj.com
112. AP, WeWork's lease amendments before and during bankruptcy; via The Republic. therepublic.com
113. EE Times, on McKinsey's "Caveat Vendor" study of telecom vendor financing. eetimes.com
114. Fortune, Oct 15, 2010, Lucent's vendor financing and Winstar. fortune.com
115. Benzinga, Global Crossing's bankruptcy and sale. benzinga.com
116. Oracle, Form 10-Q for the quarter to Aug 31, 2026 (remaining performance obligations). sec.gov
117. Wealth Professional, Oct 2026, Oracle's default insurance hits high-yield levels; ratings from S&P and Moody's. wealthprofessional.ca
118. DatacenterDynamics, JPMorgan and MUFG lead a $38B debt package for Oracle-linked data centres. datacenterdynamics.com
119. SemiAnalysis, "Nvidia's Backstop Universe", Sep 11, 2026 (free part read; paid). NVIDIA's off-balance-sheet obligations from its 10-Q. newsletter.semianalysis.com
120. OpenAI executives on podcasts, Sep 14 to Oct 22, 2026: Sarah Friar (CFO) on The Big Tech Show (Irish Independent, Oct 20), Pioneers of AI (Oct 22) and The Business Post Connected (Oct 14); Greg Brockman (president) on Odd Lots (Sep 14) and No Priors (Oct 15); Sam Altman on Fair Game (Oct 6) and Stanford GSB View From The Top (Oct 21). Quotes are from automatic transcripts, attributed from context; cited by show and minute.
121. Meta, Q3 2026 results, earnings release (Form 8-K, exhibit 99.1). sec.gov
122. Meta, Q3 2026 earnings call, remarks by CFO Susan Li and the question from J.P. Morgan's Doug Anmuth. investor.atmeta.com
123. Meta, Form 10-Q for the quarter to September 30, 2026, commitments and lease notes. Return on net new investment is our calculation from its line items: the year's after-tax operating income gain over the prior year's capex less depreciation. sec.gov
124. SpaceX, Q3 2026 results, Form 8-K and earnings release, Oct 20, 2026. sec.gov
125. SpaceX, Q3 2026 earnings call, Oct 20, 2026, transcript; per-gigawatt cost is our estimate from its reported capital and capacity. benzinga.com
126. Tomasz Tunguz, "The GPU Revenue Ladder", Oct 21, 2026; the SpaceX figures are his arithmetic from its filings. tomtunguz.com
127. Leaf Ye, "SpaceX (SPCX) Q3 2026 Earnings Analysis", GrowthStock, Oct 20, 2026; the revenue mix is the author's estimate. leafyecfa.substack.com
128. Market data, closes on October 23, 2026 for ORCL, CRWV, NVDA, AVGO, MU, TSM, AMZN, MSFT, META and GOOGL, split-adjusted, read October 25, 2026; drawdowns are our arithmetic from the peak close.
129. NVIDIA, Form 10-Q for the quarter to July 26, 2026, Note 10, "Future commitments by fiscal year": supply and capacity commitments of $92B (fiscal 2027), $87B (2028), $88B (2029), $279B in total, mostly memory and manufacturing capacity. sec.gov
130. Market data, May 4, 2026, Apollo president Jim Zelter on funding AI spending. the vendor's site
131. Apollo, The Allocation podcast, Fixed Income: The AI Issuance Boom and What it Means For Portfolios, Sep 16, 2026 (about 0:12 and 0:25). podcasts.apple.com
132. Telecom history: Richmond Fed Economic Quarterly, Fall 2003; Antoniewicz, Federal Reserve Board, for the BIS CGFS.
133. Alphabet, Form 10-Q for the quarter to June 30, 2026, Note 3: $99.0B net gain on equity securities, adding $77.1B to net income; coverage in Fortune, July 22, 2026. fortune.com
134. A market-data vendor, consensus analyst EPS estimates by fiscal year, read October 25, 2026: NVIDIA $9.29 (year to January 2027) and $15.81 (to January 2028); Broadcom $19.29 (to November 2027); Micron $175.29 (to September 2027); Oracle $8.14 (to May 2027) and $11.02 (to May 2028). Multiples are our arithmetic.
135. TechCrunch, "OpenAI unveils its first custom chip, built by Broadcom", Jun 24, 2026. techcrunch.com
136. Tom's Hardware, "The custom AI ASIC state of play", May 2026, on Google selling TPUs to outside customers. tomshardware.com
137. Nasdaq, "Microsoft Q4 earnings call highlights", Jul 2026: Amy Hood said fiscal 2027 capex will grow year over year. nasdaq.com
138. S&P 500 (^GSPC) and QQQ daily closes, via a market-data vendor. The 200-day average is our own calculation from 200 closes; the QQQ 20-week EMA is approximate.
139. Alphabet equity pricing release, Jun 2, 2026 ($84.75B). sec.gov
140. Oracle FQ1 FY27 release ($20B ATM). sec.gov
141. Renaissance Capital IPO statistics, and its 2025 review. renaissancecapital.com, 2025 review
142. Anthropic, confidential draft S-1 announcement. anthropic.com. The timing is press reporting.
143. IEA Oil Market Report, October 2026, and EIA spot prices, monthly and daily. iea.org, eia.gov monthly, eia.gov daily. The SPR level is from press reports of EIA's weekly data.
144. FOMC statement, Sep 16, 2026. federalreserve.gov
145. CME FedWatch, as quoted by Phemex and Mitrade (secondary).
146. Fitch, US private credit default rate, via Investment Executive, Sep 14, 2026. investmentexecutive.com
147. ICE BofA US High Yield OAS (BAMLH0A0HYM2), FRED. fred.stlouisfed.org
148. ICE BofA US Corporate OAS (BAMLC0A0CM), FRED. fred.stlouisfed.org
149. Chicago Fed NFCI, FRED. fred.stlouisfed.org
150. Market data Earnings Insight, Oct 23, 2026. the vendor's site PDF
151. 10-year TIPS yield (DFII10), FRED. fred.stlouisfed.org. The 10-year Treasury at 5.00% is from the Treasury par curve via market data.
152. Shiller CAPE, multpl. multpl.com
153. Aswath Damodaran, implied ERP. stern.nyu.edu
154. BLS Employment Situation, September 2026. bls.gov
155. BLS CPI, September 2026. bls.gov
156. University of Michigan sentiment, October final, via Trading Economics (secondary). tradingeconomics.com
157. @ren_stocks on X, hedge-fund tech selling, Oct 5–16, citing Goldman prime brokerage. x.com
158. X signal daily, Oct 24, 2026, including CNN Fear & Greed at 21 via @TheProfInvestor and Goldman short leverage. notion, x.com
159. Goldman CTA model, via Tickmill (secondary; the note is client-only). tickmill.com
160. BofA Global Fund Manager Survey, Oct 13, 2026 (client-only), via Crypto Briefing.
161. NAAIM Exposure Index. naaim.org. The live value sits behind a login; 87.7 is from Cain's Market sentiment page.
162. FINRA margin statistics (September debit balances $1,480B). finra.org
163. Investing.com, record $7B weekly inflow to SOXL, Oct 19. investing.com
164. Unusual Whales, volatility pages for SMH, SOXX, QQQ, SPY, NVDA, AVGO and TSM, and SPY gamma exposure, snapshot of Oct 23, 2026. unusualwhales.com
165. Term premium: Kim-Wright (THREEFYTP10), FRED. ACM from the NY Fed, read via a third-party copy.
166. Reserve balances (WRESBAL) and overnight reverse repo (RRPONTSYD), FRED. WRESBAL, RRPONTSYD
167. Monthly Treasury Statement, September 2026, via a third-party copy. PDF
168. Price data and our calculation: daily closes for SPY, QQQ, SMH and SOXX from Jan 2, 2025 to Oct 23, 2026, and weekly closes of the PHLX Semiconductor Index (SOX) from May 1994 and of the Nasdaq Composite from 1971, via market data. A downturn is a fall of 20% or more below the two-year high, ending when the index is back within 10% of it. S&P 500 bear markets are from our study of S&P 500 closes since 1927 (market data ^GSPC), with US recessions as dated by the NBER. the vendor's site
169. Amazon 10-Q, Q2 2026 (performance obligations about $496B; OpenAI and Anthropic expansions). sec.gov
170. WSJ, Jun 3, 2026, citing J.P. Morgan: over 60% of 2027 capacity not under construction (paywalled). wsj.com; readable: edwardconard.com
171. Bloomberg, Oct 9, 2026, OpenAI expects $70B or more annualised by year end; raise at $1.4T pre-money (paywalled). bloomberg.com; readable: Bloomberg Línea
172. Initial claims (ICSA), FRED. fred.stlouisfed.org
173. Microsoft, FY26 Q4 earnings call, July 29, 2026 (capex and Azure guidance, the lease reclassification); FY27 segment and investor metrics, Form 8-K, September 2026. fool.com, sec.gov
174. Alphabet, Q2 2026 earnings call (capex guide, third-party capacity); Bisnow on Meta's Hyperion joint venture. investing.com, bisnow.com
175. Bernstein note on Azure, Oct 19, 2026, as reported by KuCoin. kucoin.com
176. Google Cloud consensus (Bloomberg), as reported by moomoo, Oct 23, 2026. moomoo.com
177. BNP Paribas on AWS, as reported by Benzinga, Oct 2026; Amazon, Q2 2026 earnings call (2026 capex guide). the vendor's site, investing.com
178. Bloomberg, Oct 2026, Anthropic could start its IPO roadshow the week of Nov 9; as reported by Implicator. A public filing must be on EDGAR at least 15 days before the roadshow. implicator.ai
179. Market data earnings and economic calendars, retrieved Oct 25, 2026: Microsoft and Alphabet Oct 28 and Amazon Oct 29 (confirmed); estimated dates for CoreWeave, NVIDIA, Broadcom, Oracle, Micron and Meta; US data releases. the vendor's site
180. Anthropic, PBC, Form S-1, filed October 26, 2026: quarterly and half-year results, liquidity, contractual obligations and infrastructure arrangements, and the 2026 note purchase agreements with Google, Amazon, Broadcom and AMD. sec.gov
181. "Anthropic S-1: X sentiment check (Oct 26, 2026)", our scan of 19 posts on the morning of the filing, with key figures checked against the S-1. notion
182. Meta Platforms, free writing prospectus, October 26, 2026 (Rule 433): depositary shares on mandatory convertible preferred, Class A common under forward sale agreements, the planned Class C at-the-market program and the indications from NVIDIA, AMD and Broadcom. Press tallies of up to $95B from Barron's and Investopedia. sec.gov