At the end of July, on Microsoft's fourth-quarter earnings call, Amy Hood spent about ninety seconds on something she introduced as housekeeping. Starting in fiscal 2027, the company would depreciate its data centres and office buildings over twenty-five years rather than fifteen. A few sentences later came the second change: more of Microsoft's future data centre leases would be treated as operating leases rather than finance leases.
The two items belong together. Microsoft counts finance leases in capital expenditure. Operating leases run through operating cash flow instead. So the accounting change did something to the headline number: calendar-2026 capex guidance, which had been about $190 billion in April, came down to about $175 billion.
Then Hood said the sentence that ought to be framed and hung in a business school.
"Outside of this useful life impact, our calendar year 2026 CapEx investment expectations remain unchanged."
The guide fell by $15 billion. The spending did not move at all.
If you want a single artefact to explain how the largest capital programme in corporate history is being administered, that is it. Nothing was concealed. Two adjustments, each with a defensible rationale, produced a smaller number and a longer horizon, and management said plainly on the record that the underlying commitment was untouched.
It is one of dozens of such artefacts. Over the past year I have been reading the filings, the transcripts, the rating agency reports and the bond prospectuses of the companies building artificial intelligence, looking for the gap between what they are spending and what they are saying. The gap is not hidden. It is disclosed, quarter by quarter, in language that has been carefully arranged so that no individual ever has to be the one who admits the bet was wrong.
Table of Contents
Knee-deep
In 1976, the organisational psychologist Barry Staw ran an experiment that has been replicated, extended and argued over ever since. He gave business students a hypothetical allocation of research and development funds, then told them the project had failed. Some were told they had chosen the project themselves; others were told someone else had. Asked how much more they would invest, the students who felt personally responsible put in more money than the ones who had inherited the mess.
Staw's paper was called "Knee-deep in the big muddy", and its finding is one of the more reliable in the study of organisations: people escalate their commitment to a failing course of action in proportion to how personally responsible they feel for it. The mechanism is not stupidity. It is self-justification. Having defended the decision publicly, the decision-maker needs the decision to have been right, and the cheapest way to make it right is to keep going.
Prospect theory adds the second half. Losses loom larger than equivalent gains, and a certain, immediate loss is avoided more energetically than a probabilistic, delayed one. For a chief executive, the arithmetic is brutally simple. Admitting the mistake now produces a certain loss — the share price, the board, the profile in the business press, the career. Continuing produces an uncertain loss that may land in three years, may be smaller, and may land on somebody else.
That asymmetry is the whole story. Everything else is technique.
I want to be careful here, because the lazy version of this argument is a conspiracy theory, and the evidence does not support one. Most of the people involved appear to believe what they are saying. The infrastructure may well be useful; some of it will be. The point is not that five chief executives met in a room and agreed to lie. The point is that a set of individually defensible decisions — an accounting revision here, a joint venture there, a reclassification, a reframing — has aggregated into a system whose main function is to postpone a sentence.
And the system is working. The sentence has not been said.
The size of the bet
Start with the number, because the number is the part that is not in dispute.
The five largest hyperscalers — Amazon, Microsoft, Alphabet, Meta and Oracle — are forecast to spend roughly $602 billion on capital expenditure in 2026, up about 36% on 2025, according to CreditSights. J.P. Morgan puts global AI capex at $900 billion in 2026 alone and $5.5 trillion by 2030. Goldman Sachs thinks global AI investment passes $1 trillion in 2026, including $581 billion in the United States. Futurum counts $660–690 billion for the big spenders; MUFG says more than $600 billion, three-quarters of it AI compute.
The Bank for International Settlements put the two-year figure at over $1 trillion for the five largest hyperscalers in 2025 and 2026, and noted that AI-related investment contributed roughly one percentage point to US real GDP growth in 2025. Goldman's cumulative estimate for the big four across 2025–27 is $1.15 trillion, more than double the $477 billion spent in the previous three years combined. Bank of America's credit strategists calculated that this would consume 94% of operating cash flow after dividends and buybacks.
In the third quarter of 2025, the same five companies reported a combined $106 billion of capital investment across their most recent quarters — the first time the group had crossed $100 billion, and a 73% increase on the year before.
The debt is following the capex. Hyperscaler gross bond issuance is expected to reach a record $420 billion in 2027, up 60% from 2026, according to Goldman data reported by Reuters. Vanguard puts total AI-related debt issuance for 2026 at $300–570 billion; Goldman tracks nearly $500 billion. Consensus capex for 2027 is already $1.1 trillion.
These are not numbers that can be quietly walked back. They are commitments with counterparties, power contracts, land options, fab allocations and construction schedules attached. Which is precisely why the interesting question is not whether the spending is large. It is what happens inside an organisation that suspects the spending will not pay.
What would have to be true
The bull case has a shape, and it is worth stating fairly before examining what has been done to avoid testing it.
For the buildout to earn its cost of capital, AI revenue has to grow from something like $60 billion a year to somewhere between $600 billion and $2 trillion by 2030, depending on which scenario you pick. Bain & Co. has put the required annual figure at roughly $2 trillion, rising toward $4–6 trillion by the early 2030s in its more aggressive scenarios (Bisnow, Network World). Sequoia's David Cahn calculated that the ecosystem needed $600 billion a year against perhaps $50–100 billion actually being generated.
The coverage ratio is the cleanest way to see the problem. In 2025, direct AI revenue across the ecosystem was roughly $40–60 billion against AI-specific capex of about $300 billion — a ratio near 0.15x. By 2026, with AI revenue perhaps $80–120 billion against $450 billion of capex, it improves to roughly 0.25x. Sequoia's arithmetic implied a ratio that would not sustain a business; the trend line improves it without closing it.
Other estimates are harsher. The economist Stijn Van Nieuwerburgh has argued that at least $3.5 trillion a year of revenue is required to fund the facilities being built. Gartner forecasts $8.5 trillion of AI revenue needed by 2030 against $3.6 trillion likely to materialise. MIT's Daron Acemoglu, who won the Nobel prize in 2024, projects a total GDP increase of just 1.1–1.6% over ten years — about 0.05% a year — and estimates that only around 5% of economic tasks are cost-effectively automatable at current prices. Goldman's Jim Covello made the same point more bluntly in 2024: replacing low-wage jobs with tremendously costly technology is "basically the polar opposite" of previous technology transitions.
Then there is the most uncomfortable witness. In February 2026, on the Dwarkesh Podcast, Dario Amodei — who is building some of this infrastructure himself — described the failure mode with unusual precision:
"If my revenue is not $1 trillion, if it's even $800 billion, there's no force on Earth, there's no hedge on Earth that could stop me from going bankrupt if I buy that much compute."
He was explaining his own discipline relative to peers. The companies spending three times what Anthropic spends apparently believe they have found the hedge he could not.
None of this proves the bet fails. It defines the hurdle. And a hurdle that high is exactly the kind of thing an organisation is tempted to stop measuring.
Except that in June 2026, something odd happened: the measurement started to work. Bloomberg calculated that global AI sales outside China reached $25 billion in the first quarter, against roughly $21 billion in estimated quarterly depreciation across the industry. For the second consecutive quarter, AI revenue had edged past the depreciation line. It was, as Bloomberg put it, a pretty thin margin — but it was the first time the industry could point to a number rather than a narrative. The bull case acquired a data point. It also acquired a much more precise definition of what "working" would have to mean.
Moving the clock
Which brings us back to depreciation, the quietest lever in corporate finance and the one most exercised in this cycle.
A server's useful life is an estimate. Stretch it, and the same capital expenditure produces a smaller annual charge against earnings. Nothing about the cash changes; everything about the reported profit does. Between 2020 and 2025, every major hyperscaler extended the assumed lives of its servers and network equipment. Harvard Business School turned Meta's version into a case study: servers and network assets moved from four years to five and a half, a revision that reduced 2025 depreciation by $2.9 billion.
Microsoft's July revision was the more ambitious one. Data centres and office buildings went from 15 years to 25 — longer, as one analysis noted, than the leases on some of the sites themselves. As the accounting analyst Olga Usvyatsky and Francine McKenna noted in their analysis of the 2026 annual report, the simultaneous shift of future leases from finance to operating classification had a second effect: it moved a chunk of spending out of the capex line and into operating cash flow, where it reduces free cash flow rather than appearing as investment.
The full-year pattern at Microsoft is stark. Capex rose from $44.48 billion in fiscal 2024 to $115.95 billion in fiscal 2026, while operating cash flow rose from $118.55 billion to $182.94 billion. Calculated free cash flow fell, from $74.07 billion to $66.99 billion, because capex grew at about 80% against 34% for operating cash. Capex went from 38% of operating cash flow to 63%. Depreciation and amortisation, at $39 billion, has grown far more slowly than the spending behind it — and longer asset lives will stretch that gap further. Free cash flow fell 23% in the fourth quarter alone, partly, as Hood noted, because of higher operating lease payments.
Across the five hyperscalers, the Wall Street Journal reported $146 billion of combined depreciation and amortisation for 2025. Michael Burry, who modelled the mortgage market correctly in 2005 and has been early on most things since, estimates that hyperscalers will understate depreciation by roughly $176 billion in aggregate between 2026 and 2028, overstating earnings by more than 20%. That is an advocate's number and should be treated as one. The direction, though, is not seriously contested.
The reason it matters is obsolescence. Jensen Huang, not a man given to underselling his own products, said of the H100 once Blackwell shipped: "You couldn't give Hoppers away." Nvidia now releases a new architecture annually where it once took two years. Epoch AI measured inference costs falling at a median of 50 times a year, accelerating to 200 times a year after January 2024; processing that cost about $20 per million tokens in 2020 could be had for roughly $0.07 by early 2026. If a chip's economic life is closer to three years than six, the depreciation maths does not merely get worse. It changes the character of the reported profit.
The industry's answer is Jevons: cheaper inference will expand demand faster than it destroys the value of the hardware. That may be right. It is also unfalsifiable in the short term, which is what makes it useful.
There is one more accounting detail worth holding onto. Microsoft's 2026 10-K disclosed, for the first time, that it recorded $24.1 billion of revenue from commercial arrangements with OpenAI in fiscal 2026, with $6 billion of accounts receivable outstanding. Microsoft holds roughly a 25% equity-method stake, so OpenAI's losses flow through "other, net" rather than being consolidated. Throughout fiscal 2026, Microsoft excluded gains and losses from that investment from its non-GAAP earnings. The disclosure is welcome. It also shows how much of the reported picture depends on where a boundary is drawn.
The filing cabinet
If depreciation moves the clock, structured finance moves the asset.
In October 2025, Meta and funds managed by Blue Owl Capital closed a $27 billion joint venture to develop and own the Hyperion campus in Richland Parish, Louisiana. Meta controls the facility in every operational sense that matters. The debt sits in the vehicle. The Wall Street Journal headline was unusually direct: "AI Meets Aggressive Accounting at Meta's Gigantic New Data Center." Paul Kedrosky put the objection plainly at the time: the special purpose vehicle is not technically on Meta's balance sheet, but for all practical purposes it exists to serve Meta and is controlled by Meta.
This is the oldest move in the book, and it has a distinguished history. Enron's special purpose entities, Lehman's Repo 105, Citigroup's conduits — the structure changes, the function does not. What is new is the scale and the ordinariness. Hyperion is not a fraud; it is a financing arrangement, disclosed, with a legitimate commercial logic. That is what makes it instructive. You do not need a scandal to move a trillion dollars of obligations into a place where a leverage ratio cannot see them.
Oracle has taken the technique furthest. Its off-balance-sheet lease commitments rose from $411 million in 2020 to roughly $45 billion by fiscal 2025 and about $260 billion by the third quarter of fiscal 2026, according to Calcbench. That is an increase of roughly 24,800% in five years. Its remaining performance obligations — contracted revenue not yet delivered — went from $99.1 billion to $455.3 billion to $638 billion at the fiscal year end, up 363% year on year, against annual revenue of $67.4 billion. A software company with a decade of revenue in backlog is a different kind of company from the one Oracle used to be, and it now carries four years of revenue in lease commitments before a single chip is switched on.
Meta's disclosures tell a similar story in slower motion. Future lease obligations reached $355.7 billion at 30 September 2026, up from about $279 billion in the second quarter. Secondary analyses of its filings, reported by Reuters, put total future commitments at roughly $825 billion, with about $121.4 billion contractually due in 2027.
And then there is the securitisation market, which has been drafted in to refinance the construction loans. Data centre asset-backed securities have grown from a standing start in 2018 to $11.9 billion of issuance in 2026 so far, against $15.5 billion in all of 2025. In August 2026, the Securities and Exchange Commission responded to a letter from the law firm Latham & Watkins by agreeing that securities issued in data centre securitisations "of the type described" are not asset-backed securities — exempting them from the disclosure and reporting regime that governs the asset class they most resemble.
That is not corruption. It is a regulator smoothing a financing channel during a boom, which is a thing regulators do. It is also how a market acquires depth without acquiring visibility.
The market, to its credit, has noticed. By late September, data centre ABS spreads had decoupled from the rest of the securitised market and were trading closer to corporate AI debt — 150 basis points over the five-year Treasury, 16 basis points wider than a year earlier, while the broader ABS market tightened. Hyperscaler corporate spreads moved in the same direction, to 165 basis points over SOFR, 31 basis points wider on the year. Barclays' Elana Lipchak described the shift to The Real Deal: data centre securities are now less correlated with ABS and more correlated with "all the debt supply on the AI infrastructure trade from corporates." The supply was large enough that when the Federal Reserve raised rates for the first time in three years, Chair Kevin Warsh cited it directly. "The so-called hyperscalers are out in the market raising funding," he said. "And so the competition for capital is real. And I think it partly explains the increase in yields."
Which brings us to the escape hatch, and the most revealing single document of the year. In September 2026, Oracle sent a force majeure notice to Blue Owl regarding Project Jupiter, a $165 billion, 2.45-gigawatt campus in New Mexico, seeking to delay payments if the site is not online by 2028. Oracle cited potential delays in securing power (Reuters, CNBC, Capacity).
Read that again. The contract that was supposed to make the revenue predictable contains a clause allowing the customer to walk if the weather, the grid or the schedule turns. Force majeure is a normal commercial provision. Invoking it before the failure, on the largest project in the portfolio, is something else: it is the exit ramp being built while the road is still being laid.
The circle
The financing has a property that is easy to describe and hard to see in any single transaction: the money tends to return to its starting point.
Nvidia is the clearest case. It is investing $30 billion in OpenAI, guaranteeing up to $105 billion for an OpenAI data centre in Ohio, putting up to $10 billion into Anthropic and $2 billion into CoreWeave, and made more than 50 AI investments in 2024 alone. It is simultaneously the supplier, the investor and, in some cases, the customer. One central bank analysis cited by The Conversation found that 55.2% of investment into AI firms came from other AI firms. The Reserve Bank of Australia devoted part of its Financial Stability Review to "opaque or circular funding arrangements" in which chipmakers finance the companies that buy their products.
The Australian data centre builder Firmus shows the structure in miniature. Nvidia invested in Firmus, supplies the platform its data centres are built on, and has signed on as a customer for a facility in Batam, Indonesia, worth up to $30 billion over six years. Firmus abandoned its A$44 billion listing in early October when investors would not support the valuation. All three relationships are disclosed. None of them is improper. Together they make it difficult to say where demand ends and financing begins.
Then there is the detail that best captures the moment. When Meta raised $25 billion of common stock in the third quarter, Nvidia, AMD and Broadcom indicated interest in buying up to $5 billion of the offering. The indications were non-binding, and the companies Meta pays for chips were offering to help fund the purchases. A supplier helping to finance its own customer is not new — Cisco did exactly that during the telecom buildout in the 1990s, and the fibre that resulted was real even as the investors who paid for it were wiped out. What is new is the speed and the circularity.
At the speculative edge sits the neocloud layer. CoreWeave ended the second quarter with $16.3 billion of operating lease liabilities and $35.6 billion of debt — about $52 billion of obligations — against $4.6 billion of first-half revenue and a $1.4 billion net loss. 98% of its revenue comes from committed take-or-pay contracts, and its top three customers generated 72% of it. Its bonds briefly yielded 15% in October after Chamath Palihapitiya posted a chart asking whether the AI buildout was creating a private credit bubble. Its remaining performance obligations stand at $103.7 billion, 41% of which falls due by 2028.
The defence, offered by investor Evan Schlossman to Fortune, is that those contracts are backstopped by Microsoft and OpenAI, and that the capacity would be worth more than the contract if a customer vanished. Francine McKenna's response is worth quoting in full: "If CoreWeave is counting on Microsoft to honor its contracts and bail them out of extreme debt and losses I have a bridge in Brooklyn I'd like to sell them. Microsoft is ruthless and will eat CoreWeave for lunch and throw bones to its dogs if it makes sense for them."
Gartner expects 80% of neocloud providers to cease independent operations by 2030 because their revenues will not cover their outgoings. That is the point of the structure. The hyperscalers' commitments are take-or-pay, which protects them until the counterparty cannot deliver — at which point the loss lands on the neocloud's lenders, its landlords and its equity, not on the company that signed the contract.
The concentration runs all the way up. Nvidia is about 8.06% of the S&P 500, and chipmakers accounted for roughly 70% of the index's gains in 2026. A single company's financing decisions now move the retirement savings of a country.
The sentences
You can learn a lot about an organisation by watching how it defines the test it might fail.
Sundar Pichai, July 2024: "The risk of underinvesting is dramatically greater than the risk of overinvesting for us here." Andy Jassy, April 2026: "We're not investing approximately $200 billion in capex in 2026 on a hunch." He has also described the moment as a land rush, and said Amazon will not be conservative.
Mark Zuckerberg, September 2025, is the most interesting because he concedes the ending in advance:
"If we end up misspending a couple of hundred billion dollars, I think that that is going to be very unfortunate, obviously. But what I'd say is I actually think the risk is higher on the other side."
He then reaches for the railroads and the fibre buildout: infrastructure that turned out to be enormously valuable, built by people who took on too much debt, followed by a blip, followed by a lot of companies going out of business. The analogy is doing a lot of work. It acknowledges that the builders may fail while implying that the building was right — and that the person who ordered it was therefore right too, even if the shareholders were not.
Satya Nadella, January 2026, moves the goalposts with more elegance than anyone:
"For this not to be a bubble, by definition it requires that the benefits of this are much more evenly spread. I think a telltale sign of if it's a bubble would be if all we're talking about are the tech firms."
This is a serious point about diffusion, and it is also a test that no capex figure can fail. You could spend $2 trillion and still pass, provided a hospital somewhere reports a faster drug trial.
Jeff Bezos calls it "a kind of industrial bubble, as opposed to financial bubbles" — a frame in which losses are the acceptable price of progress. Bill Gates distinguishes tulips from the internet, concluding that something profound will happen even though "a ton of these investments will be dead ends." Sam Altman, September 2026, blames the marginal entrant: "I'm not worried about our compute buildout plans. I am worried about the world's compute buildout plans," adding that he sees "unsustainable silliness of random, new neocloud popping up." Dario Amodei, more candidly, describes "a real underlying risk" and "a lot of pressure to push things," and says some players are "YOLO-ing."
None of these men is lying, exactly. Each is describing a genuine uncertainty. But notice the grammar. The question "will this investment earn its cost of capital?" has been replaced by "is AI important?", "will the technology diffuse?", "is the infrastructure useful?", and "are the smaller players reckless?" Every substitute question has a comfortable answer. The original one does not.
That is what a well-managed organisation does under threat: it changes the question.
The confession, in installments
Here is the part that should give the optimists pause and the pessimists a strange kind of hope. The admission is already happening. It is just being delivered to progressively larger audiences, in progressively softer language.
The first audience was investors, in private. In February 2026, OpenAI told them it was targeting roughly $600 billion of total compute spend by 2030 — down from the $1.4 trillion in infrastructure commitments that Sam Altman had been touting for months. The company also projected about $280 billion of revenue in 2030, against $13.1 billion in 2025, and a burn of $8 billion (CNBC). The new number was presented not as a retreat but as a clarification, a plan "more directly tied to expected revenue growth." That is what a confession looks like when it is made to people who have signed confidentiality agreements.
The second audience was the rating agencies. On 30 October 2026, S&P Global changed its outlook on Meta to negative, projecting capex of $140 billion in 2026, over $225 billion in 2027 and $252 billion in 2028, with free operating cash flow deficits of $22 billion, $71 billion and $39 billion respectively. It expects adjusted leverage of 1.3x by 2028, above the 1.0x threshold for the current rating, and notes that Meta's ability to monetise AI is more limited than its peers' because it has no cloud business to rent capacity to. Meta's stock fell. The agency was doing, in public and in a PDF, what no executive had done on a call: putting the numbers in a row and asking whether they add up.
The third audience is the bond market, which has begun to price the difference. AI-related issuers trade around 115 basis points over Treasuries against 78 for the broader investment grade market. Alphabet had to offer a large concession to complete its August sale; Aon's $13.5 billion acquisition financing attracted $65 billion of orders in the same month. BlackRock's Russell Brownback described some AI deals as double-A credits pricing closer to triple-B. Wellington's Loren Moran said some investors are nearing single-name exposure limits once debt issued through parent-backed data centre vehicles is aggregated back to the same technology companies — which is precisely the risk that off-balance-sheet structures create, arriving in the one place that cannot ignore it.
And the fourth audience, eventually, will be everyone else. That is the one that has not happened yet.
The pattern is not accidental. It follows the structure of the incentive. The people whose money is committed hear first; the people with the least immediate leverage hear last. By the time the public is told, the write-down has usually been taken, the chief financial officer has usually left, and the story has usually been rewritten as a strategic pivot.
Who pays
The costs do not disappear while the admission is deferred. They are redistributed, and the redistribution is already visible.
At the local level, data centres have become the most effective NIMBY movement in a generation, and it cuts across party lines. Local opposition blocked or delayed 75 projects worth $130 billion in the first quarter of 2026 alone — roughly the same number as in all of 2025, according to Data Center Watch, whose research is summarised by Brookings. Across two years the group counted $64 billion of projects blocked or delayed, 142 activist groups across 24 states, and a political split among opposing officials of roughly 55% Republican to 45% Democrat. In Cascade Locks, Oregon, voters recalled two port authority officials; in Warrenton, Virginia, they voted out an entire town council.
The industry's response has been to buy the politics. "Leading the Future," a super PAC launched by industry figures, began with $140 million, including a reported $50 million each from Andreessen Horowitz and OpenAI president Greg Brockman, for the purpose of supporting pro-AI candidates. Employee groups such as the Guardrails Alliance are collecting small-dollar donations in the other direction. Senators Josh Hawley and Elizabeth Warren, arriving from opposite ends of the spectrum, have both described the concentration of capital and political power as a problem.
The fiscal subsidy is being withdrawn more quietly. At least 38 states offer dedicated data centre tax incentives, but between May and August 2026 eight states paused or ended their breaks, and New Jersey repealed $250 million of credits earmarked for large AI facilities (CBPP, FiscalNote). At the federal level the opposite is happening: AI infrastructure is being folded into national security and sovereign capability, which is the most durable subsidy of all, because it makes opposition look like disloyalty. Britain announced a £1.1 billion package in June 2026, including £750 million for a national AI supercomputer.
And then there is the personal layer. One tally of Form 4 filings put insider sales at $77.6 billion in the first half of 2026, the most since 2021, the year before the last technology correction. That is not proof of anything — executives sell for many reasons, and 2021 was a good year to own shares — but it is an odd coincidence that the people most confident about the decade are reducing their exposure to it.
Meta's capital structure is the tell here. In the third quarter it raised $37.5 billion — $25 billion of common stock at $625 a share and $12.5 billion of mandatory convertible preferred yielding 6.5% — and announced a further $60 billion at-the-market programme. It also plans a one-for-one Class C dividend in January 2027, creating a class of non-voting shares it can issue indefinitely without diluting founder control. Economically, the dilution falls on shareholders. Structurally, control does not move.
That is a company that has decided the spending must continue and that the equity market, not the operating cash flow, will fund it.
The case for the defence
A piece like this is dishonest if it pretends the evidence runs only one way. It does not.
AI revenue is growing quickly. Azure grew 43% and is guided to about 45%; Microsoft's commercial remaining performance obligations reached $678 billion; Oracle's backlog grew 363%. The $25 billion quarter that edged past depreciation is a real signal, and it has now happened twice. If the ratio keeps improving at that pace, the arithmetic that looks impossible in 2026 looks merely difficult in 2029.
Some of the capex is genuinely demand-backed. When a hyperscaler signs a take-or-pay contract with a creditworthy customer, it is not building on a hunch, whatever the rhetoric. The bond market is discriminating rather than panicking: spreads have widened, but issuers are still funding, and investors are still buying. That is what a functioning feedback loop looks like. The rating agencies are engaged. Short sellers are being answered in public. Fidelity noted in early 2026 that several classic bubble indicators — shrinking free cash flow, deteriorating credit — were not yet flashing in the way they did in 1999.
Nor is every bearish claim true. SemiAnalysis dismantled the viral statistic that half of 2026 US data centre capacity would be delayed or cancelled, showing that the underlying reports were measuring different things and that the headline number was not supported. Amazon shortened the assumed life of some servers from six years to five and took a $920 million charge, cutting 2025 operating income by a further $700 million. The accounting is not uniformly one-directional. The BIS, hardly a cheerleader, concluded that AI investment had supported global growth through a difficult period.
There is also a serious argument that the infrastructure is valuable even if the investors are not. Railroads were overbuilt. Fibre was overbuilt. Both excesses funded capacity that later ran full. Zuckerberg's analogy is not stupid; it is the strongest card in the deck. The problem is that it is a card about the asset, played in a game about the return.
So the honest position is not that the bet will fail. It is that the bet is enormous, the hurdle is high, the measurement has been flattered, and the people who made it have arranged their affairs so that the failure, if it comes, will be attributed to circumstances rather than to judgement.
What would make them say it aloud
Voluntary candour is rare because the incentives point the other way. Reversals usually require one of four things.
Liquidity or creditor stress. When the story stops being optional, it stops being told. This is the mechanism that ended the telecom buildout: not a change of heart, but a change in the cost of rolling over commercial paper.
An outside actor with no ownership of the decision. An activist, a new chief executive hired from outside, a board that turns over. The confession is much easier for someone who did not make the bet.
An external shock so unambiguous that it is not personally blameworthy. "The market turned" is a survivable sentence. "I was wrong" is not. A recession, a rate shock or a demand air pocket allows everyone to be wrong together, which is the safest way to be wrong.
A competitor breaking ranks first. If one hyperscaler guides capex down and the stock goes up, the taboo breaks for everyone. This is why the market watches the first mover so closely.
And when the admission does come, it will not use the word. It will be a "rebalancing toward the highest-return opportunities." It will be an "impairment following a disciplined portfolio review." It will be a "strategic pivot to inference." A chief financial officer will leave to spend more time with family. The chief executive will stay, having "reset the strategy." The write-down will be taken in a quarter when something else — a legal charge, a tax change, a currency move — gives the numbers cover. Meta's third quarter, with its $9.89 billion legal charge, showed how conveniently such cover arrives.
This is not cynicism about individuals. It is a description of how institutions process failure. The organisation needs a story in which the decision was reasonable at the time, and the story can be true: it probably was reasonable at the time. That is what makes escalation so durable. Each step was defensible. Only the aggregate is damning, and the aggregate is nobody's job.
Six things to watch
If the playbook is running, it should leave fingerprints. These are the ones I would watch over the next four quarters.
- A second force majeure. Oracle's notice to Blue Owl is currently an exception. A second one, from a different company or on a different project, would convert it into a norm — and would tell you that customers are pre-positioning to walk away from take-or-pay obligations.
- Capex guided down in absolute dollars while demand is described as strong. Microsoft has already shown the template: a lower number achieved by reclassification, with management insisting nothing changed. Watch for the version where the spending actually falls and the language does not.
- An auditor or SEC comment letter on useful lives. The depreciation assumptions are where the earnings are most exposed. A challenge there would force disclosure that the current estimates are a judgement rather than a fact.
- A neocloud bankruptcy that tests take-or-pay. Chapter 11 is where the contracts get read by a judge rather than a credit committee. If a CoreWeave-style counterparty fails and the hyperscaler obligations do not survive intact, the whole financing structure reprices at once.
- A chief financial officer exit. The CFO is the traditional designated adult in these situations: the person who signed the numbers and can be moved aside without the chief executive conceding anything.
- Data centre ABS spreads decoupling further. Barclays flagged the first divergence in September. If securitised data centre debt starts trading like distressed corporate credit while the ABS market tightens, the refinancing channel that the whole structure depends on will have begun to close.
There is a seventh, harder to date but more revealing. Watch what the companies say when they are not being asked about AI. Watch whether the language in the annual letter, the proxy statement and the investor day starts to include the word "discipline." Watch for the first use of "legacy capacity." Watch for the moment a data centre is described as "repurposed" rather than "expanded."
The sentence
The strange thing about this cycle is how much of it is on the record. The $638 billion backlog. The $355.7 billion of lease obligations. The 25-year buildings. The $5 billion of equity that chipmakers offered to buy from their own customer. The force majeure notice. The negative outlook. The $1.4 trillion that became $600 billion in a private meeting.
None of it required an investigative journalist to uncover. It required someone to read it all in one sitting, which is precisely what no one inside these companies is paid to do.
That is the mechanism, in the end. Not secrecy, but distribution. Each disclosure is small enough to be explained. Each explanation is reasonable. The explanations accumulate into a story in which the bet is always about to pay off, the demand is always just ahead of capacity, and the write-down is always someone else's quarter.
Staw's students escalated because they could not bear to have been wrong. The executives in this story are in a harder position, because they cannot afford to be wrong in public and they cannot stop without admitting that stopping was always an option. So they do what organisations do under that pressure: they keep the commitment, adjust the measurement, spread the risk, and change the question.
The confession is already written. It is being delivered in instalments, to audiences ordered by how much they can cost the company, and the last instalment is the only one that will be called a confession at all.
By then, the sentence will have been edited into something that can be filed under "rebalancing" — and the people who wrote it will be able to say, with perfect sincerity, that they never claimed it was certain.
Figures are as reported in company filings, earnings transcripts, rating agency reports and the journalism cited above; estimates from banks, consultancies and short sellers are attributed to their authors and should be read as estimates. Where a claim rests on a single report, the report is named. No allegation of fraud is made or implied: the structures described are disclosed, and their significance is a matter of interpretation.