Burry Tallies $664 Billion of AI Promises at Oracle Alone
The Big Short investor totted up AI obligations across five hyperscalers, put $664 billion against Oracle's name, and warned a small write-off could do outsized damage to earnings.
Michael Burry has added up the artificial-intelligence commitments at five hyperscalers, citing $664 billion at Oracle, and argues that even a modest write-off of those assets would hit reported profits harder than investors expect.
Michael Burry, the investor who built his reputation shorting mortgage credit before 2008, has turned his arithmetic on the artificial-intelligence buildout. He has tallied the AI commitments accumulating at five hyperscale computing companies and singled out a figure of $664 billion attached to Oracle (ORCL). His argument, as reported by TheStreet, is not that the spending is fake. It is that even a small write-off against obligations of that scale would land harder on reported earnings than the market has priced.
Burry has also put a date on when he expects the reckoning to arrive — a rhetorical move that separates this warning from the general unease about AI capital spending that has circulated for two years. A thesis with a calendar attached is falsifiable, and that is precisely why it travels.
Why an obligation number is not the same as a capex number
The distinction matters more than the headline figure. Capital expenditure is money already spent, sitting on the balance sheet as an asset and running through the income statement as depreciation over an assumed useful life. A commitment is a promise to spend — contracted purchases of chips, leases on data-center capacity, take-or-pay arrangements with power suppliers and colocation landlords. Those promises disclose in the footnotes, not in the headline capex line, and they are far less visible to the average screen or stock-market app.
That is the gap Burry is pointing at. A company can grow its contracted obligations far faster than its revenue and still look disciplined on the metrics investors habitually watch. The obligations only become an accounting event later — when the equipment arrives, when the lease commences, when depreciation starts. If demand for the capacity is not there by the time the bill matures, the asset gets written down and the charge hits in a single quarter rather than being smoothed across several years.
The lever is the depreciation assumption
Every hyperscaler makes a judgment call about how long an AI accelerator earns its keep. Stretch the assumed life and annual depreciation falls, earnings per share rise, and the same physical box flatters more quarters. Shorten it and the reverse happens. Nothing about the hardware changes; only the accounting estimate does.
That is the mechanism Burry is warning about, and it cuts both ways. The optimistic case is that a GPU bought this year is still generating inference revenue years from now, and the long life assumption is simply honest. The pessimistic case is that each new chip generation makes the prior one economically marginal faster than the depreciation schedule admits, and the shortfall accumulates quietly until someone has to recognize it.
The honest position for an investor is that neither case has been settled by evidence yet. What can be said is arithmetic: the larger the committed base, the smaller the percentage error that produces a material earnings impact. Against a number the size of the one Burry cites for Oracle, a write-off that would be trivial for a smaller balance sheet becomes a headline event.
Oracle's last close, and what the tape is actually saying
Oracle closed at 137.10, down 1.75% from its prior close of 139.54, as of the last trade on Friday, September 25, 2026 at 20:00 GMT. The session range ran from 136.92 to 140.88, meaning the stock gave up ground through the day and finished nearer the low than the high.
88, meaning the stock gave up ground through the day and finished nearer the low than the high.
That is a single session and should not be over-read. It also stands out against a broadly firmer market. The S&P 500, via SPY, closed at $771.35, up 0.54% on the day from $767.18. The Nasdaq 100 tracker QQQ finished at $744.50, up 0.46% from $741.10. The Dow 30 proxy DIA was the strongest of the three, closing at $517.49, up 0.94% from $512.68. So the tech-heavy benchmark rose modestly while Oracle fell — a divergence worth noting, but not on its own evidence that the market has adopted Burry's view.
Indeed, the wider read is closer to the opposite. Both the S&P 500 and the Nasdaq 100 proxies closed higher in the same session in which one of the most widely followed skeptics in the market attached a date to an AI unwind. Whatever the merits of the argument, the index level does not yet reflect it.
What would confirm the thesis, and what would break it
Readers who want to test the claim rather than simply react to it have a short list of things to watch, none of which require a forecast:
- Footnote disclosure of purchase obligations. The commitments tables in annual and quarterly filings are where promised spending lives before it becomes an asset. Their direction of travel, quarter over quarter, is the cleanest available signal.
- Changes to useful-life estimates. Any extension or shortening of assumed server and accelerator lives is disclosed, and any company doing it will explain why. A round of shortenings across the group would be the single most concrete confirmation of Burry's mechanism.
- Impairment language. The first write-down of AI infrastructure at a hyperscaler — whatever its size — would establish that the category can be impaired at all, which is currently an assumption rather than an observed fact.
- Revenue attach rate. Committed capacity is only a problem if it goes unsold. Disclosed AI or cloud backlog growing alongside the obligations is the counter-evidence.
The cost of being early
Burry's record cuts in two directions at once. He was right about subprime mortgage credit and famously endured years of losses, redemption pressure and ridicule before the position paid. That history is why his warnings get amplified. It is also the reason to treat a dated prediction carefully: the same investor's most celebrated call was correct on substance and badly wrong on timing for an extended stretch, which is ruinous for anyone using borrowed money or short-dated options to express the same view.
The more durable point in his tally has nothing to do with the date. It is that AI spending has grown to a scale where the accounting estimates governing it — useful life, residual value, impairment triggers — have become material inputs to reported profit rather than technical footnote matters. Investors who track only revenue growth and capex headlines are looking at the wrong line. The obligations disclosed beneath them are where the risk Burry is describing actually sits, and they are public. Anyone can go and add them up, which is essentially what he did.
Key facts
- Oracle last close: ORCL 137.10, -1.75%, as of Sept. 25, 2026 20:00 GMT
- Figure cited for Oracle: $664 billion in AI commitments
- Companies tallied: Five hyperscalers
- Benchmark close: S&P 500 (SPY) $771.35, +0.54%; Nasdaq 100 (QQQ) $744.50, +0.46%
Frequently asked questions
What exactly did Michael Burry claim?
Burry added up the artificial-intelligence commitments piling up at five hyperscale computing companies, citing a figure of $664 billion for Oracle. He argued that even a small write-off against obligations of that size would hit reported earnings harder than investors currently expect, and he attached a date to when he expects that reckoning.
Is $664 billion money Oracle has already spent?
No. The figure refers to AI commitments rather than completed capital expenditure. Commitments are contracted future obligations — chip purchases, data-center leases, power and colocation agreements — that typically appear in filing footnotes. They become income-statement items later, as assets are delivered and depreciation begins.
How did Oracle shares finish the most recent session?
Oracle closed at 137.10 as of the last trade on Friday, September 25, 2026 at 20:00 GMT, down 1.75% from a prior close of 139.54. The stock traded between 136.92 and 140.88 during the session and finished nearer the low of that range than the high.
Did the broader market fall on the warning?
No. In the same session, the S&P 500 proxy SPY closed at $771.35, up 0.54%, the Nasdaq 100 tracker QQQ closed at $744.50, up 0.46%, and the Dow 30 proxy DIA closed at $517.49, up 0.94%. The indexes rose while Oracle declined.
Why does depreciation matter so much to this argument?
Depreciation spreads the cost of equipment across its assumed useful life. A longer assumed life lowers annual depreciation and raises reported earnings per share; a shorter one does the reverse. Because the assumption is an estimate rather than a fact, changes to it can move profit without anything changing physically.
What should investors watch to test the claim?
Four disclosed items: the purchase-obligation tables in quarterly and annual filings, any change to assumed useful lives for servers and AI accelerators, the first impairment of AI infrastructure at a hyperscaler, and whether disclosed cloud or AI backlog is growing alongside the committed spending.
