Burry Says Five Hyperscalers Carry $3 Trillion in Hidden AI Commitments — Nvidia Already Answered, and He Won't Budge

Michael Burry has a new number and a new deadline for the AI trade. The investor behind "The Big Short" laid out his case in a Substack post published Thursday, arguing that Amazon, Meta, Alphabet, Microsoft and Oracle are carrying roughly $3 trillion in off-balance-sheet commitments tied to AI infrastructure — and warning the bill could arrive in 2028 or 2029, when even a small write-off could land with outsized force.

Michael Burry has a new number and a new deadline for the AI trade. In a Substack post published on September 24, Burry compared the current AI spending boom to past capital cycle bubbles, including the dot-com era. The investor behind "The Big Short" argues that Amazon, Meta, Alphabet, Microsoft and Oracle are carrying roughly $3 trillion in off-balance-sheet commitments tied to AI infrastructure — and warns the bill could arrive in 2028 or 2029, when even a small write-off could land with outsized force.

A $3 trillion figure from the filing footnotes

The post, titled "Capital Cycle IQ & the Forensic Files of the Big 5 Hyperscalers," follows a September 19 installment subtitled: "A study of overlooked footnotes to the 10-K and 10-Q filings from Amazon, Meta, Alphabet, Microsoft & Oracle." Burry's number is not on-balance-sheet debt. It is an aggregate of the obligations that live in filing footnotes: purchase commitments, future leases, guarantees backing third-party debt, construction in progress, and various special-purpose vehicles. In his own words: "That $3 trillion in aggregate purchase commitments, future leases and guarantees backing third-party debt, construction-in-progress and various SPVs - these all point to epic levels of spending over the next few years."

The breakdown, as reported in coverage of the posts, names Alphabet as the largest single carrier at nearly $900 billion in off-balance-sheet commitments and exposures, with Meta at roughly $700 billion in uncommenced leases and purchase obligations — a figure Burry suggests could approach $1 trillion as broader future commitments are counted. These are Burry's estimates from his reading of the filings, not the companies' own disclosures. Under current accounting rules, commitments of this kind can stay off the balance sheet until goods arrive and obligations crystallize — which is why he titled the earlier post "The Missing $3 Trillion."

The capital-cycle case: 2.07% of GDP

The post's centerpiece is a chart of S&P 500 net capital investment — capital expenditures minus depreciation — as a share of U.S. GDP. As of June 30, the last available data, Burry puts the reading at 2.07%, higher than at any point in nearly four decades except the aftermath of the March 2000 Nasdaq peak. "I have little doubt the next few quarters will set still higher and higher net investment/GDP marks, possibly even eclipsing that aftermath of the 2000 tech stock peak," he writes.

The history lesson is the point. After the late-1990s telecom-media-tech buildout, depreciation overwhelmed new capital spending so thoroughly that aggregate S&P 500 net investment went negative for twelve straight quarters, from mid-2003 to mid-2006 — a hangover Burry argues helped fell Merrill Lynch, Lehman Brothers and Bear Stearns before the housing crisis. "The most profitable companies in the world, except Apple, are betting everything on this as their free cash flow turns negative and the borrowing pace increases," he writes. The current pace: hyperscalers are on track to spend roughly $800 billion on capital expenditures this year, with combined annual spending expected to cross $1 trillion in the year ahead, according to TheStreet's reporting.

The depreciation fight: $176 billion and four to six years

Running alongside the spending argument is Burry's accounting argument. He has long contended that hyperscalers depreciate AI chips over four to six years while the realistic replacement cycle is two to three — estimating the gap understates industry-wide depreciation by roughly $176 billion between 2026 and 2028. On his math, Oracle's earnings would be overstated by about 26.9% and Meta's by about 20.8% by 2028 if current depreciation assumptions hold. "Will I continue on this idea that depreciation expense is being understated and write-downs are the future? Yes," he writes. Burry's figures are estimates that assume AI hardware becomes uneconomic to use sooner than the companies expect.

Nvidia has already answered this argument — directly and by name. In November 2025, the company circulated a seven-page memo to sell-side analysts, first reported by Barron's and detailed by CNBC, pushing back on Burry's claims about stock-based compensation, depreciation and what he calls circular financing. The Nvidia seven-page response to "questions and claims we've received," began by citing "Michael Burry on Twitter / X" as its first collection of source documents the company sought to refute.

The memo said its customers depreciate GPUs over four to six years based on real-world longevity and utilization patterns, and that 2020-era A100s continue to run at high utilization and retain meaningful economic value well beyond the two-to-three years critics claim. It corrected one Burry figure outright: Nvidia has repurchased $91 billion in shares since 2018, not the $112.5 billion he cited — Burry, the memo said, had incorrectly included taxes on restricted stock units. And it rejected the circular-financing suggestion, saying Nvidia's strategic investments are a small fraction of revenue and AI startups raise capital predominantly from outside investors.

Burry's reply, on his Substack, did not retreat: "Nvidia emailed a memo to Wall Street sell side analysts to push back on my arguments on [stock-based compensation] and Depreciation ... I stand by my analysis. I am not claiming Nvidia is Enron. It is clearly Cisco."

"Its name is Nvidia"

The Cisco line is the spine of the whole thesis. Burry believes Nvidia now occupies the position Cisco held in 1999–2000: the picks-and-shovels supplier at the center of a giant capital investment cycle. Telecommunications companies then spent tens of billions laying fiber on forecasts that internet traffic would double every 100 days; Burry notes that less than 5% of U.S. fiber capacity was ever operational. "And once again there is a Cisco at the center of it all, with the picks and shovels for all and the expansive vision to go with it. Its name is Nvidia," he wrote.

It is worth keeping the analogy honest: Cisco's forward price-to-earnings multiple was around 200 at its 2000 peak, while Nvidia's is a fraction of that, and Cisco survived as a profitable company even as its shareholders who bought the top waited a generation to break even. Burry's argument is not that Nvidia is a fraud — it is that the supplier of a capital cycle is where the cycle's arithmetic eventually shows up.

The bull-side answer, from the supply side

The strongest rebuttal to Burry's newest post has come not from Nvidia but from the order books. Micron's chief business officer told investors on a recent earnings call that customer demand for memory chips remains well above the company's ability to supply across nearly every product category through 2028 — a direct counter to the claim that AI demand is inflated. Oracle's September 10 results gave the bulls ammunition too: cloud infrastructure revenue jumped 121% to $7.39 billion, total revenue rose 30% to $19.35 billion, and the company lifted full-year guidance to at least $90 billion while booking another $30 billion in AI cloud contracts.

The next evidence arrives soon. Micron (MU) reports on September 30, and the hyperscalers report in late October, so we'll hear how they're talking about useful life and capacity soon enough. That earnings commentary — on depreciation schedules, on uncommenced leases, on whether demand is real or financed — is where Burry's thesis meets something checkable.

He has money on it

Burry has positioned his own capital behind the argument. His disclosed September portfolio shows short positions led by Oracle, Palantir, Nebius, Nvidia and the iShares Semiconductor ETF, with additional shorts on Micron, Caterpillar and CoreWeave, plus leveraged downside via put options — including January 2027 puts on the Philadelphia Semiconductor Index. Most of those short targets posted gains through August, which is the clearest test of his timing while the AI trade keeps rallying. In mid-September he trimmed positions across the portfolio while keeping the same ranking, describing the move as pulling back on risk.

He has also added a political twist. "Trump and his team know that the AI narrative and the buildout is the only thing keeping this economy going, and the most positive thing happening this year," Burry wrote in a recent Substack chat. "They cannot afford to let it fall," he added, while questioning what Washington could actually do to prevent a downturn.

No write-offs have happened. Burry's 2028–2029 window is a forecast, not a fact, and the honest read of this fight is that both sides have real points: the commitments are enormous and largely off the balance sheet, while the demand behind them is currently verifiable in order books and quarterly results. The testable part of the thesis — useful lives, lease commencements, whether the $3 trillion keeps growing in the next footnote cycle — will be adjudicated in earnings calls long before 2028.