On September 28, 2026, the investor who became famous for shorting the 2008 housing market published a newsletter post arguing that the biggest names in tech are quietly inflating their profits by stretching out how long they claim their Nvidia chips will last. Two days later, a routine regulatory filing showed Michael Burry’s firm, Scion Asset Management, had backed the argument with real money: put options with a notional value of roughly $187 million against Nvidia and $912 million against data-analytics firm Palantir, disclosed as of September 30, 2026. It is the most concrete bet yet that one of Wall Street’s most-watched contrarians thinks the AI boom’s accounting, not just its valuations, is due for a reckoning.
A Short Seller Returns to the Spotlight
Burry built his reputation finding a structural flaw markets had priced as irrelevant — mortgage-backed securities before 2008 — and the pitch of his latest warning follows the same shape. Rather than arguing AI technology doesn’t work, he is arguing that the companies spending hundreds of billions of dollars building it are using depreciation schedules that make today’s earnings look healthier than the underlying economics justify. He has been public about this for months, but the September 28 post sharpened the argument with a specific number and a specific villain: Nvidia’s GPUs, and the accounting choices hyperscalers make about how long those chips remain useful.
The Depreciation Trick, Explained
Depreciation lets a company spread the cost of an expensive asset, like a server full of GPUs, across several years of its income statement instead of booking it all at once. The longer the assumed useful life, the smaller the annual expense, and the higher reported profit looks in any given year. Burry’s newsletter cites research estimating that major technology companies could collectively understate depreciation by about $176 billion between 2026 and 2028 by assuming their GPUs last far longer than they actually do. His central claim is that Nvidia ships a new chip generation roughly every 18 to 24 months, and each new generation roughly halves the resale value of the one before it — meaning a GPU’s real economic life is closer to two or three years, not the five or six years some companies now assume.
Who’s Stretching the Clock
According to the figures circulating in Burry’s analysis, Microsoft has extended its depreciation period for GPU assets from four years to six, while Meta has adopted a 5.5-year schedule; Google, Amazon, and Oracle are described as making similar adjustments. Applied forward, Burry estimates Oracle could be overstating earnings by as much as 26.9 percent by 2028, with Meta overstating by roughly 20.8 percent over the same stretch. Those are large enough gaps that, if accurate, they would materially change how investors should read the profit margins hyperscalers have been reporting as justification for continued AI infrastructure spending.
The Bet Behind the Warning
Burry isn’t simply publishing opinions — his firm’s September 30 disclosure shows he is positioned against the trade working out well for Nvidia and Palantir specifically, the two companies whose valuations are most directly tied to the market’s belief that AI infrastructure spending will keep paying off. The timing matters: his warning lands alongside broader unease about so-called circular deals, arrangements in which Nvidia invests in OpenAI, OpenAI buys cloud capacity from Oracle, and Oracle in turn buys chips from Nvidia, with AMD granting OpenAI warrants for roughly 10 percent of its equity as part of a separate chip-supply arrangement. Critics argue these overlapping, self-reinforcing deals make it harder to tell how much underlying demand for AI compute actually exists versus how much is manufactured by the deals themselves.
Why Tech Giants Say the Critics Are Wrong
The hyperscalers’ defense, as reported by industry analysts, is that Burry’s framework oversimplifies how enterprise hardware actually gets used: GPUs retired from cutting-edge model training don’t go to the scrap heap, they get repurposed for less demanding inference workloads, extending their productive life well past the point they’d be useless for frontier research. Finance teams at these companies have also noted that accounting standards allow for judgment in setting useful-life assumptions and that external auditors have signed off on the schedules in question. Former Intel CEO Pat Gelsinger has staked out a middle position, telling interviewers he agrees the AI sector is in bubble territory but expects any correction to unfold gradually rather than in a single dramatic break.
What Happens Next
Burry’s warning arrives against a backdrop of growing investor jitteriness: a late-June 2026 selloff knocked the Nasdaq down 2.2 percent and the S&P 500 down 1.4 percent in a single session, with AI-memory supplier Micron Technology — up nearly 800 percent over the prior year — falling 13 percent. It also follows a widely cited MIT study finding that roughly 95 percent of corporate generative AI pilots have failed to produce a measurable profit impact, reinforcing the sense that there’s a gap between AI spending and AI payoff. None of this proves Burry is right about depreciation specifically, and hyperscalers will report their next round of quarterly earnings with their existing schedules intact. But his bet guarantees the question won’t go away quietly: either the depreciation assumptions hold up under scrutiny from auditors and regulators, or a chunk of the profits Wall Street has been crediting to the AI boom turns out to have been an accounting artifact all along.
Photo: Gijs Peijs / BY via flickr
