Michael Burry does not make small bets. The Scion Asset Management founder, who turned a contrarian mortgage short into one of the most profitable trades of the 2008 financial crisis, has now staked roughly $1.1 billion against two of the AI boom’s highest-flying stocks. His argument goes beyond valuation: Burry contends that the largest technology companies in the world are using an accounting lever to quietly inflate their earnings while spending record sums on AI infrastructure.
The wager surfaced in Scion’s quarterly 13F filing with the Securities and Exchange Commission, which disclosed put-option positions on Nvidia and Palantir Technologies as of September 30, 2025. The filing was made public on November 3 of that year. Put options increase in value when a stock declines, making the positions a direct bet that both companies are overvalued.
But the options are the mechanism, not the thesis. The thesis is about depreciation, and it implicates a much wider swath of Big Tech than just two stocks.
How a bookkeeping change worth billions actually works
Every company that buys physical equipment, whether it is a delivery truck or a rack of AI servers, spreads the purchase cost over the asset’s expected useful life. That annual charge, called depreciation, reduces reported operating income. The longer a company assumes the equipment will last, the smaller the annual hit to profits.
This is not exotic accounting. It is a basic requirement under U.S. Generally Accepted Accounting Principles (GAAP). But when a company decides to extend an asset’s useful life, the effect on earnings can be enormous, especially at the scale of Big Tech capital spending.
Alphabet provided the clearest example. In its 2023 annual report (10-K), the company disclosed that it had extended the estimated useful life of its servers from four years to six years and adjusted certain networking equipment from five years to six. That single change reduced Alphabet’s depreciation expense by $3.9 billion for fiscal 2023 and added $3.0 billion to net income, according to the filing’s notes.
Oracle made a similar move. In its 10-K for the fiscal year ending May 31, 2025, the company reported extending the useful lives of servers and networking equipment from five years to six years. Microsoft disclosed comparable adjustments to the depreciable lives of cloud and AI-related hardware in its fiscal 2024 annual report, as did Meta Platforms. Each revision followed the same logic: newer hardware lasts longer, so the write-off period should be longer too.
None of these changes are concealed. They appear in the footnotes of audited financial statements, reviewed and signed off on by Big Four accounting firms. But footnotes are where information goes to be technically disclosed and practically ignored. Most investors never read past the earnings-per-share line.
The $176 billion claim: what we can verify and what we cannot
Burry has pointed to $176 billion as the cumulative scale of depreciation being obscured across the technology sector. That figure has circulated in financial commentary tied to his public statements, but it does not appear in any single SEC filing or audited financial statement.
The number appears to originate from third-party analysis that extrapolates the company-level depreciation changes at Alphabet, Oracle, Meta, Microsoft, Amazon, and others into a sector-wide total. The full methodology and the complete list of companies included have not been published in a form that can be independently verified from primary filings alone.
What is verifiable from those filings is the pattern: at least four trillion-dollar technology companies extended depreciation schedules on AI-critical hardware between 2023 and 2025, a period when their combined capital expenditures on data centers and AI infrastructure reached historic highs. Each extension mechanically boosted reported profits. Whether the aggregate effect approaches $176 billion depends on assumptions about which companies are included and over what time horizon, but the direction of the distortion is not in question.
What Scion’s 13F filing reveals, and what it leaves out
A 13F is a regulatory snapshot, not a trading diary. It captures a fund’s long equity and options positions on a single date and is typically filed 45 days later. Scion’s disclosure tells us that as of September 30, 2025, the fund held put options on shares of Nvidia (NVDA) and Palantir (PLTR) with a combined reported value near $1.1 billion.
Crucially, the filing does not disclose strike prices, expiration dates, the premiums Burry paid, or whether the positions have been modified or closed since. The $1.1 billion figure represents the market value of the options as reported on the filing date, not the notional value of the underlying shares and not the capital Scion put at risk. The actual amount spent to acquire the contracts could be substantially less.
Scion’s public ADV record with the SEC describes a small advisory firm that has historically favored concentrated long positions in sectors like energy, healthcare, and traditional financials. A derivatives-heavy, macro-style bet of this magnitude is a departure from that profile, which suggests Burry sees this as a high-conviction opportunity.
There is also a nuance worth noting: Nvidia and Palantir are not the companies changing their depreciation schedules. Nvidia sells the GPUs; it is the cloud operators and enterprise buyers of those chips who are extending useful lives. Burry’s bet on Nvidia and Palantir appears to target the downstream consequence: if the AI infrastructure buildout is being subsidized by accounting flattery, the companies supplying and servicing that buildout are the most exposed when reality catches up.
The case that the depreciation changes are legitimate
Burry’s skeptics have a straightforward rebuttal: the hardware actually does last longer now.
Server manufacturers have improved chip durability, thermal management, and firmware update cycles over the past decade. Workload orchestration software allows companies to redistribute processing tasks across aging machines rather than retiring them on a fixed schedule. Google, which pioneered the design of custom server hardware for its data centers, has published research showing that its equipment remains productive well beyond the four-year window it previously assumed.
GAAP does not just permit companies to update useful-life estimates when evidence supports a change. It requires them to do so. From this perspective, the revisions are a sign of responsible accounting, not manipulation. The Big Four auditors who reviewed and approved the changes presumably reached the same conclusion.
Then there is the revenue side of the ledger. Nvidia’s data-center segment posted consecutive quarters of record revenue through late 2025, driven by insatiable demand for GPUs used in training and running large language models. If that demand persists and the infrastructure generates cash flows that justify the spending, the depreciation timeline becomes a secondary concern.
Burry’s post-2008 public record also gives bulls reason for skepticism. He warned of a broad stock market crash multiple times between 2020 and 2023, took a short position against Tesla, and raised alarms about passive-investing bubbles. Some of those calls were early. Others simply did not play out. Options, unlike long stock positions, carry expiration dates, and being directionally right on a thesis but wrong on timing can be financially indistinguishable from being wrong entirely.
The case that the bill is coming due
Bears counter that the timing of the depreciation changes is too convenient to dismiss. AI-related capital expenditures at the five largest U.S. tech companies exceeded $200 billion in 2024 alone, according to their combined earnings reports. Under the old depreciation schedules, that spending would have generated significantly larger annual charges, pressuring the profit margins that investors use to justify premium stock valuations.
Extending useful lives allowed companies to report stronger earnings per share at precisely the moment Wall Street was rewarding AI exposure with historically rich multiples. Whether intentional or not, the effect was to smooth the cost of the AI buildout across more years, making the investment look less painful on a quarterly basis.
The deeper risk is obsolescence. AI hardware evolves rapidly. The GPU architectures dominating data centers in 2025 may be supplanted by newer chips, custom silicon, or entirely different computing paradigms within a few years. If current-generation servers become economically obsolete before their six-year depreciation window closes, companies will face impairment charges or accelerated write-downs that compress earnings in a single quarter rather than spreading the pain over time.
That scenario is exactly what Burry appears to be positioning for: a moment when deferred costs resurface, reported earnings contract, and the stocks that benefited most from the AI premium reprice sharply lower.
What investors should actually watch next
As of June 2026, the outcome of Burry’s trade is still unresolved. The September 2025 filing remains the most recent public window into Scion’s positioning. The fund’s next 13F, covering holdings as of March 31, 2026, should reveal whether Burry held, expanded, or unwound the bet. Nvidia and Palantir shares have swung in both directions since the filing became public, driven by shifting expectations around AI monetization timelines, U.S. chip export controls, and broader equity market volatility.
Regardless of how the trade resolves, Burry has forced a question into public view that most earnings coverage ignores: when a company reports record profits while simultaneously spending record amounts on infrastructure, how much of that profit is real and how much is a function of how the spending is spread across the calendar?
For anyone evaluating AI stocks on fundamentals, the answer is buried in the same place it has always been: the footnotes of quarterly and annual SEC filings. The gap between reported earnings and the economic cost of the hardware generating them may be the most consequential number that never appears in an earnings headline. Reading past the first page has rarely mattered more.
