Is Michael Burry Right About AI Depreciation?

On 10 November 2025, Michael Burry posted that the biggest AI spenders "will understate depreciation by $176 billion 2026-2028." Eleven months later, one of the five companies he named has already reversed course. In its 2025 annual report, Amazon shortened the life of part of its server fleet from six years to five, citing "the increased pace of technology development, particularly in the area of artificial intelligence."

That reversal is the strongest evidence for Burry's argument. It isn't proof of his number. We read the filings behind his claim to separate the two.

The short version

• The direction is right. Microsoft, Alphabet, Amazon, Meta and Oracle all lengthened the accounting life of their servers between 2020 and 2025. Those changes lifted reported profit by billions a year, all disclosed in their filings.
• Amazon blinked. It is the only one to shorten lives again, and it named AI as the reason.
• His $176 billion has no published method. Our rough check shows the figure is plausible only if AI servers wear out economically in about 3–4 years rather than the 5–6 the companies assume.
• Today's rental market cuts against him. Nvidia's H100, launched in 2022, rents for more this year than last.
• It is not "fraud" in any legal sense. Every change was disclosed with its dollar effect. The real question is whether the earnings the market is paying for are front-loaded.

What Burry actually said

The claim came in a post on X, reported by Investing.com:

"Understating depreciation by extending useful life of assets artificially boosts earnings - one of the more common frauds of the modern era. Massively ramping capex through purchase of Nvidia chips/servers on a 2-3 yr product cycle should not result in the extension of useful lives of compute equipment. Yet this is exactly what all the hyperscalers have done... By my estimates they will understate depreciation by $176 billion 2026-2028. By 2028, ORCL will overstate earnings 26.9%, META by 20.8%, etc."

Three things matter here. First, the target is accounting estimates, not hidden numbers. Second, the core premise is economic: Nvidia now releases a new generation every year or two, so a server bought today may be obsolete well before it is fully depreciated. Third, he did not publish how he got from that premise to $176 billion, 26.9% or 20.8%.

He backed the view with money. Scion Asset Management's last quarterly holdings report, for the quarter ended 30 September 2025, showed put options on 5 million Palantir shares and 1 million Nvidia shares, with notional values of $912.1 million and $186.6 million (Sherwood News). Notional value is the value of the shares the options cover, not the cash he paid. Days later, Scion's registration as an investment adviser was terminated, and Burry turned to writing.

What depreciation is, and why the life matters

When a company buys a $1 billion server fleet, it doesn't book a $1 billion expense that year. It spreads the cost over the years it expects to use the hardware. That yearly charge is depreciation. Choose six years and the charge is about $167 million a year. Choose four and it's $250 million.

Nothing about the cash changes. The full $1 billion left the bank on day one. What changes is when the cost shows up in reported profit. A longer life makes today's earnings bigger and pushes cost into later years. If the hardware really does last longer, that's accurate. If it's replaced sooner, the remaining cost arrives later as faster depreciation or a write-down.

What the filings show

Here is every server useful-life change we could confirm in the companies' own annual reports and releases:

Add up just the first-year effects disclosed by Microsoft, Alphabet and Meta and you get roughly $10 billion of depreciation that never hit the income statement in the year of each change. That effect grows as the server fleet grows, and the fleet is growing very fast.

Microsoft has since extended another estimate. From fiscal 2027 it will depreciate data center and office buildings over 25 years instead of 15. Its CFO described the profit benefit as "minimal" (CFO Dive). Buildings last far longer than chips, so this change is easier to defend, but it moves in the same direction.

Amazon's reversal is the tell

Accounting lives are management estimates, and companies review them every year. For five years, every review at these companies ended the same way: longer. Amazon's 2025 decision is the first to go the other way, and its explanation is almost Burry's argument in the company's own words: AI is making hardware obsolete faster.

Amazon only shortened a subset of its fleet, and only by one year. But it shows that at least one of the people with the best view of real server usage concluded that six years was too long for some AI equipment.

Depreciation is already catching up

Even with longer lives, depreciation is rising fast because so much new hardware is going into service:

Sources: Meta's Q2 2026 10-Q, Alphabet's Q2 2026 10-Q.

That's before most of the record 2026 spending starts depreciating. As we showed in AI spent $132 billion; earnings show only part of the bill, assets still under construction aren't depreciated until they're placed in service. Neither company's latest quarterly report shows a new change to server lives in 2026.

Does $176 billion add up? Our rough check

Burry didn't show his math, so we built the simplest version ourselves. Treat this as a sense check, not a forecast.

• Spending. Microsoft, Alphabet, Amazon and Meta put about $131.6 billion into infrastructure in the first quarter of 2026. Annualized, that's about $526 billion.
• Equipment share. We assume half of that, about $263 billion a year, is servers and network gear. The rest is buildings, land and power. The real split varies by company and isn't fully disclosed.
• Timing. We assume the same amount each year from 2026 to 2028, ignore hardware bought before 2026, and leave out Oracle.

Under those assumptions, the gap between a six-year life and a shorter "true" life looks like this:

The batches stack up. In the four-year case, 2026 carries one batch of extra charges, 2027 two and 2028 three, so $22bn × (1 + 2 + 3) ≈ $132bn.

Burry's $176 billion falls between our four-year and three-year cases. So his number isn't wild. It's what you get if you believe AI servers lose their economic value in roughly three to four years. The whole debate comes down to that one assumption.

The case against Burry

The hardware is still earning money. In a memo reported by CNBC, Nvidia said older A100 chips "continue to run at high utilization rates and retain meaningful economic value." The rental market backs that up for now. GetDeploying's tracker puts the median on-demand price of an H100, launched in 2022, at $3.39 an hour in the last week of September 2026, up 14% from a year earlier (GetDeploying). If three-year-old chips are renting for more each year, a five- or six-year life isn't absurd.

Old chips move to easier work. A data center can train new models on the newest chips and run inference, the day-to-day work of answering requests, on older ones. Microsoft CEO Satya Nadella has said he deliberately spreads purchases across generations because he "didn't want to go get stuck for four or five years of depreciation on one generation."

Nothing was hidden. Each change appears in the annual report with its dollar effect. Calling it "fraud" overstates it. The fair criticism is that the estimates are optimistic, which investors can adjust for themselves.

Depreciation timing moves profit between years; it doesn't destroy it. If lives turn out too long, earnings are higher now and lower later. Value is destroyed only if the hardware earns less than it cost, which is a question about AI demand rather than accounting.

How his bets have done

Burry's accounting critique and his trades are separate questions, but readers ask about both. Using closing prices:

The Nvidia and Palantir puts from Scion's last filing haven't paid off on these prices, although we don't know exactly when he traded. Oracle, the company he singled out as most exposed, has fallen by half. In August 2026 he wrote that he had closed Oracle puts at a "substantial" profit and opened a new short in the stock, and in Nebius, another AI cloud provider (Business Insider).

Oracle fits his argument best. It depends on borrowed money and long leases to build AI capacity, its margins are thinner than the big platforms', and it uses a six-year server life. When the depreciation bill arrives, it has the least room to absorb it.

Our read

Burry is right about the direction and unproven on the size. The useful-life extensions were real, they flattered earnings by billions a year, and Amazon's reversal shows at least one insider thinks six years is too long for some AI hardware. Depreciation is already growing around 40% a year, and the biggest wave of spending hasn't reached the income statement yet.

But his $176 billion depends on AI servers losing their value in about three to four years, and the rental market doesn't show that yet. The practical takeaway is narrower than "the AI boom is a fraud." Price-to-earnings ratios for the biggest AI spenders are calculated on earnings that assume long hardware lives. If those lives shorten, the same business looks more expensive.

The most exposed companies are those that combine long accounting lives, heavy borrowing and thin margins. That points to Oracle and the newer AI cloud providers more than Microsoft or Alphabet.

What we'd watch

1. Another company shortening server lives. Amazon went first. A second would confirm the trend.
2. Rental prices for two-generation-old chips. If H100 and A100 prices start falling fast, Burry's three-to-four-year case gets stronger.
3. Depreciation growth versus cloud revenue growth. If depreciation keeps growing around 40% a year while cloud revenue grows more slowly, margins get squeezed.
4. Write-downs and early retirements. These show up when hardware is retired before its accounting life ends.
5. Construction in progress. It's the depreciation already locked in for the next few years.

You can test each scenario yourself with the free AI capex, depreciation and return model: set the equipment life to four years and watch what happens to earnings.

Questions readers ask

What is Michael Burry's Substack?

Cassandra Unchained, which he launched in November 2025 after closing Scion to outside investors. He reported 300,044 subscribers after 231 days (Burry). He hasn't disclosed how many pay.

Is extending useful lives illegal?

No. Accounting rules require companies to estimate useful lives and update them when evidence changes, with the effect disclosed. The debate is whether the estimates are realistic, not whether they're allowed.

Which companies are most exposed if server lives shorten?

The ones where depreciation is a large share of costs relative to profit. Burry named Oracle (26.9% overstatement by 2028) and Meta (20.8%). We can't reproduce those figures because he didn't publish his method, but Oracle's leverage and thin margins make it the most sensitive by our read.

Does this mean Nvidia is a bubble?

Not directly. Shorter lives would lower the earnings of Nvidia's customers, not Nvidia's. The risk to Nvidia is indirect: if customers conclude each generation wears out faster, they may slow purchases to protect returns.

Go deeper with member research

• Issue 03: The AI bill is arriving late to earnings: the full depreciation clock for all four hyperscalers, company by company, with the earnings tests that come next.
• Issue 05: Nvidia's credit layer: how the AI buildout is being financed, and where borrowed money makes the depreciation risk sharper.
• Meta jumped 11%. What must happen next to justify the price?: one of Burry's named companies, valued with the depreciation bill included.

Limits of this data

• Our $176 billion check assumes a fixed spending level, a 50% equipment share and no pre-2026 hardware. Change any of these and the totals move a lot. It shows what the claim implies; it doesn't verify it.
• Burry's company-level figures can't be reproduced because his method isn't public.
• GPU rental prices vary by provider, contract length and region. The figure above is one tracker's median on-demand price.
• Burry's trade timing and profits are self-reported or inferred from filings. We used quarter-end closing prices.