Investing Frameworks11 min read
The $176 Billion Question: AI's Biggest Earnings May Be Hiding in a Depreciation Footnote
Burry estimates that by stretching the useful life of AI chips, the hyperscalers will understate depreciation by roughly $176 billion from 2026–2028 — inflating profits by the same amount. This isn't a conspiracy theory; the numbers are public, audited, and sitting in plain sight in the footnotes. Using Buffett's "owner earnings," Munger's accounting skepticism, and Damodaran's "story vs. numbers," we decode the most overlooked crack in the AI boom.
A Footnote Nobody Wants to Read Holds the AI Boom's Biggest Question Mark
In November 2025, Michael Burry — famous for shorting the 2008 housing market — broke his silence on AI and aimed at the giants. His accusation sounds technical and boring, yet it may be the single most important sentence of this entire AI rally: the hyperscalers are using "depreciation," an accounting lever, to quietly shift enormous outlays that belong in the cost column over into the profit column.
The logic isn't complicated. A GPU costing tens of thousands of dollars isn't expensed all at once in the year of purchase; its cost is spread over its "useful life." Depreciate a $40,000 chip over 3 years and the annual expense is roughly $13,333; over 6 years, it's about $6,667. Same cash outlay, same chip — but flip one assumption and reported profit is worlds apart.
And the scale of that assumption is staggering. In 2025, the five big hyperscalers are expected to spend roughly $300 billion in capex, up sharply from $217 billion in 2024, with a large chunk aimed at AI chips. At that magnitude, every extra year on the depreciation schedule moves billions from "expense" to "profit." Burry's conclusion: across 2026–2028, the industry cumulatively understates depreciation by about $176 billion — precisely where a value investor should press the magnifying glass.
Same Chip, Opposite Answers: Depreciation Is a "Choice," Not a "Fact"
The most damning evidence came in early 2025, when Amazon and Meta made completely opposite calls on the very same technology. Amazon shortened the useful life of a subset of its servers from 6 years to 5, stating it plainly — "the increased pace of technology development, particularly in artificial intelligence and machine learning" — and took a ~$700 million hit to operating income for it.
At almost the same time, Meta extended its server lives to 5.5 years, booking a $2.9 billion reduction in depreciation expense — nearly 4% of that year's estimated pre-tax profit. Same Nvidia chips, same technological environment, one shortened, one extended. What does that tell you? That within the guardrails of GAAP, "useful life" is largely management's discretion, not a physical fact etched into the silicon.
Aswath Damodaran — the dean of valuation — loves to remind investors that financials sit at the intersection of "story" and "numbers," and management always has an incentive to make the numbers serve the story. When a company's story is "AI, high growth, high profit," and it conveniently "discovers" its chips last another year and a half, the question isn't "is it legal." It's this — is that assumption reflecting economic reality, or managing earnings?
Burry's Arithmetic: Oracle's Profit May Be Overstated 27% by 2028
Scale the logic to the whole industry and the numbers get uncomfortable. Say a company invests $10 billion in GPUs: under conservative 3-year depreciation, annual expense is roughly $3 billion; under aggressive 6-year depreciation, only about $1.2 billion. That $1.8 billion annual gap flows almost untouched into operating income — after tax, about $1.4 billion in "extra" profit that comes purely from accounting treatment, not real operations.
From this, Burry estimates that if the true life is 3 years rather than 6, the industry understates depreciation by roughly $50–60 billion a year — compounding over three years to right around his $176 billion figure. At the single-stock level, he calculates that by 2028 Oracle's earnings could be overstated ~26.9% and Meta's ~20.8%. Which means: you think you're buying a "high-profit" company at 30x earnings, when the true multiple may be far higher.
The bulls have their rebuttal, of course — the "value cascade" theory holds that year-one and -two frontier training chips step down to years three and four for inference, then to years five and six for batch processing, so lives are naturally longer. CoreWeave's CEO says returned H100s were re-leased at 95% of original price. But this is exactly the point of Damodaran's frame: both sides are telling "stories." Your job is to translate each story into a concrete depreciation assumption, then watch where it pushes the valuation.
Buffett's "Owner Earnings": Profit Can Be Dressed Up, Cash Cannot Lie
The ultimate arbiter of this debate is a concept Buffett stressed his whole life — "owner earnings." He never trusts accounting profit on its own; he demands: net income, plus depreciation and amortization, minus the true capex a business needs to maintain its long-term competitive position. The whole game is in that last term: if a chip must be replaced every 2–3 years to stay competitive, then that continuous, enormous capex must be truly deducted from profit — no matter how many years the financials stretch the depreciation over.
Here's a brutal real-world data point: an H100's rental rate fell from over $8/hour in 2023 to $1–2 at auction prices by late 2024. The chip still boots up, still runs "fully booked" — but the money it earns per hour has been halved and halved again. That is exactly what owner earnings is built to catch: an asset's earning power is depreciating fast, and that should govern the amortization pace far more than its physical lifespan. Even Microsoft CEO Satya Nadella admitted it: "Nvidia's pace increased… I didn't want to get stuck with four or five years of depreciation on one generation."
Charlie Munger would put it more bluntly. He kept a near-instinctive suspicion of accounting tricks his whole life, famously saying: "Every time you see the word EBITDA, substitute the words 'bullshit earnings.'" To him, talking about profitability while excluding the most real cost of a capital-heavy industry — depreciation — is a dangerous act of self-deception. When an entire industry leans toward "stretch it longer," the Munger response isn't anger; it's heightened alert: if profit looks this good, where did the cash flow go?
What Ordinary Investors Should Do: Read the Footnote as a "Profit Stress Test"
So, facing this "$176 billion" case, what an ordinary investor should do isn't pick a side and scream bull or bear — it's change how you read the financials. This article won't tell you to buy or sell any AI stock — but it wants to hand you a check-up you can run yourself.
Step one: find that footnote. Turn to the "Property, Plant & Equipment" depreciation policy in the annual report and see how many years it spreads servers and equipment over. Step two: run a stress test — cut the useful life from 6 years to 3, recompute profit and free cash flow, and see how far the valuation collapses. Step three: reconstruct with owner earnings — net income + D&A − (the true capex needed to stay competitive) — and see how much real gold is left. Those three steps turn a scary headline into a cash-flow balance sheet you calculated yourself.
This is exactly the skill Valuefocus is built to give you. When the market's social sentiment is screaming over AI revenue and arguing whether Burry is a prophet or a doomsayer, our intrinsic value calculator lets you adjust the depreciation and capex assumptions yourself and see what the same company is worth under "aggressive" versus "conservative" accounting. Then the master frameworks take over, guiding you to use Buffett's owner earnings, Munger's accounting skepticism, and Damodaran's story-vs-numbers to read a footnote nobody wants to read — into the question the AI boom most needs answered: are these profits earned, or calculated?
Recent posts
%22%2F%3E%3Cpath%20d%3D%22M0%20610%20370%20370l300%20160%20400-330%20530%20360%22%20fill%3D%22none%22%20stroke%3D%22%23c99a62%22%20stroke-width%3D%2218%22%20opacity%3D%22.65%22%2F%3E%3Ccircle%20cx%3D%221070%22%20cy%3D%22200%22%20r%3D%2246%22%20fill%3D%22%23c99a62%22%2F%3E%3C%2Fsvg%3E)
%20-1.jpg?width=1600)
Frequently asked questions
What exactly is Michael Burry accusing them of?
Per Burry's public analysis from November 2025: hyperscalers like Meta, Amazon, Microsoft, Google and Oracle depreciate their Nvidia GPUs over 5–6 years, while the chips' true economic life is closer to 2–3 years. He estimates this understates depreciation by about $176 billion over 2026–2028, inflating profits by the same amount, and calculates that by 2028 Oracle's earnings could be overstated ~26.9% and Meta's ~20.8%. This article is for educational purposes only and is not investment advice.
Why does stretching the depreciation life "inflate" profit?
Because a GPU's cost is spread across its useful life as annual expense. The same $40,000 chip costs ~$13,333/year over 3 years but only ~$6,667/year over 6. The longer you stretch it, the less cost you book each year, so reported profit rises — the cash outlay is unchanged, but statement profit gets pulled up.
What framework best evaluates this kind of accounting dispute?
Stack three master lenses: Buffett's "owner earnings" — net income plus D&A, minus the true capex needed to stay competitive, to recover real cash; Munger's accounting skepticism — an instinctive distrust of "profit" that excludes depreciation; and Damodaran's "story vs. numbers" — translate each side's narrative into a depreciation assumption and watch where it drives the valuation.
Are these AI profits earned, or calculated? Use Valuefocus to read a depreciation footnote as a calm profit stress test:
Download Valuefocus