For most of its history, the useful life of a server has been a topic reserved for technical accounting teams and ignored by everyone else. Over the past year it became a market-moving question, the subject of a public fight between a famous short seller and the largest companies in the world, and the proximate cause of a sixty percent collapse in one listed company's share price. There is a lesson in it for businesses that will never own a GPU, and it is not the lesson the headlines suggest.
Key Takeaway
On November 11, 2025, Michael Burry publicly argued that hyperscalers were inflating earnings by depreciating AI hardware over five to six years when its real economic life is closer to two or three, estimating roughly $176 billion of understated depreciation across 2026 to 2028 and profit overstatement above 20%. The industry's defence, including Nvidia's, is that observed utilization supports four-to-six-year lives. The decisive evidence for how to read this is not either argument but a divergence: in 2025 Amazon completed a useful life study and shortened the life of a subset of servers, while Meta extended its estimate further, under identical technological conditions. That confirms useful life is a subjective management estimate rather than an observable fact. For a Canadian business the transferable lesson is that depreciation policy is a judgment that flows directly to reported earnings, that yours was probably set years ago and never revisited, and that the same reasoning applies to your own technology assets at a smaller scale.
The Cleanest Fact In The Whole Debate
Before the arguments, the observation that settles what kind of question this is.
Between 2020 and 2024, many large technology companies steadily extended the useful lives of their servers and networking equipment, only for the trend to diverge in 2025, when Amazon shortened the useful life of a subset of servers while Meta extended its estimate further[1].
Two of the most sophisticated finance organizations in the world, operating comparable hardware in the same market at the same moment, reached opposite conclusions about how long that hardware lasts. As one analysis puts it, this divergence occurring under identical technological conditions confirms that useful life is a subjective management estimate reflecting corporate strategy and accounting conservatism, or the aggressive lack thereof[2]. The same source quantifies the gap between the two positions as a $3.6 billion earnings quality difference, and argues that gap is the story rather than the absolute depreciation numbers[2].
This matters for how a reader should approach everything that follows. If useful life were an observable engineering fact, one of these companies would simply be wrong. Because it is an estimate about future economic benefit, both can be defensible while producing materially different earnings. That is the nature of the accounting, not a defect in it.
What Burry Actually Said
The claim, stated precisely, since it is frequently paraphrased loosely.
On November 11, 2025, Michael Burry, manager of Scion Capital and known for predicting the 2008 housing collapse, posted on X that hyperscalers were artificially boosting earnings by extending useful life assumptions beyond what two-to-three-year product cycles justify[3]. The argument is that large cloud providers flatter earnings by depreciating Nvidia-based data centre hardware over five or six years even though Nvidia's fast chip cycle means the real economic life is closer to two or three[4].
The quantified estimate: roughly $176 billion of understated depreciation and overstated profits across the industry between 2026 and 2028[4], with one summary adding that this implies hyperscalers overstating profits by more than 20%[5].
Burry also made a separate criticism of Nvidia regarding stock-based compensation destroying owner's earnings, and drew a specific historical comparison, likening the setup not to Enron but to Cisco at the peak of the dot-com bubble, where exuberant capital spending and optimistic assumptions later unravelled[4]. That comparison is worth noting because it is considerably milder than the fraud framing often attributed to him: Cisco's problem was over-optimism, not deception.
The Mechanism, In Plain Terms
For readers who do not spend their days in fixed asset registers, the arithmetic is simple and worth stating explicitly.
Depreciation spreads the cost of an asset across the period it is expected to generate benefit. A $6 million asset depreciated over three years produces $2 million of annual expense; the same asset over six years produces $1 million. The asset, the cash outflow and the economics are identical. Only reported profit differs, by $1 million a year.
Extending useful life therefore reduces annual depreciation expense and increases reported earnings, without any change in cash. Burry's argument is that hyperscalers ramped capital expenditure to unprecedented levels while simultaneously lengthening depreciation schedules to five and six years, which mechanically flatters earnings[5].
The consequence if the shorter life is correct is not that the expense disappears. It arrives later and abruptly. As one analysis puts it, if the true economic life is closer to three years a cliff eventually arrives, either via large impairments when stranded chips are written down, or via the realization that what is considered growth capex is actually permanent sustaining capex, locking these companies into structurally lower free cash flow[5].
The Industry Answer
The counterargument deserves fair statement, and it is not weak.
Nvidia pushed back, arguing that customers consistently use four-to-six-year depreciable lives based on observed utilization and longevity[1]. The industry position more broadly is described as a computing cascade: even where a chip generation is no longer state of the art for frontier training, it remains economically useful for less demanding workloads, inference, internal applications and smaller customers, so the asset continues generating benefit well past its front-line obsolescence[2].
This is a coherent argument and it maps onto how depreciation is supposed to work. Useful life is not the period during which an asset is best-in-class; it is the period over which the entity expects to derive economic benefit. A GPU displaced from training work but running inference profitably for three more years is still generating benefit, and depreciating it as though it were scrapped would misstate the position in the other direction.
One balanced assessment splits the difference on plausibility: extending server lives from three to four years as data centre design improves and workloads diversify is plausible, while stretching all the way to five or six years for rapidly advancing AI accelerators is more debatable, especially given effective economic life shrinking for leading-edge GPUs as upgrades accelerate[5].
The Quote That Complicated The Defence
One remark from inside the industry gave the sceptical case more weight than any external analysis.
Microsoft CEO Satya Nadella said: "I didn't want to go get stuck with four or five years of depreciation on one generation"[2]. The analysis reporting it observes that this single sentence, acknowledging technological obsolescence faster than accounting schedules, validates the question Burry posed[2].
The remark is significant because it comes from an operator describing a commercial concern rather than an accounting position. Someone worried about being stuck with four or five years of depreciation on one generation is describing a generation that stops being useful before four or five years elapse, which is the premise of the sceptical argument rather than the industry defence.
It would be unfair to treat one comment as an admission against interest on a technical accounting matter, and we do not. But it does illustrate a tension that runs through the whole debate: the operational language executives use about hardware refresh cycles frequently sits awkwardly beside the useful lives their financial statements assume.
Why This Is Not A Fraud Story
The framing matters, and one analysis puts the point better than we could.
Public discourse has devolved into a noisy binary: either hyperscalers are committing massive fraud, or their six-year schedules are perfectly justified by the computing cascade, and both extremes are oversimplifications that obscure the nuanced reality, with the allegation of fraud being hyperbolic[2]. A second commentator reaches the same conclusion independently: it is a judgment call, not outright cooked books, so far[5].
The reason is structural rather than charitable. Useful life is an estimate that management is required to make, reviewed by auditors against a reasonableness standard rather than verified against a fact. Two managements can reach different estimates on the same asset class and both can pass audit, which is precisely what the Amazon and Meta divergence demonstrates. Calling an estimate at the aggressive end of a defensible range "fraud" misdescribes the mechanism and, more practically, makes the situation sound rarer than it is.
The more useful characterization is that this is a story about the width of the range within which reasonable estimates sit, and about the fact that a company's position within that range is a choice with earnings consequences.
The Conflict Worth Naming
Intellectual honesty requires stating the obvious about the source of the argument.
Burry has short positions in various AI-related companies and would benefit from a decline in AI stocks, and his accusation is an accusation, with the possibility that there may be a reasonable rationale behind the useful lifetimes being used[6]. Concretely, in September 2025 Scion disclosed roughly $1.1 billion notional in put options against Nvidia and Palantir, about 79.5% of his reported portfolio, and by November he shuttered the fund[7].
That is a substantial financial interest in the conclusion. It does not make the analysis wrong, and one commentator notes that while the specific $176 billion figure is difficult to verify, the directional logic is sound[7], with another describing the estimate as widely circulated but unconfirmed in details[7].
The appropriate posture is the one a reader should adopt toward any interested analysis: weigh the argument on its structure, treat the specific quantum sceptically, and pay more attention to verifiable disclosures than to estimates. Which is what the leading-indicator section below recommends.
CoreWeave And What Happened Next
The debate stopped being theoretical when it met a company whose economics depend entirely on the answer.
Jim Chanos turned his attention to CoreWeave, arguing from public filings and conservative assumptions, even a generous ten-year GPU life, that the numbers did not add up: annualized adjusted EBITDA around $3.4 billion versus roughly $1.2 billion in interest on an estimated $20 billion of GPU assets[7]. CoreWeave fell from $187 in June 2025 to $72 by mid-December, a 61% collapse erasing about $33 billion in market capitalization in six weeks[7].
Two features of CoreWeave's position made it the sharpest test case. It had implemented a pre-IPO increase in useful lifetimes[6], and it is not profitable on a GAAP basis as it stands, meaning those losses would look worse if depreciation expense were higher[6]. A six-year useful lifetime for a category dominated by data centre GPUs does seem at least somewhat optimistic, and if the company eventually reduces it to reflect real-world replacement rates it would have a large negative impact on the bottom line[6].
The generalizable point for any business is that depreciation assumptions matter most where they interact with leverage. A company with substantial debt service against depreciating assets has very little margin for an estimate that proves optimistic, because the interest obligation is contractual while the earnings supporting it are partly an assumption.
The Blind Spot Almost Nobody Mentions
A technical point with large forecasting consequences, and the most underreported element of the whole subject.
Under GAAP, balances are not depreciated until the assets are placed in service, meaning today's depreciation expense reflects past investment cycles and may not fully reflect the current wave of AI-driven infrastructure spending; given the materiality of AI-related capex and the relatively long period required to place some of these assets in service, construction-in-progress balances are increasingly significant for forecasting depreciation trends[1].
Follow the implication. A company that spent enormous sums in 2025 on data centres not yet operational is not yet depreciating any of it. Its current reported depreciation reflects an older, smaller asset base. When those assets are placed in service, depreciation steps up mechanically, entirely independent of whether the useful life assumption is right or wrong.
So there are two distinct effects being conflated in public discussion: a possible understatement from optimistic useful lives, and a certain future increase from assets not yet in service. The second requires no controversy at all, and an analyst looking only at current depreciation relative to current capex will misjudge the trajectory regardless of where they land on the first question.
The Leading Indicator
The most practically useful idea in the literature, and one that converts an unresolvable argument into something observable.
One analysis argues that the instinct is to wait for AI revenue to validate the spending or for productivity statistics to confirm the picture, but those are lagging signals, while the leading one is already public: the first place the thesis breaks will not be a revenue miss, it will be a depreciation disclosure. Amazon shortening a subset to five years was the bull case quietly cracking in real time, ahead of any income statement, and the next hyperscaler to shorten useful life, or to take an asset retirement charge on stranded silicon, is telling you something[8].
This is genuinely good analytical advice and it generalizes. A change in accounting estimate is a disclosure a company must make, and it reveals management's revised view of the future before that revised view shows up in operating results. Anyone tracking this sector, whether as an investor, a customer negotiating multi-year commitments, or a supplier assessing counterparty durability, should be reading estimate-change disclosures rather than waiting for revenue.
Note that Amazon's move came in February 2025, described as the first crack[7], roughly nine months before Burry's post made the issue famous. The signal preceded the noise by three quarters.
Growth Capex Or Sustaining Capex
A distinction that determines what these businesses are actually worth, and which applies equally to a small business buying equipment.
The sceptical case includes the possibility that what is considered growth capex is actually permanent sustaining capex, locking companies into structurally lower free cash flow[5].
The distinction is fundamental to valuation. Growth capex is discretionary spending that expands capacity and can be reduced without shrinking the business. Sustaining capex is the spending required to stand still. A business whose assets last six years replaces roughly a sixth of them annually; one whose assets last three years replaces a third. If the shorter figure is right, a large part of what has been presented as investment in growth is actually the cost of maintaining current capability, and the free cash flow available to shareholders is correspondingly smaller and permanently so.
The same question is worth asking about any capital-intensive business, including a small one. Equipment replaced every three years because it wears out is not growth investment however it appears in a capital budget, and a business that has been treating recurring replacement as discretionary expansion has been overstating its own free cash flow to itself.
What This Means For Your Own Books
The transferable content for a Canadian business that owns no data centres.
Depreciation policy is a management estimate at your scale exactly as it is at Meta's. The rates in your fixed asset register were chosen by someone, often years ago, sometimes by a bookkeeper applying a default, occasionally by matching capital cost allowance classes that exist for tax purposes rather than to reflect economic life. Very few small and mid-sized businesses revisit those rates, and both IFRS and Canadian accounting standards contemplate that useful life estimates be reviewed rather than set once and left, a point worth confirming with your accountant for your specific framework.
The categories where the estimate is most likely to be wrong are the ones this debate concerns: technology assets. Computers, servers, software and specialized equipment frequently carry rates set when replacement cycles were longer. If your business is replacing laptops every three years and depreciating them over five, your reported earnings are flattered and your asset register overstates what you own, in exactly the manner under dispute at a vastly larger scale.
The consequences are not merely presentational. Depreciation affects reported EBITDA, which affects covenant calculations, which affects lender relationships, and it affects the earnings figure a buyer applies a multiple to. As this publication has discussed in the context of exit multiples, a buyer's quality of earnings review will examine whether depreciation reflects genuine replacement cycles, and a business with understated depreciation is presenting an EBITDA that a diligence process will adjust downward.
A Worked Case: The Policy Nobody Revisited
A Canadian professional services firm with a significant technology asset base: laptops, servers, and specialized software. The reconstruction below illustrates a common pattern rather than reporting a specific engagement.
Its depreciation rates had been established roughly a decade earlier, applying straight-line lives that matched the tax classes its accountant used rather than any assessment of how long the equipment actually lasted. Nobody had revisited them, because nothing prompts a review of depreciation policy in a business where the accounting is otherwise routine.
In practice the firm replaced staff laptops on a three-year cycle and its specialized software was subscription-based and rewritten roughly every four years. Its books depreciated the hardware over five years and carried capitalized software development over a longer period still. The result was a fixed asset register showing net book value for equipment that had been disposed of or superseded, and reported EBITDA modestly higher than the firm's actual replacement economics supported.
The issue surfaced during a lender's review, when a covenant calculation based on EBITDA came under scrutiny and the lender's analyst asked how the firm's depreciation compared to its actual replacement spending. The gap was not large in absolute terms and nothing improper had occurred. But it was an avoidable conversation, and correcting it afterward under scrutiny was less comfortable than reviewing the policy would have been.
The Customer Angle
One reason a Canadian business should care about hyperscaler economics even as a pure customer.
Hyperscalers have collectively committed to over $300 billion in AI infrastructure capex for 2025 alone, with the majority allocated to GPU acquisitions[7]. That capital has to earn a return, and the mechanism by which it does is the pricing of cloud and AI services to customers.
If the sceptical case is directionally right and useful lives compress, the cost base underlying those services is higher than currently reported, which puts upward pressure on pricing over time. That connects to the software repricing dynamic examined elsewhere in this publication, where AI-related uplifts of 20 to 37 percent have been appearing on enterprise renewal quotes. A business negotiating multi-year commitments for AI-dependent services has an interest in whether its provider's underlying economics are as favourable as reported.
We would not overstate this: the transmission from depreciation assumptions to customer pricing is indirect and slow, and competition may absorb it. But a buyer signing a long commitment on the assumption that current pricing reflects a stable cost base should at least know that the stability of that cost base is publicly contested.
The Limits Of This Analysis
Several caveats matter. The $176 billion figure is an estimate by an interested party, described in the sources themselves as difficult to verify and unconfirmed in details, and it should not be treated as a measurement. Most sources cited here are market commentary, Substack analysis and financial media rather than company filings, and we have not verified the specific useful life changes, quoted figures or share price movements against primary filings or market data. Michael Burry held short positions in AI-related companies and Jim Chanos is a short seller; both had financial interests in their conclusions, which we have flagged rather than discounted. Accounting standards differ between US GAAP, IFRS and Canadian ASPE in ways this article does not address, and the discussion of depreciation review obligations is general rather than a statement of any particular standard's requirements. Nothing here is accounting, audit or investment advice, and any change to depreciation policy should be made with your accountant and considered for its covenant, tax and comparative-period consequences.
Frequently Asked Questions
What is the AI depreciation debate about?
Is anyone committing fraud?
Should I discount Burry's argument because he was short?
What is the most reliable thing to watch?
Why does construction in progress matter?
What should a small business take from this?
References
- National Law Review. Deep Quarry: Useful Lives Of GPUs, Key Considerations, on the 2020-2024 extension trend, the 2025 divergence, Nvidia's pushback via CNBC, and construction-in-progress treatment under GAAP. natlawreview.com/article/deep-quarry-useful-lives-gpus-key-considerations
- Stanley Laman Group. (2025, November 21). Why GPU Useful Life Is The Most Misunderstood Variable In AI Economics, including the Nadella quote, the subjective-estimate conclusion, the $3.6 billion earnings quality gap and the false-binary argument. stanleylaman.com/signals-and-noise/gpus-how-long-do-they-really-last
- Footnote Brief. (2026, May 21). The $200 Billion Question Hiding In Big Tech's AI Spending, on the November 11, 2025 post and cross-company aggregation caveats. footnotebrief.com/hyperscaler-depreciation-ai-capex-circularity
- Deep Quarry. (2025, December 7). Depreciation Of GPUs: Between Useful Lives And Useful Myths, on the $176 billion estimate, the stock-based compensation criticism and the Cisco comparison. deepquarry.substack.com/p/depreciation-of-gpus-between-useful
- Levelheaded Investing. (2025, December 4). Are AI Chip "Useful Lives" Creating Useless Earnings?, on the profit overstatement estimate, the impairment cliff, growth versus sustaining capex and the judgment-call assessment. levelheadedinvesting.com/p/are-ai-chips-useful-lives-creating-useless-earnings
- The Motley Fool. (2025, November 11). Michael Burry's Latest Warning Could Be Bad News For CoreWeave, on Burry's short positions, the pre-IPO useful life increase and CoreWeave's GAAP position. fool.com/investing/2025/11/11/michael-burrys-latest-warning-could-be-bad-news-fo
- Outerspeak. (2026, February 14). The Melting Ice Cube: Michael Burry And The Great AI Depreciation War, on Scion's disclosed positions, the Chanos CoreWeave analysis, the share price collapse, the Amazon February 2025 useful life study and the $300 billion capex figure. outerspeak.substack.com/p/the-melting-ice-cube-michael-burry
- K4I. (2026, June 29). AI Benefits Outrun Capex Only If GPUs Last Six Years. Burry Says Three., on the leading-indicator argument and the capex step function. k4i.com/ai-benefits-outrun-capex-only-if-gpus-last-six-years
This article discusses market commentary and reported company disclosures and is provided for general informational purposes. It is not accounting, audit or investment advice, and does not constitute a recommendation regarding any security. Figures reported here derive from secondary commentary rather than primary filings and several originate from analysts holding short positions; verify against filings before relying on them.