From Shubhi K | Product & Market Analysis
GPU Depreciation Schedules: The Accounting Choice Holding Up AI Earnings
On this page
Hyperscalers depreciate AI hardware over five to six years. Nvidia ships a new architecture roughly every year. That gap between the accounting schedule and the technology cycle is worth an estimated $176 billion of understated depreciation across the industry between 2026 and 2028, and it is the single most consequential judgement call in AI financial reporting.
Key takeaways
- Useful life is an estimate, not a fact. Companies choose how long they expect hardware to generate benefit, and that choice moves reported profit without changing the business.
- The contested figure is $176 billion. Michael Burry estimated that understated depreciation would inflate industry earnings by that amount between 2026 and 2028, with Oracle and Meta most exposed.
- The companies do not agree with each other. Amazon shortened the useful life of some servers in 2025 while Meta extended its estimate further, under identical technological conditions.
- Depreciation is non-cash, which is the strongest counter-argument. Alphabet generated roughly $165 billion of operating cash flow in 2025 regardless of how the schedule was set.
How the mechanic actually works
Depreciation spreads the cost of a long-lived asset across the period it is expected to generate benefit. Buy a server for $60,000 and depreciate it over six years, and $10,000 hits the income statement each year. Depreciate the same server over three years, and $20,000 does.
Nothing about the business changes. The cash left the building on the day of purchase either way. What changes is reported profit, and reported profit is what most valuation models are built on.
Useful life is a management estimate. Accounting standards require it to reflect the period over which the asset is expected to be used, supported by evidence. Auditors test that evidence. They do not set the number.
Why AI hardware makes this contentious
Most equipment ages predictably. Accelerators do not. A new architecture arriving annually means a chip can remain physically functional while becoming uneconomic for the workload it was bought for.
That creates a genuine question with no clean answer. Is the useful life the period before the chip stops being competitive at frontier training, or the period before it stops generating any revenue at all? Those are very different numbers and both are defensible.
The $176 billion claim
In November 2025, Michael Burry argued publicly that hyperscalers were inflating earnings by depreciating Nvidia-based hardware over five to six years when the real economic life sat closer to two or three.
His arithmetic put the gap at roughly $176 billion of understated depreciation across the industry between 2026 and 2028. He singled out Oracle and Meta, estimating their profits could be overstated by around 27% and 21% respectively by 2028.
He also disclosed short positions in AI-exposed names, which is a material fact when weighing the argument. It does not make the accounting point wrong. It does mean the framing was chosen by someone with a position.
The claim moved a niche accounting topic into mainstream coverage within days. That is worth noting on its own, because the underlying disclosures had been public and audited the entire time.
The companies do not agree with each other
This is the detail that makes the debate real rather than rhetorical.
Between 2020 and 2024, large technology companies steadily extended the useful lives of servers and networking equipment. Then in 2025 the trend split. Amazon shortened the useful life of a subset of servers. Meta extended its estimate further.
Two companies, the same hardware generation, the same market conditions, opposite conclusions. That divergence is the clearest available evidence that useful life is a judgement reflecting management strategy rather than an observable property of the asset.
Microsoft's chief executive has publicly described not wanting to be left carrying four or five years of depreciation on a single hardware generation. That is an operating executive acknowledging the same tension the accounting debate is about.
The defence, stated properly
The industry response is not a denial. It is a different model of how the hardware is used.
The cascade argument
An accelerator does not go from frontier training to scrap. It moves down a chain. Frontier training first, then inference on production workloads, then serving smaller models, then general compute. Each step is less demanding and each still generates revenue.
If that cascade genuinely operates, six years of economic benefit is a fair estimate even though eighteen months of frontier competitiveness is also true. Both statements can hold simultaneously.
The evidence requirement
Nvidia has argued that customers consistently use four to six year depreciable lives based on observed utilisation and longevity. Companies defending extended schedules must produce utilisation data and failure analysis to auditors, and they have continued to pass those reviews.
Passing an audit is not proof of economic accuracy. It does mean the estimates are supported by evidence rather than asserted, which is a meaningfully different situation from fraud.
Where the defence is weakest
The cascade requires demand at every step. It assumes there is always a lower-value workload waiting for hardware that has aged out of the higher-value one. In a period of aggressive capacity building, that assumption is doing a great deal of unexamined work.
Why cash flow is the better test
Depreciation is a non-cash charge. It moves reported earnings and it does not touch a single dollar of operating cash flow.
Alphabet generated roughly $165 billion in operating cash flow in 2025, and that figure is indifferent to how the accountants scheduled a server. An investor valuing these businesses on cash rather than earnings is largely insulated from the entire argument.
This is where the bear case overreaches. There is a defensible claim that earnings quality is worse than reported. There is a much weaker claim, often made in the same sentence, that the businesses are not working. Only the first survives contact with the filings.
The genuine risk is different and slower. If real economic life is closer to three years, then what is currently classified as growth capital expenditure is actually sustaining capital expenditure. That would mean permanently lower free cash flow rather than a one-off writedown, and it would show up gradually rather than as a shock. The broader cash flow picture is examined in the analysis of 2026 capital expenditure.
Why this matters if you just buy software
Two consequences reach you even though none of this is your accounting.
The first is pricing. If depreciation schedules shorten, the cost base of every cloud and inference provider rises on paper, and providers under margin scrutiny pass that through. The mechanism is slow and it is directional.
The second is capacity. A company that concludes hardware ages faster than assumed has an incentive to slow purchasing, which tightens availability of the newest capacity. That affects who can access frontier models and at what rate limits.
Neither is a reason to act today. Both are reasons to understand why your provider's costs might move in a direction that has nothing to do with your usage.
There is a third consequence that is easy to miss. Depreciation policy shapes how aggressively a provider is willing to build. A company confident that hardware earns for six years commits to capacity a company assuming three years would not, and that difference determines whether there is spare capacity in the market when you need it.
What to watch
| Signal | Why it matters | Where to find it |
|---|---|---|
| Any change to stated useful life | The single most direct evidence. A shortening signals management no longer believes the cascade holds. | Property and equipment note in annual filings |
| Impairment charges on compute assets | A writedown is the schedule breaking in public rather than gradually | Quarterly results, non-recurring items |
| Construction in progress balances | Large balances mean depreciation has not started yet and future charges are still building | Balance sheet detail |
| Divergence between companies | Two firms reaching opposite conclusions under identical conditions is the tell that this is judgement | Comparing filings side by side |
The first row is the leading indicator and almost nobody reads it. It appears in a note most investors skip, and it will move before any revenue number does.
There is a second-order signal worth adding. Watch whether companies begin disclosing useful life separately for accelerators rather than bundling them with servers and networking equipment. Separate disclosure would suggest management expects the two to diverge, which is itself an admission about the pace of obsolescence.
Finally, watch the language. Filings that previously described useful life estimates in a single sentence have started adding paragraphs about utilisation data and secondary deployment. Expanded disclosure usually precedes a change in the number, because companies build the justification before they move it.
None of these signals requires an accounting background to read. They require knowing which note to open, which is the only real barrier to a topic that has been public and audited throughout.
Frequently asked questions
How long do hyperscalers depreciate GPUs for?
Most large cloud providers use five to six year useful lives for AI data centre hardware, though the figure varies by company and by asset category. Nvidia has said customers consistently use four to six year depreciable lives based on observed utilisation and longevity. The trend extended steadily between 2020 and 2024, then diverged in 2025 when Amazon shortened some estimates and Meta extended its own.
What is the $176 billion depreciation claim?
In November 2025, investor Michael Burry argued that hyperscalers were understating depreciation by extending useful life assumptions beyond what a two to three year chip cycle justifies. He estimated the gap at roughly $176 billion of understated depreciation across the industry between 2026 and 2028, with Oracle profits potentially overstated by around 27% and Meta by around 21% by 2028.
Does depreciation policy affect cash flow?
No. Depreciation is a non-cash charge, so changing the schedule alters reported profit without touching operating cash flow. Alphabet generated roughly $165 billion of operating cash flow in 2025 regardless of how servers were scheduled. This is the strongest counter-argument to the overstatement claim, and it is why earnings quality and business quality are separate questions here.
Is extending GPU useful life fraud?
Not on the available evidence. Companies must support useful life estimates with utilisation data and failure analysis, and auditors test that evidence. The industry defence rests on a cascading use model where hardware moves from frontier training to inference to lower-value compute over several years. That is a documented operational pattern, though it assumes demand exists at every step.
Why did Amazon and Meta go in opposite directions?
Because useful life is a management estimate rather than an observable property. In 2025, Amazon shortened the useful life of a subset of servers while Meta extended its estimate further, under the same technological conditions. That divergence is the clearest evidence that these numbers reflect corporate strategy and accounting conservatism rather than a single correct answer.
What should I watch to know if this matters?
The property and equipment note in annual filings, where stated useful lives appear. A company shortening its schedule is telling you management no longer believes hardware generates benefit for as long as previously assumed. That disclosure will move before any revenue figure does, and it appears in a section most investors skip entirely.
Where to start this week
One reading exercise, about twenty minutes.
Open the most recent annual filing of any large cloud provider and find the property and equipment note. Locate the stated useful life for servers and networking equipment. Then find the same note in the filing from three years earlier and compare.
Whichever direction it moved, you now understand more about the economics of this cycle than most of the commentary on it. If it moved down recently, that is the most informative signal available and it was published without a headline.
References
- CNBC, Big Short investor Michael Burry accuses AI hyperscalers of artificially boosting earnings, 11 November 2025. Used for the $176 billion estimate and the Oracle and Meta figures.
- Deep Quarry, Useful lives of GPUs: key considerations. Used for the 2020 to 2024 extension trend, the 2025 divergence and Nvidia's response.
- Real Investment Advice, The AI bear case: what sceptics get right and wrong, 2026. Used for the cash flow counter-argument and Alphabet operating cash flow.
- CNBC, Tech AI spending approaches $700 billion in 2026, February 2026. Used for the capital expenditure context.
The $176 billion figure is one investor's estimate, published alongside disclosed short positions in AI-exposed companies. It is reported here as a contested claim with its counter-arguments, not as an established number.
Related reading