How Long Does an AI GPU Actually Last? The Quiet Number Holding Up Big Tech Earnings

There is a single number buried in the footnotes of every hyperscaler 10-K (the annual report a public company files with the SEC) that does more work than any revenue line on the front page. It is the estimated useful life of a server. Microsoft says six years. Alphabet says six. Meta says five and a half. Amazon has moved its number twice in two years. CoreWeave says six.^[1,2,3,5]

Change that number by one year and billions of dollars of reported profit appear or disappear across the S&P 500. Nothing about the business changes. No extra chip gets sold, no extra token gets served. Just an assumption in a footnote. That is the whole argument, and it is why I think GPU depreciation is the most underrated variable in the entire AI trade.

Depreciation in plain English

Say you buy a delivery van for $60,000 and you think it will run for six years. You do not book a $60,000 loss the day you buy it. You spent the cash, sure, but the van is going to earn money for six years, so accounting spreads the cost across those six years. That is $10,000 a year of depreciation expense hitting your income statement. Cash out the door on day one, expense recognized in slices.

Now decide the van lasts ten years instead of six. Same van, same cash, same job. But your annual expense drops from $10,000 to $6,000, and your reported profit goes up by $4,000 a year for the first six years. You did not get better at running the business. You changed a guess.

And it is a guess. Useful life is an estimate management makes, audited for reasonableness, not a fact. When a company changes it, GAAP treats it as a "change in accounting estimate," applied going forward with no restatement of history. There is no penalty for having been wrong. There is just a new number.

Here is the part people miss. Stretching the schedule does not create profit. It moves profit forward in time. The van still dies. If it dies in year six and you were depreciating it over ten, you take a writedown in year six for the leftover book value you never expensed. You borrowed earnings from the future and the future eventually asks for them back.

What the hyperscalers actually disclosed

This is not theory. The moves are on the record, with the dollar effects the companies themselves quantified.

Microsoft extended the useful life of server and network equipment from four years to six, effective fiscal 2023. Microsoft's own disclosure: the change increased fiscal 2023 operating income by $3.7 billion and net income by $3.0 billion, or $0.40 per share.^[1,17] Alphabet did nearly the same thing in January 2023, moving servers and certain network equipment to six years, and disclosed a $3.9 billion reduction in depreciation expense and a $3.0 billion increase in net income for 2023, worth $0.24 per share.^[2]

Meta assessed its servers in January 2025 and pushed certain servers and network assets to a 5.5-year life, which it expected to cut full-year 2025 depreciation expense by roughly $2.9 billion.^[3]

Amazon is the interesting one, because Amazon went the other way. It moved servers from five years to six effective January 1, 2024. Then, effective January 1, 2025, it moved a subset of servers and networking equipment back down from six years to five, telling investors the change would reduce 2025 operating income by roughly $0.7 billion. It also took about $920 million of accelerated depreciation in Q4 2024 for equipment it decided to retire early.^[18,19] Through the first nine months of 2025, the shortened life added $889 million of depreciation and cut net income by $677 million.^[19]

The bear case, stated fairly

Michael Burry, the investor best known for shorting subprime mortgages in 2007, has made this his signature argument. His claim, published in late 2025, is that Meta, Amazon, Microsoft, Alphabet, and Oracle are depreciating Nvidia GPUs over five to six years when the real economic life is closer to two or three. His estimate is that the group will understate depreciation by about $176 billion between 2026 and 2028, and that by 2028 Oracle's earnings would be overstated by roughly 27% and Meta's by roughly 21%.^[8,14]He called the practice "one of the more common frauds of the modern era."^[14]

These are Burry's claims, not established facts, and the auditors of all five companies signed off on the current schedules. But the underlying mechanics he is pointing at are real, and there are two separate reasons a GPU might die young.

Reason one: obsolescence.Nvidia ships a new architecture roughly every year now. Rubin, the successor to Blackwell, was launched at CES in January 2026 with the company claiming large multiples of Blackwell's inference performance per watt.^[15] Performance per watt is the whole ballgame in a data center, because power and floor space are the binding constraints, not the chip price. If a new rack does the same work in a third of the power, the old rack does not just become slower. It becomes the wrong thing to have plugged into a socket you could be renting to something better. That is economic death while the hardware is still perfectly functional.

Reason two: physical wear. These parts run hot and they run constantly. Meta published the engineering detail from its Llama 3 training run on a 16,384-GPU H100 cluster: over a 54-day window there were 419 unexpected interruptions, roughly one every three hours, and 58.7% of them traced back to GPU or HBM3 memory failures. Two CPUs failed in the same period.^[12]An H100 pulls around 700 watts. That is a lot of thermal cycling on a part you are telling shareholders will still be earning in 2031. A Princeton CITP analysis cited a Google architect's assessment that GPUs at the 60% to 70% utilization typical of AI workloads survive one to two years, with three as a ceiling.^[10]That is one engineer's view, reported secondhand, and I would not build a thesis on it. But it is not nothing.

The bull case, which is better than the bears admit

The strongest counterargument is not an accounting argument. It is an operational one, and it is simply this: old GPUs keep earning.

Nvidia's response to Burry was that A100s, shipped roughly six years ago, are still running at full utilization.^[13]CoreWeave's CEO made the same case on CNBC, saying the company's A100 fleet was fully booked and that older cards still hold most of their value as new workloads find them.^[7] Reporting on the debate has noted that Google still runs TPUs that are seven and eight years old at full utilization.^[9]

The mechanism is what people in the industry call the cascade. A chip starts life on frontier training, where only the newest silicon is competitive. It gets pushed off that job in eighteen to thirty months. Then it moves to fine-tuning, then to inference (running a trained model to answer queries, which is far less demanding than training one), then to serving smaller models, then to batch jobs, rendering, research clusters, academic access. Each step down is a lower price per hour. But it is not zero, and the asset was already paid for.

There is a second point in the bulls' favor. Inference demand is growing faster than training demand, and inference is exactly the workload old chips are good at. The cascade only works if there is something at the bottom of it. Right now there is a lot at the bottom of it.

Where I actually land

I think both sides are arguing about the wrong thing. The question is not "does an A100 still work in year six." It obviously does. The question is whether the revenue it earns in year six, minus the power and the rack space and the cooling it consumes, is enough to justify carrying it on the balance sheet at the value the schedule implies.

Straight-line depreciation says a GPU delivers the same economic value in year six as in year one. That is almost certainly false. The value curve is front-loaded and steep. A more honest schedule would be accelerated, expensing more in the early years. Nobody does that, because nobody wants to be the first company to report worse earnings than a competitor with identical hardware.

So my honest read: the schedules are probably a little too long, not fraudulently so. Not two-versus-six too long. Maybe five is right where six is claimed, and the shape is wrong even where the length is roughly right. That is not a fraud. It is a soft assumption that everyone has an incentive to keep soft, in an industry where nobody has yet operated a large GPU fleet for six full years and can prove what year six looks like.

The reason this matters more every quarter is the gap between what hyperscalers are spending and what they are expensing. Capex is running far ahead of depreciation, because you depreciate what you bought years ago and they are buying vastly more now than they were then. That gap is not permanent. Depreciation catches up to capex mechanically, with a lag. The bill arrives whether or not the useful-life assumption was right. If the assumption was also too generous, the bill arrives bigger.

How to check it yourself in a 10-K

You do not need a Bloomberg terminal. Pull any hyperscaler's 10-K off EDGAR (the SEC's free filing database) and do four things. It takes about ten minutes.

1. Search the document for "useful lives." It lives in Note 1, Summary of Significant Accounting Policies, under Property and Equipment. You will get a sentence naming the schedule for servers and network equipment. Write the number down.
2. Search for "change in accounting estimate." If the company moved the number recently, it has to tell you the dollar effect on operating income, net income, and earnings per share. This is where Microsoft's $3.7 billion and Alphabet's $3.9 billion live. If a company changed the schedule and its earnings beat came in around the same size as the disclosed benefit, you have learned something.
3. Compare capex to depreciation.Both are on the cash flow statement. Depreciation is an add-back near the top; capex shows up as "purchases of property and equipment" in investing activities. If capex is running at two or three times annual depreciation, you are looking at a company whose expense line has not caught up to its spending, and it will.
4. Compute the implied life.Take gross property and equipment from the PP&E footnote, divide by annual depreciation expense. That gives you a blended average life across everything they own, buildings included. It will not match the server number, buildings drag it up. But track it across years. If it is quietly drifting longer while the asset mix is getting more GPU-heavy, that is a flag worth understanding.

Do that for four companies and you will know more about the real economics of the AI buildout than most of the people arguing about it on television.

The one thing that would settle it

A liquid secondary market for used AI accelerators would end this argument in a week. If there were a public price for a three-year-old H100 the way there is for a three-year-old truck, we would simply look up what the market thinks these things are worth in year three and mark the books to it. Princeton's CITP has pointed out that this market is thin and opaque, which is exactly why the debate stays unresolved.^[11] The absence of a price is not evidence that the price is high.

Until that market exists, the useful-life number is a management estimate that flows straight into earnings, and every company in the sector has the same incentive pointing the same direction. That is not an accusation. It is a structure. And when you find a number that important, that soft, and that incentivized, you do not have to call it fraud to think it deserves a harder look than the market has been giving it.

The AI buildout might well be worth every dollar. I lean toward thinking the infrastructure gets used. But "the demand is real" and "the earnings are correctly stated" are two different claims, and the second one rests on a footnote almost nobody reads.