GPU Depreciation: The $18 Billion Accounting Fight Behind AI Earnings

One Number Hides Inside Every AI Profit

A single accounting estimate has quietly decided roughly $18 billion of annual profit across the biggest AI spenders on earth. It is not revenue, not a margin, not headcount. It is how many years a data-center GPU is written off before it disappears from the books.

This is the number behind the loudest fight in tech markets right now — the one that has the man who bet against the housing bubble shorting the AI trade, and CEOs disagreeing with him in public. It is worth understanding, because it decides whether the AI earnings boom is as solid as the headlines make it look.

One machine, two answers

When a company buys a $50,000 GPU server, accounting rules do not let it expense the whole thing in year one. It spreads the cost across the machine's "useful life," recognizing a slice each quarter as depreciation. Stretch that life out, and each year's expense shrinks — reported profit rises, with not one extra dollar of cash. Shorten it, and the opposite happens. This is why the useful-life estimate was always going to become a battleground.

Historically, general-purpose servers were written off over about three years. As AI spending exploded, the hyperscalers pulled that far out. Over a few years, Microsoft, Google, and Oracle moved their AI server estimates out to as long as six years, CoreWeave to six since 2023, and Meta stepped its way from 4.0 years up to 5.5 by early 2025.

Add that up across the group and the movement is enormous. The collective swing from a three-year to a six-year schedule cut the hyperscalers' annual depreciation from roughly $39 billion to about $21 billion — a 46% reduction, or close to $18 billion a year lifting reported earnings. That is the whole game.

Then Amazon did the unexpected: it moved against the crowd. In January 2025 it shortened the useful life for a subset of servers and networking gear from six years back to five, citing the "increased pace of technology development, particularly in the area of artificial intelligence." It swallowed the consequences — roughly $1.6 billion of hits in one quarter.

Amazon's CFO and Microsoft's CEO are on opposite sides of the same machine. Satya Nadella put Microsoft's position plainly: "I didn't want to go get stuck with four or five years of depreciation on one generation."

And Michael Burry, of Big Short fame, thinks even the conservative camp is wrong. He argues the true economic life of Nvidia-based hardware is two to three years, calls the longer schedules aggressive accounting, and estimates it has overstated AI industry profit by roughly $176 billion through 2028. He is short the trade.

Both sides cannot be right. Here is the thing worth noticing: they are each a little bit right, and the reason is the same.

The escape valve: the value cascade

Burry's two-to-three-year case has real force — but it stretches only part of the story. A GPU does not simply die at three years. It cascades down a ladder of jobs.

In its first two years it runs frontier training, where only the newest, most efficient chips qualify. In years three and four it shifts to production inference — answering live queries on models it can still run fine. In years five and six it settles into batch analytics, where latency barely matters and old hardware is cheap enough to be economical.

The evidence says this cascade is real. CoreWeave, which rents out Nvidia gear, re-books GPUs whose contracts expire at roughly 95% of their original price, and reports that A100 chips from 2020 are still fully booked. Google says TPU-class hardware as much as seven or eight years old still runs at "100% utilization." Microsoft's older K80-class servers kept earning for nine years, from 2014 to 2023.

That is the strongest case the hyperscalers can make, and it is a legitimate one: the machine keeps generating revenue long after its frontier days are over. Burry's short side underweights it.

Where the cascade breaks

But a five-to-six-year life is only as solid as the weakest rung of that ladder, and the bottom rungs are getting shaky.

The cascade only works if inference demand grows fast enough to swallow all the retired training chips. Inference is projected to climb from roughly 20% to 80% of AI computing workload by 2030 — a near-16-fold expansion. If growth stalls at eight to ten times instead, modelers estimate 40% to 50% of the old GPUs would have no economical home, and write-downs would follow.

Compounding that, the hyperscalers are building custom inference chips of their own — Amazon's Trainium, Microsoft's Maia, Meta's MTIA. If a purpose-built, cheaper chip outperforms a three-year-old H100 at the inference jobs that are supposed to be the GPU's retirement plan, the bottom of the cascade collapses, and the older Nvidia silicon has nowhere left to go.

The market is already pricing some of this. Rental rates for H100 GPUs fell from roughly $8 to $10 an hour to about $2.85 to $3.50, down around 70% from their 2024 peaks.

Read the cash, not the accounting

Strip the noise away and one practical conclusion survives. Depreciation is a non-cash charge, so a company's free cash flow is completely untouched by whether it writes GPUs off over two years or six. When reported profit and cash flow diverge this much, cash is the more honest mirror of the business.

Use the accounting choice itself as a signal about the person running the numbers. Amazon's reversal — taking a $1.6 billion hit to keep its books closer to reality — is a rare case of a company choosing honesty over a flattering number. Meta's string of yearly extensions leans toward the optimistic end. And for Nvidia, the same annual cadence that makes its newest chips twice as efficient — 25 times more efficient on some inference work — is at once its moat and the source of every customer's write-down anxiety; no hardware maker is more responsible for the speed at which its own products become yesterday's news.

That is what the $18 billion fight is actually about. It is not a scandal and not a fraud — it is an estimate, and honest people can disagree over it. The discipline it demands is straightforward: do not price a company's AI business off its reported earnings. Price it off what the machines actually earn in cash and how the owner talks about how long they will run. That single footnote, more than any revenue print, tells you who is being straight with you.