The most honest number in the AI buildout is not a revenue figure or a capex commitment. It is an accounting assumption that one company revised downward while everyone else was revising theirs up. To see why that matters, you have to understand the boring lever that has been quietly inflating Big Tech earnings for three years.
The lever
When a company buys a server, it does not expense the whole cost at once. It spreads the cost across the asset's "useful life"—the number of years management decides the thing will earn its keep. Stretch that estimate and the annual depreciation charge shrinks, which means reported profit grows. No new customer, no new product, no new chip. Just a longer number in a footnote.
Between 2022 and 2025, the largest cloud operators all pulled this lever in the same direction.
| Company | Change | When | Reported effect |
|---|---|---|---|
| Microsoft | Servers/network 4 → 6 years | FY2023 | +$3.7B operating income, +$3.0B net income |
| Alphabet | Servers 4 → 6 yrs; network 5 → 6 yrs | FY2023 | −$3.9B depreciation; +$3.0B net income; +$0.24 EPS |
| Meta | Certain servers/network → 5.5 years | FY2025 | −$2.9B depreciation expense |
These are not estimates from a short-seller. They are the companies' own disclosures, reported through Computer Weekly and The Stack and confirmed in their filings. Each firm offered the same justification: better software, more efficient operation, hardware that simply lasts longer than it used to.
Add it up and the lever moved roughly ten billion dollars of reported profit into existence across three companies, through nothing but a revised opinion about how long a machine works.
The justification has a problem
The "our hardware lasts longer now" story collides with the other story these same companies tell investors every quarter: that AI hardware is improving so fast you must buy the newest generation immediately or fall behind. Nvidia ships a new architecture roughly every year—Hopper, then Blackwell, then Vera Rubin due late this year—each pitched as dramatically better per dollar than the last.
You cannot comfortably hold both beliefs at once. If a 2023 GPU is still economically competitive in 2029, the upgrade urgency is overstated. If the upgrade urgency is real, the six-year useful life is a fiction. The depreciation schedule and the sales pitch are in direct tension, and most of the industry resolved it by simply not mentioning the tension.
You cannot simultaneously argue that last year's chip is obsolete and that this year's chip will still be earning its keep in 2031.
Then Amazon broke ranks
Here is the part that did not get a press release. Amazon raised its server useful life to six years in February 2024, like everyone else. Then, eleven months later, it reversed course. In its Q4 2024 disclosure, Amazon said it had completed a fresh useful-life study and "observed an increased pace of technology development, particularly in the area of artificial intelligence and machine learning." So it cut the assumed life of a subset of servers from six years back to five, effective January 2025.
That reversal is expensive when you are the one telling the truth. Amazon said the change would reduce 2025 operating income by about $700 million. It separately retired some equipment early, booking a roughly $920 million accelerated-depreciation charge in Q4 2024, with another expected $600 million hit to 2025 operating income. Call it $2.2 billion of profit that Amazon chose to recognize—while its peers were busy making profit appear.
The direction is the whole story. Amazon looked at the same AI hardware everyone else owns and concluded it ages faster, not slower. That is a company contradicting the industry consensus against its own short-term earnings interest, which is the rare kind of disclosure worth believing.
Enter the short-seller, and the rebuttal that proves the point
In November 2025, Michael Burry—of The Big Short—put a number on the gap. He estimated the hyperscalers would collectively understate depreciation by $176 billion between 2026 and 2028, and projected that by 2028 the practice would overstate Oracle's earnings by 26.9% and Meta's by 20.8%, calling extended useful lives "one of the more common frauds of the modern era," per CNBC and Fortune. Treat the precision skeptically; Burry is talking his book and the percentages depend on assumptions you cannot audit. The directional claim is harder to dismiss.
Nvidia's defense is the most revealing artifact in the whole exchange. In a note to analysts, it argued that hyperscalers depreciate GPUs over four to six years and that older products "continue to run at high utilization rates and retain meaningful economic value." CFO Colette Kress offered the proof point: A100 chips shipped six years ago are still running at full utilization today.
Read that carefully. "Full utilization" is not "full economic value." A six-year-old A100 may run flat-out and still be a money-loser if the power and rack space it occupies could host a Blackwell doing ten times the work. Utilization measures whether a chip is busy. Depreciation is supposed to measure whether it is worth keeping busy. Nvidia answered the second question with evidence about the first—which is exactly the substitution that lets a depreciation schedule drift away from reality.
What the tell is telling you
The interesting figure here is not Burry's $176 billion. It is the asymmetry: three companies extended useful lives and booked the profit; one company shortened them and ate the cost; and the firm that cut its estimate is the one that runs the largest cloud and presumably knows best how its servers age.
When the accounting and the sales pitch disagree, watch which one a company is willing to pay for. Amazon paid. That is the number to trust.
