In the same three months of early 2025, two of the companies driving the AI buildout looked at the exact same machines and reached opposite conclusions.
Meta studied the servers and networking gear humming inside its data centers and decided they would last longer than it used to think. It extended their assumed useful life to 5.5 years, a stroke of the pen it expected to reduce its 2025 depreciation expense by about $2.9 billion. Amazon studied the same category of equipment and decided the opposite. It shortened the assumed life of a subset of its servers and networking gear from six years to five, a change it said would cut 2025 operating income by roughly $0.7 billion. Amazon even gave a reason: “an increased pace of technology development, particularly in the area of artificial intelligence and machine learning.”
Same hardware. Same year. Opposite direction. That contradiction is the whole story, because it proves the thing the AI boom would rather you not notice: how long a chip “lasts” is not a fact of physics. It is a choice, made in a footnote, and it flows straight to the bottom line. Investor Michael Burry, the man who shorted the 2008 housing bubble, estimates that this one choice is hiding around $176 billion of overstated profit across Big Tech between 2026 and 2028.
The Dial Nobody Talks About
When a company buys an AI server, it does not subtract the whole price from profit that year. Accounting spreads the cost across the machine’s assumed “useful life,” the number of years it is expected to earn its keep. Choose a longer life and the yearly charge against profit shrinks; choose a shorter one and it grows. The cash itself left the building the day the server arrived. The expense is a schedule the company sets for itself, reviewed by auditors but chosen by management.
That dial is a powerful lever on a hyperscaler’s income statement, since these companies are spending on compute at enormous and still-accelerating scale. Stretch the schedule and this year’s profit swells; every dollar of depreciation pushed into the future is a dollar of reported earnings pulled into the present. Shorten it and profit shrinks. Meta turned the dial one way. Amazon, staring at the same AI-driven obsolescence, turned it the other and said so out loud.
Alphabet’s Wednesday Tell
You do not need Burry’s thesis to see the pressure building. You just need Alphabet’s numbers from this Wednesday.
Revenue rose 24 percent to $119.8 billion, the company’s twelfth straight quarter of double-digit growth, with Google Cloud accelerating 82 percent to $24.8 billion on AI demand. By the headline, it was a blowout: operating income climbed 30 percent and operating margin widened to 34 percent. The reported net income of $112 billion looked surreal, and it was. Most of it came from a $99 billion unrealized gain on Alphabet’s equity stakes, a paper mark that added $77 billion to net income and tells you nothing about the operating business.
Look one line down and the strain shows. Capital expenditures hit $44.9 billion in a single quarter, and free cash flow swung to negative $5.9 billion as the spend on property and equipment outran the cash the business generated. Management told analysts to expect $195 billion to $205 billion of capex for the full year, with more to come in 2027.
Here is the number that matters. Alphabet’s depreciation of property and equipment jumped from $5.0 billion in the year-ago quarter to $7.1 billion. That is 42 percent growth in the cost of aging hardware, against 24 percent growth in revenue. Depreciation is rising nearly twice as fast as sales, and that is with the generous useful-life assumptions already in place. It is the wave arriving in slow motion.
Why the Chips Age in Dog Years
The reason this debate is sharp, rather than academic, is Nvidia’s release cadence. The company has moved to a roughly annual rhythm: Hopper in 2022, Blackwell in 2024, and Rubin arriving in 2026. Each generation is dramatically more efficient per unit of work, and in a data center where power is the dominant running cost, a two-year-old chip can become uneconomic for frontier work long before it physically fails.
That is the crux of Burry’s argument. He contends hyperscalers depreciate Nvidia GPUs over five to six years when the real economic life is closer to two or three, and that the gap understates depreciation and overstates profit by more than $176 billion across Meta, Amazon, Microsoft, Google, and Oracle from 2026 through 2028. The arithmetic is unforgiving. Picture roughly $200 billion of AI infrastructure, about one year of Alphabet’s guided build. Depreciate it over six years and you book about $33 billion of annual expense. Depreciate it over three and you book $67 billion. That $34 billion swing is not a rounding error; it is the difference between a triumphant quarter and a soft one, produced entirely by an assumption. And because each year’s spending stacks a fresh layer on top of the last, the gap compounds.
The Best Argument That Burry Is Wrong
There is a serious rebuttal, and it is not corporate spin. It is the reason Meta’s auditors, and Microsoft’s, and Oracle’s, keep signing off.
A GPU that is obsolete for training the newest model is not scrap. It cascades down a “waterfall” of less demanding work (inference, fine-tuning, batch jobs, internal tools) and can keep earning revenue for years at the back of the fleet, just not at the frontier. If a chip genuinely produces income for six years, depreciating it over six years is not a trick; it is the rule. Under accounting standards, management is required to revise useful lives when new evidence says an asset will serve longer, and companies defend these schedules to auditors with real utilization and failure data. Burry’s case is therefore not “the chips are worthless.” It is narrower and harder to prove: that the blended economic life is shorter than the books assume, and that the cushion is thinning as Nvidia’s cadence speeds up.
That nuance keeps this from being a fraud story, and it makes the honest verdict unsettling rather than damning. Nobody is cooking books. They are making an aggressive, defensible estimate near the top of a spending cycle, and the same evidence that lets them defend it is the evidence Amazon used to move the other way. When the company with the deepest inference business on earth looks at its own fleet, shortens the schedule, and points the finger at AI, the waterfall defense springs a leak.
What Breaks the Spell
An estimate can hold right up until it can’t. If real-world utilization data forces a broad revision toward shorter lives, the adjustment does not arrive gently. Depreciation catches up all at once, and it lands on the income statement as a sudden bite out of the very margins the market is now paying a premium for. See the mechanics of that spending cycle in why 2026 is AI’s capital “valley of death”, the balance-sheet version in the circular financing loop funding the chips, and the capex scale in Meta’s vast AI budget.
For now, the tell to watch is simple, and it is about to get its next reading. When Meta and Microsoft report on July 29, skip past the revenue line and the AI superlatives and go straight to two figures: how fast depreciation is growing against sales, and whether either company has quietly touched its useful-life assumption again. Those footnotes, not the headlines, are where you will find out how much of Big Tech’s profit is real.
🦋 Discussion on Bluesky
Discuss on Bluesky