AI's Profits Are Hiding in a Depreciation Footnote

The chips wear out in three years. The books say six. That gap is where the earnings come from.

By Jake Taylor · 5 min read

Big Tech keeps posting monster AI quarters, and a chunk of that shine comes from a decision buried in a footnote nobody reads: how long they say a computer lasts. Stretch the guessed lifespan of a server and yesterday's expense becomes next decade's problem, and this quarter's profit goes up. Nobody bought anything. Nobody sold anything. They changed a number, and the number changed the earnings.

Quick refresher, because this is the boring part they count on you skipping. When a company buys a forty-thousand-dollar server, it does not expense the whole thing the day it arrives. It spreads the cost across the machine's "useful life." Call that life four years and you write off a quarter of the cost annually. Stretch it to six and you write off a sixth. The gear did not get cheaper. It did not last longer. The estimate moved, and the annual cost dropped, and everything below that line on the income statement got prettier.

This is not a hypothetical. Microsoft extended the useful life of its servers and networking gear from four years to six in 2022. Google did effectively the same thing in 2023. Amazon and Meta stretched theirs too. Each of those changes added billions of dollars to reported operating income in the year it happened, money that materialized out of an assumption and nothing else. Auditors signed off. It was all disclosed. That is exactly what makes it so effective.

The gear didn't last longer. The story about the gear did.

Now hold that against the other thing these same companies are telling you. While they assure investors a server earns its keep for six years, Nvidia is shipping a new flagship architecture roughly every year. A100 in 2020. H100 in 2022. Blackwell in 2024, and another jump already on the roadmap. Every generation makes the last one look slow and power-hungry by comparison. The entire pitch of this industry is that the frontier moves violently fast. You cannot stand on stage and claim the chips are obsolete in eighteen months and age like fine oak in the ledger. Pick one.

Ask anyone who actually runs these clusters. Top-end training GPUs run flat out, hot, around the clock, and they fail. The economic life of an H100 is not six years. It might be three before something twice as fast at half the power makes keeping the old rack plugged in a losing trade. Nobody is going to run 2022 silicon in 2028 out of loyalty. They will rip it out the second the electricity bill beats the resale value. The datacenter knows this. The depreciation schedule pretends it doesn't.

Here is why the trick matters at this particular moment. The hyperscalers are pouring hundreds of billions into AI infrastructure, and depreciation is swelling into one of the largest lines on the whole income statement. When your single biggest cost is governed by a guess, and you keep nudging that guess in the one direction that flatters the quarter, the profit you report becomes partly an editorial decision. It is not a lie. It is a choice about which story to tell, made by the people who benefit most from the flattering version.

And there is a bill at the end of it. If those chips really are cooked in three years instead of six, the back half of every depreciation schedule is a cliff. Two options when you hit it. You keep dead-weight hardware racked and "in service" long past its real usefulness so the books stay clean, or you take a giant write-down and admit a mountain of "assets" quietly turned into space heaters. Neither of those outcomes is anywhere in the price when the market is cheering another record capex number. It gets discovered later, all at once, the way these things always do.

I want to be precise here, because the reflex is to scream fraud, and this isn't that. Useful-life estimates are legal, disclosed, and blessed by every auditor in the building. That is the whole lesson. The most powerful way to move a public company's earnings is not some offshore shell or a shredded document. It's a single sentence in a footnote, revised at the exact moment you need a better number, in full view of everyone, protected by how mind-numbingly dull it is to read.

So the next time a trillion-dollar company posts a blowout AI quarter and the headline screams margins, skip the headline. Go find the property-and-equipment note. Find the useful-life assumption and check whether it moved. That is where the actual story lives, and it is usually the opposite of the one on the front page. The chips wear out right on schedule. The schedule is the fiction.