Depreciation Is an Opinion. The AI Boom Runs on It.

There's a number underneath every hyperscaler earnings beat this year that nobody can verify, because it isn't a measurement. It's a guess. How long does a GPU stay useful? Pick six years and your margins look great. Pick three and a chunk of this year's profit evaporates. Same hardware, same revenue, same electric bill — the only thing that changed is an assumption, and the assumption is worth tens of billions of dollars.

That's the part people miss when they read the income statement like it's a measurement. Depreciation is the one big line on it that isn't. It's an opinion about the future wearing the costume of a fact.

Start from what the number is supposed to do. Depreciation spreads the cost of an expensive asset across the years it earns its keep. Buy a $40,000 server that works for four years, expense $10,000 a year. The logic is clean: match the cost to the period it actually produces value. The whole thing hinges on one input — useful life — and useful life is, by the accounting rules' own definition, a management estimate. Not a fact about the hardware. An entity-specific judgment call about how the company intends to use it. GAAP says so explicitly.

So watch what the big cloud providers did with that judgment call. Over the last couple of years Amazon, Google, Microsoft, and Oracle all stretched the assumed life of their server fleets from the old three-to-four-year standard out to six. Quietly, in footnotes. Each extension does the same mechanical thing: it shrinks this year's depreciation expense, which lifts this year's reported profit, without anyone selling a single additional unit of anything. You can manufacture earnings growth by editing a number in a model. They did.

The melting ice cube

Here's where it stops being abstract. The asset they're stretching to six years is an Nvidia GPU on a roughly three-year cadence, where every new generation lands two to three times more efficient per watt than the last. That efficiency curve isn't cosmetic. It's the whole game. Power is the binding constraint on this entire buildout — I've made that case before — and when a constraint is that tight, you do not keep last generation's chips drawing premium electricity to do frontier work. You can't afford to. The old silicon doesn't fail. It just becomes uneconomic, which for accounting purposes is the same death, only quieter.

That's the contradiction sitting in plain sight. You can't simultaneously tell the market the new chips are a generational leap and book the old ones as if they'll earn at full value for six years. One of those stories is load-bearing for the stock price and the other is load-bearing for the margin, and they point in opposite directions. A melting ice cube doesn't get a longer useful life because you wrote one in the footnotes.

Michael Burry put a number on the gap and aimed it at the room. His estimate: roughly $176 billion of understated depreciation — and therefore overstated profit — across the industry between 2026 and 2028, with specific names taking the worst of it, profits overstated on the order of a quarter by the back end. He called extending useful lives "one of the more common frauds of the modern era." That's the heavy word and I'd hold it at arm's length, because there's a real counterargument: a chip that's no longer state of the art still serves inference for years, cascading down to cheaper, less demanding work. The waterfall is real. Six years of some revenue isn't crazy.

But notice that the counterargument concedes the actual point. Even the bull case admits the chip stops doing the high-value job long before year six. The debate isn't whether the curve bends down early. It's how steep. And when reasonable people are arguing over a slope worth $176 billion, you are no longer looking at a fact. You're looking at a dial, and the people who own the dial also own the earnings it produces.

Where the bill lands

Strip it to first principles and the mechanism is almost dumb in its simplicity. Stretching useful life doesn't make a cost disappear. It moves it. Every dollar of depreciation you don't book this year is a dollar you book later — on top of the gear you're buying later, at a capex run rate that's still climbing. The optimistic assumption isn't free. It's a loan against a future quarter, and the future quarter doesn't get a vote.

So the discipline is the same one that applies to any number a counterparty hands you with a confident face: ask what's an estimate and who benefits from the estimate landing where it did. On these statements, the most important line is the softest, and it's soft in the exact direction that helps the people who set it. That's not an accusation. It's just where the incentive points, and incentives don't need permission to work.

The capex is real. The revenue is real. The margin is the part somebody chose.

A GPU has a useful life. So does the assumption you booked it on — and that one expires on a schedule too.

— Dustin