There's one number that quietly decides whether the entire AI buildout actually makes money — and once you see it, you can't unsee it. So let's do the math, with real figures.
Building a one-gigawatt AI data center costs about $38 billion. The chips — the Nvidia accelerators that do the work — are the single biggest slice: roughly $20 billion, over half. The rest is the building, the power, the cooling, the networking.
Here's the catch. The building and the power last 15-20 years. The chips? Effectively obsolete in about three.
$NVDA ships a faster generation on a relentless cadence, and older chips lose 40-60% of their value within 18-24 months of the successor arriving.
So more than half of the most expensive thing humanity is building right now melts in three years. Picture a $38B warehouse where $20B of it is ice sculptures.
That used to be fine — because the rent was insane. In 2023-24, an Nvidia H100 paid for itself in under a year. Easy money.
Then everyone built. Capacity flooded in and the rent collapsed: H100 rental rates are down roughly 64-70% from their peak, to around $2-4 an hour. At those prices the payback period has stretched from under a year to seven-to-ten years — on an asset that's obsolete in three. You can't earn back a three-year chip over ten years of rent. At today's prices, the math just doesn't close.
Don't take our word for it — read the purest example's own filings.
$CRWV does nothing but rent GPUs. Last quarter revenue grew 112% to $2.1 billion. Spectacular, right? Except depreciation alone — the chips wearing out — ate $1.15 billion, more than half of all revenue. The result: a $740 million net loss. Booming demand, losing money on every dollar, because the melting eats more than half the rent.
Then there's the accounting. Most big players write these chips down over five or six years, not three — which makes reported profits look fatter than the cash reality. One prominent short-seller estimates roughly $176 billion of "missing" depreciation across the industry through 2028 — enough to flatter the reported profits at names like Oracle by 20% .
The demand is unquestionable. But the returns, at today's collapsed rents and honest depreciation, are underwater for the pure players and propped up by generous depreciation schedules for the big ones.
This is exactly why our system keeps rejecting the debt-funded builders.
$ORCL sits below our quality floor not because demand is weak — but because borrowing tens of billions to buy a melting asset whose rent is falling is a fragile way to make money.
To be fair to the other side — and we always try to be — this flips if a few things go right: if rental prices stabilize as demand finally outruns the supply glut, if the chips stay useful past three years (older ones still rent), and if utilization stays near-maxed. Any of those, and the math closes again.
But the honest read today: the prettiest demand story in tech is sitting on the fastest-melting asset in tech, and the rent is heading the wrong way. So forget revenue growth — everyone has that. Watch one thing: whether GPU rental prices stop falling. The day they stabilize is the day this becomes a business instead of a race.
Would you borrow billions to buy something half-gone in three years — while the rent keeps dropping?
10q.capital/feed