Anthropic signed a ten-billion-dollar contract with cloud startup Volta to supply compute for the next six years, with a crypto miner somewhere in the arrangement and a data centre designed from scratch. Impressive number, serious infrastructure.
And the person using the model to summarise a meeting? They don't know what Volta is, or that there's a miner and a purpose-built data centre along the way. They know the answer arrived. Or that it didn't.
The contrast sharpens against what Volta promises: dedicated GPU clusters, low latency, long contracts. All of it for a model that, at the point of use, competes on speed and price. The money is going into the foundation, not into the visible experience. It holds for any operation: what people feel is response time, not the theoretical capacity of the hardware.
The common practice for this kind of cost is to tier the work: simple jobs go to the cheap model and only escalate when the result demands it. Calibrating that routing takes months, and it isn't elegance. It's that the person on the other side can't wait.
The usual way to lose money on this bill is the silent failure: calls landing on the expensive model with no error anywhere in the logs, discovered on the monthly invoice. The system works, nobody is warned, and the price breaks. Which is why a spending ceiling has to live in the code: the vendor dashboard warns, it doesn't cut.
Ten billion in cloud, a six-year contract, half a dozen agents running on a modest server. What reaches the customer isn't the data centre. It's the system staying up.