NVIDIA says power, not FLOPS, sets AI factory profits

NVIDIA is telling the market to stop counting FLOPS and start counting tokens per watt. The company's blog argues that performance-per-watt, not raw compute, decides how much revenue an AI factory can pull out of a fixed power budget. Power is the constraint that doesn't bend, so it's the number that should drive infrastructure decisions.

Worth saying out loud: this is the company that sells the most efficient accelerators on the market making the case for the metric that flatters its own product line. That doesn't make the framing wrong.

Datacenter buildouts right now aren't bottlenecked on chip supply. They're bottlenecked on grid interconnects, the physical process of getting enough power committed and wired to a site. You can buy all the GPUs you want. You can't buy a faster interconnect queue from a utility. That's a real constraint, and it's why performance-per-watt has teeth as a metric even coming from a self-interested source.

Why the floor won't move

This is also why inference pricing has settled where it has. If power is the hard limit on how many tokens a given site can generate, then the cost floor for serving those tokens is set by watts, not by whichever accelerator ships next quarter. A faster chip helps, but it doesn't remove the ceiling, it just lets you hit the same ceiling with fewer racks.

For anyone budgeting AI infrastructure spend, the practical move is to price capacity in tokens-per-watt now, not FLOPS, because that's the number that will still be true after the next hardware refresh. The interconnect queue is the thing to watch. It's a slower-moving, harder-to-fix bottleneck than anything on a spec sheet.