A More Productive Megawatt

NVIDIA just reported that Vera Rubin can deliver up to 30× more agentic AI throughput per megawatt than GB300. Two days later, AWS announced plans to deploy 2 million additional NVIDIA GPUs after demand ran ahead of its earlier expectations.

When useful compute becomes cheaper, we do not simply run yesterday’s workloads with fewer resources and stop. We make new workloads economical. Agents reason longer, use more tools, take on smaller tasks, and get embedded into processes where the economics previously did not work. NVIDIA cites OpenRouter data showing agentic workloads already consume roughly 15× the tokens of a simple chat request.

Jevons described the mechanism more than 150 years ago. Greater efficiency can reduce the resource required for each unit of work while increasing total use by expanding what is worth doing. For AI, that means a more productive megawatt is not necessarily a megawatt we no longer need. It may be a megawatt whose economic value just increased.

Performance per watt matters enormously, particularly where power is constrained. But treating greater efficiency as evidence that infrastructure demand must eventually fall quietly assumes that demand for machine intelligence will saturate.

There is not much evidence that we are close to that point. Efficiency determines how much intelligence a megawatt can produce. Demand determines how many megawatts we build.