Nvidia's Vera Rubin 30x More Energy-Efficient Than Blackwell

Nvidia just revealed test results showing its next-gen AI platform Vera Rubin is a huge leap in energy efficiency. Running the DeepSeek-v4-PRO 1.6T model, Vera Rubin NVL72 delivers 30 times more tokens per megawatt than the current Blackwell system (GB300 NVL72). That means you get 30 times more AI output for the same power.

The tests used SemiAnalysis's AgentX workload, which checks how well AI agents handle coding tasks, looking at things like response speed and time to first token. The numbers show Blackwell is already 15 times more efficient than the older Hopper system, with cost per million tokens dropping to about one-tenth. Vera Rubin pushes that further, cutting cost per million tokens to just one-35th of Blackwell's.

This efficiency boost lets data centers run more AI agents with the same power and infrastructure, or save a lot on operating costs. Vera Rubin is expected to launch in the next couple of years, keeping Nvidia ahead in the AI computing race.