Nvidia claims 30x throughput gain on agent coding sessions

Vera Rubin NVL72 reaches 30x the throughput per megawatt of the current GB300 NVL72, per Nvidia's Vera Rubin efficiency post. The company also claims a 35x drop in cost per million tokens. That number matters most for teams sizing agent deployments, where a session's context grows across tool calls and sub-agent spawns before finishing a task.

Nvidia measured both figures on SemiAnalysis's AgentX benchmark, built from recorded agentic coding sessions rather than single-turn prompts. DeepSeek V4 Pro and Qwen3.5 were among the models tested. AgentX scores a full session. It counts context growth, tool calls, and sub-agent spawning, not a one-shot question and answer.

The results are vendor-reported and still waiting on SemiAnalysis's own review. There's no GA date for Vera Rubin NVL72, and Nvidia's post doesn't say when that review will publish.

The benchmark choice signals a shift in how chip vendors market inference economics, from single-turn token counts to full agent sessions. Watch for SemiAnalysis to publish its own AgentX numbers on the same hardware, independent of Nvidia's release, before treating 30x as settled.