Up to 30x More Work Per Watt: NVIDIA Vera Rubin NVL72 Sets a New Efficiency Standard for AI Agents

NVIDIA reports up to 30-times the throughput per megawatt of GB300 on DeepSeek V4 Pro AgentX and up to 35-times lower token cost. AgentX replays coding sessions with growing context, tools and sub-agents.

The claimed gains combine disaggregated serving, expert parallelism, KV caching, cache-aware routing, fused kernels, NVFP4 and NVLink. NVIDIA says Vera CPU tool-calling performance is not included.

This is NVIDIA’s own blog and explicitly says the Vera results were measured by NVIDIA and await SemiAnalysis review. They are vendor-reported early data, not settled independent benchmarks.

Signal Desk's read: Agent-shaped benchmarking is the right direction, but a 30-times headline needs the trace, power accounting and interactivity target. The omitted CPU work is material because agents spend time outside decode.

The durable contribution is methodological: useful agent work per megawatt is more relevant than fixed-sequence tokens per second.

AgentX is part of SemiAnalysis InferenceX and uses prerecorded Claude Code sessions through AIPerf. The post says GB300 already outperformed Hopper by up to 15-times on one workload.