← Back to feed

NVIDIA released the first publicly disclosed on-chip benchmark data for Vera Rubin NVL72 on August 25 — the first time the company has published its own measured numbers for the next-generation cabinet. The test ran DeepSeek-V4-Pro, the 1.6T open-weight MoE model, on the AgentX agentic coding benchmark developed by SemiAnalysis.

The numbers

Against the current flagship GB300 NVL72:

• Throughput per megawatt: Vera Rubin NVL72 is 30x higher than GB300 NVL72.
• Token cost: Vera Rubin NVL72 produces tokens at 35x lower cost than GB300 NVL72.

For context, GB300 NVL72 itself delivers 15x better throughput per megawatt than the H200. On a compounded basis, Vera Rubin reaches approximately 450x the per-megawatt efficiency of the H200 — a number NVIDIA has not stated directly, but which follows from its own published figures.

Why agentic workloads

NVIDIA used AgentX rather than a single-inference benchmark. The company stated explicitly that traditional throughput measurements — a single request, fixed context — no longer represent how frontier AI is deployed. Agent workflows involve growing context windows, tool calls, sub-agent spawning, and multi-turn processing. AgentX replays real coding sessions that include those patterns.

Running DeepSeek-V4-Pro under an agent harness tests Vera Rubin at the workload that actually drives datacenter planning in 2026: not chat inference, but long-running agentic tasks with branching compute requirements.

The choice of DeepSeek-V4-Pro as the test model is also notable. It is an open-weight, 1.6T MoE model that holds the #2 position on SWE-bench Verified at 96.4% — a rigorous coding capability benchmark. Running the strongest open-source agentic coding model on the most powerful GPU cabinet using an agent-specific benchmark creates a credible performance reference.

Context: prior Vera Rubin coverage

This is not the first Vera Rubin benchmark to appear publicly. In June, CoreWeave, the first operator to deploy Vera Rubin NVL72 at rack scale, posted MLPerf v6.0 inference results — an industry-standard single-request throughput benchmark. NVIDIA’s August 25 release is distinct: these are NVIDIA’s own on-chip measurements, not an operator’s MLPerf submission, and they use an agentic workload instead of MLPerf’s request-based methodology.

NVIDIA says the results are from the Vera Rubin NVL72 cabinet, not the standalone GPU. NVL72 packs 72 Vera Rubin GPUs into a single connected system using NVLink, which is what allows the 1.6T MoE model to be loaded and executed efficiently.

What this means for capacity planning

The 30x throughput gain over GB300 per megawatt, at 35x lower token cost, has direct implications for operators deciding between GB300 builds now and Vera Rubin contracts for 2027. If NVIDIA’s figures hold at scale — and the CoreWeave deployment is the first data point that they do for sustained workloads — then Vera Rubin changes the unit economics of running frontier-scale open-weight models by roughly an order of magnitude.

That is also the subtext of deploying DeepSeek-V4-Pro as the test model. A 1.6T MoE model at 35x lower cost per token means the cost to run frontier open-weight models on Vera Rubin is approaching the per-token cost of smaller proprietary models on current hardware.

Key numbers

• Vera Rubin NVL72 vs GB300 NVL72: 30x throughput per megawatt
• Token cost: 35x lower on Vera Rubin vs GB300
• GB300 vs H200: 15x throughput per megawatt (previously reported)
• Test model: DeepSeek-V4-Pro (1.6T MoE, SWE-bench Verified #2 at 96.4%)
• Benchmark: AgentX (SemiAnalysis agentic coding workflow)
• This is NVIDIA’s first publicly released on-chip measurement for Vera Rubin NVL72