The first agentic AI benchmark just dropped. Blackwell runs 20x more agents per megawatt than Hopper.

Your inference benchmarks are lying to you.

Existing AI benchmarks measure one LLM call: how fast an LLM responds to a single request. An agent chains dozens to hundreds of LLM calls together, with tool calls at every handoff. The complexity isn't additive. It's multiplicative.

Artificial Analysis just released AgentPerf — the industry's first benchmark built on real coding agent trajectories. Not synthetic prompts. Real tasks: read files, write code, execute commands, iterate on results. Across 12+ programming languages.

The results changed the math.

NVIDIA GB300 NVL72 runs 20x more concurrent agents per megawatt than HGX H200. At both 20 and 60 tokens/sec per agent SLOs. Using DeepSeek V4 Pro, a frontier MoE model.

This isn't a hardware flex. It's an infrastructure decision.

If you're deploying agents at scale, the question isn't "how fast is one LLM call?" It's "how many agents can I run simultaneously per dollar and per watt?" That's what AgentPerf measures. And the answer just shifted from Hopper to Blackwell.

The benchmark methodology: real coding trajectories, simulated tool calls with representative CPU processing, measured concurrent capacity at defined responsiveness thresholds. Tool call delays are baked in. Growing context is baked in. This is what agentic workloads actually look like in production.

Partners already running this: Together AI powers Cursor's agents on Blackwell. DeepInfra powers Pam.ai's dealership workforce. Baseten serving frontier models.

Vera Rubin is now in full production. The next generation is already here.

Audit your inference stack. If you're still sizing infrastructure off single-call benchmarks, your capacity planning is off by an order of magnitude.

SOURCE: https://blogs.nvidia.com/blog/nvidia-blackwell-agentperf-artificial-analysis/

VERIFIED: NVIDIA Blog (June 12, 2026), Artificial Analysis AgentPerf methodology, NVIDIA developer blog technical deep-dive

SIGNAL: Agentic workloads stress infrastructure differently than chat. First benchmark that measures what actually matters — concurrent agents per megawatt — and the answer changes your hardware procurement timeline.