⚡NVIDIA GB300 Runs 20x More AI Agents per Megawatt
TL;DR
Artificial Analysis launched AgentPerf, the first agentic-AI infrastructure benchmark, and NVIDIA's GB300 NVL72 topped it, running up to 20x more agents per megawatt than an H200 on DeepSeek V4 Pro. Agentic jobs chain dozens to hundreds of LLM and tool calls, stressing systems differently than chat.
Artificial Analysis launched AgentPerf, the first agentic-AI infrastructure benchmark, and NVIDIA's GB300 NVL72 topped it, running up to 20x more agents per megawatt than an H200 on DeepSeek V4 Pro. Agentic jobs chain dozens to hundreds of LLM and tool calls, stressing systems differently than chat.
Key Points
GB300 NVL72 ran up to 20x more agents per megawatt than an NVIDIA H200 system
Benchmark runs DeepSeek V4 Pro, a frontier MoE model, on real coding-agent trajectories
72 GPUs link into one rack-scale system, tested at 20 and 60 tokens per second per agent
Baseten, DeepInfra, and Together AI already serve agentic workloads on Blackwell
AgentPerf results published June 12, 2026 by Artificial Analysis
Why It Matters
As agents replace single chatbot calls, agents-per-megawatt becomes the number that decides how much an AI infrastructure dollar actually buys.
Quick Facts
Frequently Asked Questions
Why does this matter?
As agents replace single chatbot calls, agents-per-megawatt becomes the number that decides how much an AI infrastructure dollar actually buys.
What happened?
Artificial Analysis launched AgentPerf, the first agentic-AI infrastructure benchmark, and NVIDIA's GB300 NVL72 topped it, running up to 20x more agents per megawatt than an H200 on DeepSeek V4 Pro. Agentic jobs chain dozens to hundreds of LLM and tool calls, stressing systems differently than chat.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 2,085 builders reading daily.