Artificial Analysis' new AA-AgentPerf benchmark, using DeepSeek V4 Pro, found Nvidia's Blackwell GB300 NVL72 at 61,400 concurrent agents per megawatt versus 2,600 for Hopper HGX H200, underscoring data center power-efficiency gains.
Artificial Analysis launched AA-AgentPerf, which it described as the first hardware benchmark for AI agents, using concurrent agents per megawatt to measure how much agent workload a system can support for a given power budget. In initial tests with the 1.6 trillion-parameter open-source model DeepSeek V4 Pro, Nvidia's Blackwell GB300 NVL72 delivered 61,400 concurrent agents per megawatt versus 2,600 for Hopper HGX H200, supporting the claim that Blackwell can provide about 20x more agents per megawatt than Hopper. The results frame Blackwell as a potentially significant advance for data center scalability, power efficiency and operating economics as electricity and capacity constraints become more important for AI infrastructure.