Blackwell has fully maxed out energy efficiency this round, and data center electricity costs can finally catch their breath.

The first AI agent hardware benchmark aa-agentperf published by the evaluation organization Artificial Analysis shows that NVIDIA Blackwell is significantly more energy-efficient in AI agent hardware testing. With a power consumption of 1 megawatt, the GB300 NVL72 can handle 61.4k concurrent AI agents, an increase of over 20 times versus the previous-generation Hopper HGX H200, which supports only 2,600. The concurrent capacity of a single graphics card has increased by 41 times. Under the same power budget, data centers can support about 20 times more concurrent AI agents, significantly reducing application deployment costs.