منذ 1سا
Nvidia's Blackwell Delivers 20x More AI Agents per Megawatt Than Hopper, Benchmark Shows
Nvidia's Blackwell platform is posting a step-change in inference efficiency versus the prior Hopper generation, according to a new benchmark focused on agentic workloads. Results from AgentPerf indicate Blackwell systems can support 20 times more AI agents per megawatt than Hopper—meaning the electricity that previously powered one agent on Hopper-class hardware could now run roughly 20 on Blackwell.
AgentPerf was introduced by Artificial Analysis as AAAgentPerf in March 2026. The benchmark is designed to reflect real-world agent performance, including concurrent users per accelerator and efficiency per rack.
The 20x agents-per-megawatt result is consistent with other published efficiency claims. SemiAnalysis InferenceX reported in February 2026 that Blackwell's GB300 NVL72 configuration can deliver up to 50x higher throughput per megawatt compared with Hopper, alongside a 35x reduction in cost per token for complex workloads such as agentic reasoning.
Several architectural shifts underpin the gain. Blackwell supports FP4 precision, a second-generation Transformer Engine, and updated NVLink designs intended to increase intra-system GPU communication speeds. Power draw per GPU is higher—about 1,200 to 1,400 watts for Blackwell versus roughly 700W for the H100—so the efficiency jump reflects substantially more work per watt, not lower per-chip consumption.
Nvidia CEO Jensen Huang highlighted the trend in late 2025, pointing to inference and agentic AI as key drivers as Blackwell sales outpaced expectations. With data centers increasingly constrained by available power, a 20x improvement in agents per megawatt implies materially more deployment capacity without new power infrastructure, reshaping the unit economics for running AI agents.
For investors, efficiency gains on the order of 20x to 50x in a single generation can expand the total addressable market by making previously uneconomic applications viable. The cited 35x reduction in cost per token improves the feasibility of scaled use cases such as personalized financial advisory tools, real-time supply chain optimization, and autonomous customer service.
Some observers have attempted to link Nvidia's AI infrastructure leadership to AI-associated crypto tokens such as TAO, NEAR, ICP, and RNDR. No direct connection between those tokens and AgentPerf or Blackwell's performance claims has been established.
Competition remains intense as AMD, Intel, and a growing set of custom-silicon startups target the inference market. Still, a 20x edge in agents per megawatt represents a substantial lead in a metric that data-center operators increasingly prioritize.