Nvidia’s Blackwell architecture has achieved a massive efficiency milestone, running 20 times more AI agents per megawatt than the previous Hopper generation. The new AgentPerf benchmark confirms that electricity powering a single agent on Hopper hardware now supports twenty agents on Blackwell systems.
This performance leap aligns with broader industry data showing Blackwell’s GB300 NVL72 configuration delivers up to 50x higher throughput per megawatt. Crucially, complex AI workloads like agentic reasoning have seen a 35x reduction in cost per token. These gains stem from architectural advances including FP4 precision and second-generation Transformer Engines rather than reduced power consumption.
For data center operators facing strict power constraints, this efficiency allows for significant scaling without new infrastructure investment. Nvidia CEO Jensen Huang has already cited inference and agentic AI as primary drivers behind Blackwell sales surpassing expectations.
The economic implications are substantial. Dramatic efficiency improvements make previously unviable use cases, such as personalized financial advisors and real-time supply chain optimization, profitable at scale. While competitors like AMD and Intel chase the inference market, Nvidia’s 20x advantage in agents per megawatt establishes a formidable lead in the metric most critical to modern facility operators.