NVIDIA just reset the bar for AI infrastructure. Its Blackwell GB300 NVL72 platform runs 20 times more agentic AI workloads per megawatt than the previous Hopper generation. This isn’t a minor speed bump—it’s a seismic shift in how AI agents scale.

Agentic AI isn’t your typical chatbot sprint. Instead of a single question and answer, agents juggle dozens to hundreds of model calls, chaining reasoning, tool invocations, and context updates. This makes measuring performance with old benchmarks pointless. Existing tests focus on single LLM calls, missing the complexity of agentic workflows.

Enter AgentPerf, the first benchmark designed solely for agentic AI. It simulates real-world coding agents that read files, write and edit code, execute commands, and iterate. The benchmark measures how many such agents a system can run simultaneously while maintaining strict performance goals. NVIDIA’s Blackwell GB300 NVL72 dominates here, handling up to 61,400 concurrent agents per megawatt—compared to just 2,600 on Hopper.

This leap comes from smart rack-scale design, not just a faster chip. The GB300 NVL72 links 72 GPUs into one system, spreading mixture-of-experts model execution efficiently. Communication and compute overlap within CUDA kernels absorb coordination overhead instead of adding latency. NVIDIA’s TensorRT LLM software further boosts efficiency by decoupling input processing from output generation, scaling smoothly as agent sessions multiply.

The efficiency gains translate directly into cost and capacity advantages. For enterprises, more agents per watt means more productive AI work per data center investment. Power is the real bottleneck as AI deployments grow, not raw silicon speed. This makes Blackwell’s energy efficiency lead a key competitive edge.

Several AI service providers are already running production agentic workloads on Blackwell. Together AI’s Cursor platform uses Blackwell-powered agents to debug code and generate features in real time. DeepInfra’s Pam.ai deploys AI agents on Blackwell to handle calls and sales for car dealerships. These real-world applications prove Blackwell’s performance isn’t just lab hype.

Beyond the hardware, NVIDIA’s software plays a crucial role. The platform supports advanced AI training benchmarks too, winning every test in the latest MLPerf Training v6.0 round. Blackwell Ultra GPUs scale up to 8,192 units in production cloud environments, training massive mixture-of-experts models like DeepSeek-V3 and GPT-OSS-20B. Software innovations such as CUDA graph iterations, CuTe DSL kernel fusions, and optimized communication overlap drive these gains.

Network fabric improvements also matter. NVIDIA’s Spectrum-X Ethernet and Quantum InfiniBand reduce congestion and maximize bandwidth during massive all-to-all GPU communication. These advances enable record-breaking training throughput and time-to-train metrics on models with hundreds of billions of parameters.

Looking ahead, NVIDIA’s upcoming Vera Rubin architecture promises even higher AI capacity. With 50 PFLOPs of compute and a new CPU designed for LLM tool calls, Rubin will extend Blackwell’s lead in agentic AI workloads. The focus on rack-scale integration and full-stack optimization aligns perfectly with the realities of deploying agent fleets at hyperscale.

The AI world has long needed benchmarks that reflect real production challenges. AgentPerf finally does that, exposing the difference between sprints and marathons in AI reasoning. Blackwell isn’t just fast—it’s built for the complexity and scale of today’s agentic AI. Expect competitors to scramble, but for now, NVIDIA owns this race.