NVIDIA Blackwell Dominates AI Agent Benchmark

An AI agent asked to evaluate a company for investment doesn't just answer a question. It queries financial databases, pulls news and filings, spins up a sub-agent to run peer comparisons, models valuations, then stitches all of it into a recommendation. Every one of those steps feeds tokens into the next, and OpenRouter data shows that pattern pushes agentic workloads to consume 15 times more tokens than a simple chat request. Software development, customer service, deep research: the same loop of reasoning, tool calls, and sub-agent spawning shows up everywhere agentic AI gets deployed.

That token appetite is now a hardware problem. As agentic systems move from demos into production, the infrastructure underneath them has to keep pace with demand that looks nothing like a chatbot query. NVIDIA says its Vera Rubin NVL72 platform was built with that shift in mind, and new measured results, using SemiAnalysis's AgentX benchmark built from recorded real-world coding sessions, put a number on the gap between generations. The comparison centers on NVIDIA's own GB300 NVL72 systems, and the gains show up specifically in throughput per megawatt, the metric that matters most for power-constrained data centers.