Brief

AgentPerf posts put AI-chip efficiency claims on an agent-workload footing

Artificial Analysis said in an X post on June 12 that it was releasing first AA-AgentPerf results, describing the benchmark as a new agentic inference test initially covering DeepSeek V4 Pro across Nvidia Blackwell, Hopper and AMD. Nvidia and Nvidia AI Infrastructure posts framed AgentPerf as a way to measure concurrent AI agents under performance targets, with Nvidia AI Infrastructure claiming GB300 NVL72 delivered 20x more concurrent agents per megawatt than Hopper on DeepSeek V4 Pro. That makes “agents per megawatt” a notable public marketing frame for AI infrastructure, but the supplied evidence does not include a readable Artificial Analysis methodology page or result table, so exact rankings, scores, power accounting and comparison rules remain unverified.

Technology · June 12, 2026

Sources & context

Evidence, open claims, and reader-facing editor notes.

Dek

Artificial Analysis’s reported AA-AgentPerf benchmark gives Nvidia a headline efficiency win, but the harder story is whether agent workloads need a new public yardstick for data-center economics.

Synthesis angle

Investigate whether agentic coding benchmarks are changing chip-market incentives by making agents-per-megawatt a competitive metric for AI data centers.

Source

NVIDIA Blackwell Leads First Agentic AI Benchmark

Preferred source plan

Artificial Analysis AA-AgentPerf benchmark page and methodology notes
Artificial Analysis benchmark result tables for Nvidia GB300 NVL72, H200, and AMD MI355X systems
Nvidia technical documentation or blog posts on GB300 NVL72, NVLink, and inference optimizations
AMD Instinct MI355X product documentation and benchmark response if available
DeepSeek model documentation for the workload named in the benchmark
Independent MLPerf or data-center efficiency benchmarks for comparison

Reading the dossier...