NVIDIA tops first agentic-coding benchmark — built by a vendor it works closely with

incrementalDeveloperCompute

Monday, June 15, 2026

Conviction

Medium

Time horizon

This quarter

Risk

Procurement decisions anchored to a single-workload benchmark

The News

Artificial Analysis's AgentPerf, billed as the industry's first agentic AI benchmark, published its first round of results showing the NVIDIA Blackwell Ultra NVL72 platform delivering leading performance across the tested workloads, running 20x more agents per megawatt than NVIDIA Hopper. The benchmark replays real coding-agent trajectories and measures how many concurrent agents a system can serve while meeting production service-level targets, with Agents per Megawatt as its lead metric. First results for DeepSeek V4 Pro at the easiest defined service level were 61,354 Agents/MW on GB300 rack-scale, 21,053 on B300 single node, 3,551 on AMD MI355X, and 2,594 on H200. Sources: NVIDIA blog, Artificial Analysis, NVIDIA developer blog.

The Read

Agents-per-megawatt is the right metric — the conversation has finally moved from tokens-per-second to throughput-per-watt, which is what actually constrains a 2026 deployment. The verification gap is harder to close: Artificial Analysis itself notes the AMD MI355X configurations are roughly two weeks older than the Blackwell configs and couldn't stably use speculative decoding, which is a meaningful asterisk on a 17x lead. Treat the 20x-over-Hopper number as directionally correct — it is a real generational step — and the AMD comparison as preliminary until a controlled rerun lands. The bigger takeaway is that the benchmark exists at all: enterprise procurement now has a public reference point for "how many concurrent agents per watt of power I draw," and that question is going to dominate the next round of data-center capacity planning.

Watch For

A second independent AgentPerf submission from a non-NVIDIA-sponsored lab, or an AMD rerun with parity-vintage configs, by end of Q3. If the gap holds within 2x once methodology is normalized, Blackwell is genuinely the agentic-inference floor and AMD has a real generational problem. If it compresses to under 1.5x, the AgentPerf headline was a launch artifact.

For Developer