Noesion Radarnoesion.ai
Research2026-08-25

OpenAI presented Jalapeño, an inference chip it designed with Broadcom, at the Hot Chips conference, and allowed SemiAnalysis to benchmark it in its labs. On the tests run, the chip produced more tokens per megawatt of power than Nvidia, AMD and Google parts, and roughly matched Nvidia's Vera Rubin on cost per token. All figures were supplied by OpenAI, cover only simple single-turn workloads, and the chip exists so far only as engineering samples.

What changed

OpenAI ran its models entirely on chips bought from Nvidia, AMD and Google.

  • 13.4 PFLOPs MXFP4 (B0 stepping)
  • 700W TDP vs Rubin 900-1,150W
  • 15.4TB/s memory bandwidth
  • ~700 tok/sec/user on Kimi K2.5

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.