Free daily briefing on global business news.
The custom inference chip, built with Broadcom, led on tokens per user and throughput per kilowatt in third-party tests
OpenAI shared the first benchmark results for its custom Jalapeño chip on Tuesday, saying the processor outperformed Nvidia $NVDA's GB300 on two key inference metrics: the amount of AI work handled per unit of power and the speed at which it returns responses to users.
The results were presented at the Hot Chips conference at Stanford University. Richard Ho, OpenAI's head of hardware, said on a press call that Jalapeño delivers greater AI output for every watt consumed while also cutting the time users wait for responses. The chip was tested using SemiAnalysis' InferenceX benchmark, according to TechCrunch.
Ho told Bloomberg that Jalapeño's performance at low voltage — 700 watts — should reduce data center power costs. "It's a really good chip — it should drop it by a lot," he said of the expected cost savings. OpenAI ran public tests against one of its own smaller open-source models and external models from DeepSeek and Moonshot AI; the performance gap was most pronounced on Moonshot's Kimi model, which was the biggest of the three.
