OpenAI's Jalapeño AI chip claims efficiency lead over Nvidia GB300 in first public benchmarks
qz.com

OpenAI's Jalapeño AI chip claims efficiency lead over Nvidia GB300 in first public benchmarks

Tech News
2 min read

Published by AINave Editorial • Reviewed by Ramit

TL;DROpenAI revealed benchmark results for its custom Jalapeño inference chip, claiming it outperforms Nvidia's GB300 on tokens per watt and throughput per kilowatt. Builders should watch for cost implications but note the timeline and lack of independent verification.

OpenAI has released the first public benchmark results for its custom AI inference chip, Jalapeño, claiming it outperforms Nvidia's GB300 on tokens per watt and throughput per kilowatt. The results, presented at the Hot Chips conference at Stanford, are significant for AI builders because inference cost and power consumption are major constraints at scale. But the benchmarks are OpenAI's own, and deployment is still over a year away.

Jalapeño's performance claims

OpenAI's Jalapeño chip, developed with Broadcom and manufactured by TSMC, is designed specifically for inference workloads. Richard Ho, OpenAI's head of hardware, said the chip delivers greater AI output per watt while also cutting user response times. The benchmarks used SemiAnalysis' InferenceX test and ran models including OpenAI's smaller open-source model, DeepSeek, and Moonshot AI's Kimi model. The largest performance gap was on Kimi, the biggest model tested.

OpenAI normalized results to each accelerator's published TDP. Jalapeño is rated at 700 watts, but OpenAI said measured sustained power stayed at or below 550 watts. In contrast, Nvidia's GB300 is rated at 1,200 watts or more. On throughput per kilowatt, OpenAI reported [up to 1.9x better performance](https://247wallst.com/investing/2026/08/26/openai-says-its-new-chip-outperforms-nvidias-blackwell-as-nvidia-prepares-earnings-release

Sources

Latest Tech News