OpenAI Unveils Jalapeño Inference ASIC, Benchmarks Against Nvidia
OpenAI revealed Jalapeño, a custom AI inference ASIC built with Broadcom, at Hot Chips 2026, benchmarking it against Nvidia's GB200 and GB300.
OpenAI has revealed a custom AI inference ASIC called Jalapeño, according to Tom's Hardware. The chip is built for OpenAI's own compute needs, said Richard Ho, OpenAI's Head of Hardware.
Ho said the main design goal for Jalapeño was efficiency. While the chip could be used by others, OpenAI's priority is meeting its own compute demand first.
At Hot Chips 2026, OpenAI shared benchmarks comparing Jalapeño to Nvidia's GB200 and GB300. The benchmarks were run on SemiAnalysis' InferenceX benchmark. Jalapeño accelerated OpenAI's open-weight GPT-OSS model as well as DeepSeek R1 and Kimi K2.5.
Ho said a team got Kimi and DeepSeek running on Jalapeño in the two months between the A0 sample and the Hot Chips presentation. He noted that codesign with OpenAI's internal models is difficult to do with a third-party silicon merchant because of research IP in the models.
Ho said the comparison against Blackwell was because those were the best published results available. OpenAI has run more benchmarks internally against Vera Rubin and for larger context windows. OpenAI and Broadcom revealed their partnership last year.
Quick answers
What is OpenAI's Jalapeño?
Jalapeño is a custom AI inference ASIC revealed by OpenAI, built for OpenAI's compute needs.
What did OpenAI benchmark Jalapeño against?
OpenAI shared benchmarks at Hot Chips 2026 comparing Jalapeño to Nvidia's GB200 and GB300, run on SemiAnalysis' InferenceX benchmark.
Who is OpenAI's partner on Jalapeño?
OpenAI and Broadcom revealed their partnership last year.