OpenAI has revealed its first self-developed AI chip, built in partnership with Broadcom and nicknamed "Jalapeño," according to EdTech Innovation Hub and HotHardware. The chip is designed for LLM inference — the work of actually running large language models to answer user queries — rather than the training of models from scratch.
The move is widely framed as a challenge to Nvidia, whose GPUs currently dominate AI computing. HotHardware describes the launch as OpenAI "bringing the heat" to Nvidia, while 36Kr (which calls it the "Spicy Chip") frames it as a bet that an AI-designed part could rival Nvidia's Blackwell line and bypass Nvidia's main battlefield. Moomoo characterizes it as OpenAI's first self-developed inference chip.
Speed is a recurring theme. Yahoo Finance reports that Broadcom built OpenAI's first chip in record time — though it adds, pointedly, that "the money went elsewhere." 36Kr poses the question of whether AI-designed chips can match Blackwell in as little as nine months.
The financial angle is drawing as much attention as the technology. The Motley Fool flags the development as significant news for both Broadcom and Nvidia investors, reflecting how a major OpenAI customer designing its own silicon could reshape demand. Broadcom's role as the manufacturing partner positions it to benefit even as OpenAI reduces its dependence on off-the-shelf GPUs.
It's worth noting these reports are early and some details — including performance and timelines — remain framed as claims or open questions by the outlets themselves.
Why it matters: if OpenAI can run its own models on custom inference chips, it could loosen Nvidia's grip on the most lucrative part of the AI hardware market and shift billions in spending toward partners like Broadcom.