OpenAI has a chip now, and it runs hot — at least in name.
The company revealed its first custom "intelligence processor" on Wednesday, built in partnership with Broadcom. The chip is called Jalapeño and is an ASIC — an application-specific integrated circuit — designed exclusively for AI inference, the step where a deployed model processes a user's request and generates a response. It is not a training chip. OpenAI says Jalapeño will power current and future large language models, including the infrastructure behind ChatGPT and the Codex coding agent.
The move puts OpenAI in a club it has long paid dearly to enter from the outside. Nvidia has collected enormous margins supplying the GPUs that train and run models across the industry, and custom silicon from Google (TPUs) and Amazon (Trainium, Inferentia) has let those hyperscalers cut their own dependency on Nvidia for inference. A dedicated inference ASIC gives OpenAI more control over cost and latency at the exact moment a user is waiting for an answer — which, at ChatGPT's scale, is the moment that matters most.
The partnership with Broadcom arrived fast: OpenAI announced the collaboration only nine months ago. Whether Jalapeño delivers enough performance-per-dollar to meaningfully reduce OpenAI's GPU bill — or whether it ends up as a niche complement to Nvidia hardware — is a question the company has not yet answered with public benchmarks.