OpenAI unveils Jalapeño, its first custom chip, built with Broadcom
OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom silicon — a reticle-sized inference ASIC purpose-built for running large language models rather than training them. The chip went from initial design to manufacturing tape-out in about nine months, which the companies call one of the fastest advanced-ASIC development cycles ever, helped along by using OpenAI’s own models in the design process. Engineering samples are already running workloads in the lab, including a GPT-5.3 Codex variant, with performance per watt said to be substantially better than current state-of-the-art. It is the first step in a multi-generation platform — OpenAI accelerators plus Broadcom silicon and Celestica systems — slated for initial deployment by the end of 2026.
Why it matters: OpenAI is moving to build its own full stack rather than rent all of it: designing the chip its models run on is the clearest sign yet that frontier labs see custom inference silicon as the path to cheaper, more controllable compute — and another front in the contest with Nvidia.