OpenAI and Broadcom revealed Jalapeño, an "Intelligence Processor" custom ASIC designed for LLM inference, marking OpenAI's first entry into silicon. The chip was designed in nine months from concept to tape-out using software-hardware co-development, and aims to deliver better performance per watt than current state-of-the-art accelerators. Broadcom will manufacture the silicon while Celestica builds the server systems and racks.
OpenAI plans initial deployment of Jalapeño by the end of 2026, with a roadmap toward gigawatt-scale data center deployments across OpenAI facilities and partner data centers, including Microsoft. The platform will use Broadcom's Tomahawk networking solutions for scale. Early testing at production target frequency and power shows improved efficiency on key LLM workloads including GPT-5.3-Codex-Spark. Broadcom shares rose about 2% on the announcement.
For OpenAI, the move is a step toward building the full AI stack and reducing dependence on Nvidia. Architects shipping production AI systems care because custom inference silicon—optimized for a narrow workload rather than general-purpose GPUs—can lower latency and per-inference costs at scale, assuming Jalapeño's performance per watt claims hold and deployment timelines hold.