OpenAI and Broadcom announce Jalapeño ASIC chip designed for LLM inference at scale
Tags AI · Infrastructure · Enterprise
OpenAI and Broadcom unveiled Jalapeño, a custom ASIC designed from scratch for large language model inference in data centers, developed over nine months based on OpenAI's insights about its future model roadmap. The chip is designed to deliver substantially better performance per watt than current state-of-the-art solutions, though detailed benchmarks are pending. Deployment in data centers is expected by end of 2026. This is the first in a planned long-term series of custom chips as OpenAI moves to own its full hardware-software stack.
Technical significance
OpenAI's vertical integration into custom silicon follows Google's TPU and AWS's Trainium/Inferentia path, signaling that frontier AI labs now view proprietary inference hardware as a competitive moat. The nine-month development cycle is remarkably fast for an ASIC, suggesting Broadcom's design expertise compressed the timeline. If Jalapeño delivers on its performance-per-watt claims, it could reduce OpenAI's inference costs and decrease dependence on Nvidia's GPU roadmap.