OpenAI and Broadcom unveil Jalapeño AI accelerator for LLM inference

OpenAI and Broadcom announced the Jalapeño AI accelerator, the first custom‑designed processor built specifically for large‑language‑model (LLM) inference. Developed in partnership with Celestica, the chip was designed and taped‑out in about nine months, marking one of the fastest ASIC development cycles in high‑performance semiconductors. OpenAI described Jalapeño as “its first processor for intelligence,” optimized for the compute, memory, and networking demands of its models such as GPT‑5 and Codex. Early testing shows the chip may surpass current leading products in power‑efficiency, with continued performance validation underway.

The accelerator is intended for deployment in data‑center clusters, leveraging Broadcom’s Tomahawk networking series, and aims for initial installations by the end of 2026. The joint effort is positioned as the first phase of OpenAI’s broader “full‑stack” strategy that combines its models, hardware, and services, and will be offered to partners like Microsoft for large‑scale AI infrastructure.

OpenAI and Broadcom plan to expand the platform over multiple generations, targeting gigawatt‑scale AI data‑center deployments in the future.