OpenAI's Jalapeño inference chip completes nine-month design cycle with Broadcom
OpenAI unveiled Jalapeño, its first custom inference chip, built with Broadcom in nine months. The chip is in validation and reportedly outperforms competing hardware, but won't be sold or leased externally.
OpenAI has unveiled Jalapeño, its first custom inference chip, developed in partnership with Broadcom. Early testing shows the chip delivers better efficiency than competing hardware, though it remains in validation and has yet to power production servers at scale.
Broadcom has become the foundry partner of choice for frontier AI labs building custom silicon. The firm already manufactures Meta's inference accelerators and Google's TPUs, and is working with Anthropic on its own chip roadmap. OpenAI's timeline was unusually fast — Jalapeño went from initial design to fabricated silicon in nine months.
Internal deployment only
The largest model currently running on OpenAI infrastructure is GPT-5.3 Spark, and Jalapeño has not yet been integrated into full production server racks. The company plans a detailed performance report in the coming months but made clear it has no intention of selling or leasing the chips to third parties. The hardware is being built strictly for internal inference workloads.
Inference-only chips represent a strategic bet that serving costs, not training costs, will define the economics of frontier AI over the next few years. OpenAI is following the playbook Meta and Google have already executed — vertical integration to control the full stack from model weights to silicon.
