OpenAI unveiled Jalapeño, a custom AI inference chip developed with Broadcom, targeting approximately 50% cost reduction compared to current GPU-based approaches for LLM inference workloads.

OpenAI has taped out its first in-house AI chip, Jalapeño, fabricated in partnership with Broadcom and optimized for the transformer workloads behind ChatGPT, Codex, and the API. The move marks the next step in OpenAI's custom silicon development history, reducing reliance on third-party GPUs while leveraging Broadcom's AI ASIC design partnerships with hyperscalers to scale inference capacity for future agentic systems.