OpenAI unveils first custom AI inference chip built by Broadcom, codenamed Jalapeño
Tags AI · Infrastructure
OpenAI announced its first custom-designed processor, codenamed Jalapeño, built in partnership with Broadcom and optimized specifically for AI inference workloads. The chip represents OpenAI's push to reduce dependence on general-purpose GPU clusters and tailor silicon to its model serving infrastructure. Broadcom CEO Hock Tan joined OpenAI CEO Sam Altman at the unveiling. The announcement signals that leading AI labs are following Google's TPU path toward custom inference hardware to improve cost-efficiency and latency at scale.
Technical significance
Custom inference silicon designed specifically for transformer serving patterns could meaningfully reduce per-token costs compared to general-purpose GPUs. This accelerates the trend of AI labs vertically integrating hardware, potentially reshaping the inference market dominated by NVIDIA and creating a new axis of competition around inference efficiency rather than just training scale.