OpenAI and Broadcom unveil Jalapeño, a custom inference ASIC co-developed in a 9-month sprint. Samples in hand; claimed ~50% cost savings vs. traditional GPUs on inference workloads.

OpenAI and Broadcom have unveiled Jalapeño, a custom inference ASIC developed through a nine-month collaboration. OpenAI has received first samples of the chip and is currently testing its ability to run AI tasks.

Broadcom CEO Chen Fuyang stated that the custom chip delivers approximately 50% cost savings compared to traditional AI GPUs on inference workloads. The co-development represents OpenAI's effort to secure competitive advantage through optimized hardware for its inference operations.

The move underscores an emerging trend in AI infrastructure where large operators pursue custom silicon to reduce inference costs and operational expenses. With samples now in hand and under evaluation, the chip's path to production deployment and its potential impact on OpenAI's cost structure remain to be demonstrated.