OpenAI Unveils Jalapeño AI Inference Chip Developed With Broadcom

OpenAI presented its new Jalapeño chip at the Hot Chips conference, reporting higher throughput and token generation per kilowatt compared to Nvidia Blackwell systems on the InferenceX benchmark.

Developed in collaboration with Broadcom, the chip is designed to minimize data movement bottlenecks during the prefill and communication phases of AI inference processing.

OpenAI plans to begin small-volume deployment of the Jalapeño platform by the end of 2026, with expectations for more significant scaling throughout 2027.