OpenAI co-designs with Broadcom Jalapeño, its first inference chip

OpenAI and Broadcom co-design Jalapeño, an inference chip tailored for LLMs, aiming for a gigawatt-scale production rollout with Microsoft by late 2026.

OpenAI and Broadcom co-design Jalapeño, OpenAI's first inference accelerator, which OpenAI calls an "Intelligence Processor". The chip was designed not as a general-purpose accelerator repurposed from old workloads, but from scratch around large language model inference, based on what OpenAI says it knows about the needs of its own models, their kernels, their serving systems, and its products. Broadcom provides silicon fabrication and networking technologies, including its Tomahawk chip, and Celestica provides board and rack integration.

OpenAI remains cautious about final performance, which is still being measured, but states that initial tests show significantly higher performance per watt than the current state of the art, with the architecture reducing data movement to approach theoretical peak. Samples are already running workloads in the lab at the targeted frequency and power consumption, including GPT-5.3-Codex-Spark. The two companies claim a nine-month cycle from initial design to manufacturing, which they believe is the fastest ever achieved for an ASIC of this level, partially thanks to the use of OpenAI's models to accelerate design.

Jalapeño opens a computing platform that the partners intend to develop over several generations, with a first production rollout targeted by the end of 2026, at a gigawatt scale. Hock Tan, Broadcom's CEO, mentions data centers built with Microsoft and other partners starting in 2026.