OpenAI and Broadcom unveil Jalapeño, a custom LLM inference chip

Category: AI
Status: confirmed
Sources: OpenAI, Broadcom press release, discussion
Summary: On 2026-06-24 OpenAI and Broadcom announced Jalapeño, OpenAI's first custom Intelligence Processor, an accelerator designed around OpenAI's LLM inference workloads (kernels, memory movement, networking, and serving patterns). The companies say it went from design to tape-out in about nine months and that engineering samples are running ML workloads in the lab at production target frequency and power, including GPT-5.3-Codex-Spark. They report early testing shows performance per watt substantially better than current state of the art (vendor figure). It is the first accelerator in a multi-generation platform pairing OpenAI-designed silicon with Broadcom implementation and Celestica systems, targeting initial deployment by the end of 2026 at gigawatt-scale data centers with Microsoft and other partners.
Why it matters: OpenAI moving to custom inference silicon reduces dependence on merchant GPUs and signals vertical integration of frontier-model serving.
Follow-up: Track tape-out to production timeline, independent performance-per-watt validation, and deployment partners beyond Microsoft.