OpenAI and Broadcom detail the Jalapeño inference chip

Category: AI
Status: confirmed
Sources: OpenAI, Broadcom, discussion
Summary: OpenAI's first custom inference accelerator, co-designed with Broadcom around OpenAI's LLM serving patterns, drew the top Hacker News thread of the day (545 points). The companies report a roughly nine-month design-to-tape-out, lab engineering samples running ML workloads including GPT-5.3-Codex-Spark, and a performance-per-watt claim said to beat current state of the art (vendor figure, unverified). Initial deployment targets end of 2026 at gigawatt-scale data centers with Microsoft and other partners.
Why it matters: A frontier lab taping out its own inference silicon is a vertical-integration move that reduces merchant-GPU dependence and pressures the inference accelerator market.
Follow-up: Watch for tape-out to production, independent performance-per-watt validation, and deployment partners beyond Microsoft.