OpenAI launches custom inference chip Jalapeño, built with Broadcom

Key Points

• OpenAI introduced Jalapeño, its first custom inference processor, co‑developed with Broadcom.
• The chip is optimized for running pre‑trained AI models, offering higher performance‑per‑watt than current alternatives.
• Early tests suggest significant cost savings for real‑time AI services, especially coding assistants.
• Jalapeño targets inference workloads; pre‑training will likely continue on Nvidia GPUs for now.
• The move aims to reduce OpenAI’s reliance on third‑party GPUs and tighten control over its hardware stack.
• Broadcom handles design and manufacturing, expanding its presence in the AI accelerator market.
• Industry analysts view the chip as a potential game‑changer for AI economics and service pricing.