Particle.news
Download on the App Store

OpenAI and Broadcom Unveil Jalapeño, Their First Custom Inference Chip

OpenAI says the chip will cut inference costs by raising performance per watt to support larger-scale deployments.

Overview

  • The companies publicly unveiled the Jalapeño intelligence processor on Wednesday and said engineering samples are running internal workloads in OpenAI labs at target frequency and power.
  • OpenAI and Broadcom say they designed the ASIC in roughly nine months and used OpenAI’s own models to speed parts of the chip design process.
  • Broadcom’s CEO claimed early samples show about 50% cost savings versus typical AI GPUs but the firms have not released independent benchmark data to verify those figures.
  • TSMC is manufacturing the chips and system partners including Celestica and cloud/data‑center partners such as Microsoft are lined up, with small prototyping and initial deployments planned by the end of 2026 and broader scale‑up through 2027–2028.
  • The move is meant to lower OpenAI’s reliance on Nvidia GPUs, cut operating costs for inference, and join a wider industry trend of hyperscalers building custom silicon that depends heavily on high‑bandwidth memory and foundry capacity.