Particle.news

OpenAI and Broadcom Unveil Jalapeño Inference Processor

OpenAI says the chip is intended to lower per-query compute costs by improving performance per watt for large language model inference.

Overview

  • The companies unveiled the physical Jalapeño sample on Wednesday, June 24, 2026, and said engineering chips are running in OpenAI’s labs at target power and frequency with GPT-5.3‑Codex‑Spark.
  • OpenAI and Broadcom said they co‑developed the ASIC in about nine months, using OpenAI models to speed parts of the design and optimization process.
  • The chip is an application‑specific inference accelerator designed to run LLMs efficiently rather than serve as a general‑purpose GPU, and Broadcom will handle implementation and production with TSMC as the foundry and Celestica building server systems.
  • OpenAI and Broadcom report early gains in performance per watt and position Jalapeño as the first of a multi‑generation platform aimed at initial deployment by the end of 2026 and expansion toward gigawatt‑scale data centers over coming years.
  • Company claims are forward‑looking and currently lack independent benchmarks, so adoption, timeline execution, supply of high‑bandwidth memory, partner offtake and commercial economics remain key risks to watch.