OpenAI and Broadcom Reveal Jalapeño Inference Chip for LLMs

OpenAI and Broadcom Inc. have unveiled Jalapeño, an inference processor designed for large language models, announced in a press release. The chip represents the first generation in a multi-year platform aimed at improving performance and efficiency for AI workloads.

Jalapeño was developed from initial design to manufacturing in nine months. It was created through collaboration between OpenAI’s hardware and software teams and Broadcom’s silicon engineering group. The accelerator is optimized for inference tasks across current and future AI models and is already running machine learning workloads in the lab at production target frequency and power.

According to early testing, Jalapeño delivers significantly better performance per watt than current leading chips. The architecture minimizes data movement and balances compute, memory, and networking resources to achieve near-peak utilization. Broadcom contributed its networking technologies, including Tomahawk silicon, to enable large-scale deployment.

The companies plan to deploy Jalapeño at gigawatt scale in data centers starting in 2026, with further generations under development. Celestica is involved in system integration and production, supporting OpenAI’s goal to expand its infrastructure stack from models and products to custom hardware.