OpenAI and Broadcom unveil Jalapeño, a chip built for LLM inference

From models to silicon

OpenAI and Broadcom unveiled Jalapeño on June 24, 2026. The custom accelerator targets large language model inference and begins a multi-generation platform, with initial deployment planned by the end of that year.

OpenAI designed the processor, Broadcom contributes silicon implementation and networking, and Celestica provides board, rack and system-integration expertise.

What is inference?

Inference is the stage in which an already trained model processes new input and produces a response. It is different from training the model itself.

For users, its effects appear in how long an answer takes. For a service operator, another question is how much computing capacity and electricity it takes to serve those answers repeatedly to a large audience.

This helps explain why competition extends beyond building better models to improving the infrastructure that runs them continuously.

Why a purpose-built design?

Our interpretation is that designing hardware around a specific workload creates opportunities to coordinate computation, memory and data movement. A fast arithmetic unit alone cannot solve a bottleneck if it spends its time waiting for information.

The companies say the design reached manufacturing tape-out in nine months. Tape-out is a manufacturing-design milestone, not confirmation that every OpenAI service moved onto the new processor on announcement day.

The distinction between an engineering sample and a deployed system matters too. Running a laboratory workload does not, by itself, establish the cost or reliability of serving a large production service under sustained demand.

What still needs measuring?

The initial performance claims rely on early tests, with a detailed technical report promised later. A final comparison with alternatives therefore needs both results and the methods used to obtain them.

We would judge an inference accelerator through three linked questions: how much useful work does it complete, how long does a user wait, and how much energy does it consume? An advantage in one measure alone may not describe the complete service experience.

The announcement expands OpenAI's involvement from models and products into the computing infrastructure beneath them. Its eventual effect on product speed or the price paid by users depends on deployment results, rather than the announcement alone.

Sources

• OpenAI — 24 June 2026
• Broadcom — 24 June 2026