OpenAI and Broadcom's "Jalapeño" Chip: Custom Silicon for Inference

OpenAI's First Custom Chip Targets the Cost of Thinking

On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first piece of custom silicon. It is what the industry calls an Intelligence Processor: an ASIC, an application-specific chip, architected specifically for large language model inference. If you want to understand why this matters, it helps to separate two very different jobs a chip can do.

Training a model is the expensive, one-time act of teaching it. Inference is what happens every single time you actually *use* the model, and it is where the ongoing cost and energy consumption pile up. Jalapeño is built for that second job. General-purpose GPUs are remarkable, but they are generalists. A chip designed around one task can strip away everything it does not need.

Better Performance Per Watt

Early testing shows Jalapeño delivering substantially better performance-per-watt than current state-of-the-art hardware. That phrase, performance-per-watt, is the metric I would watch most closely. It measures how much useful computation you get for each unit of energy, and in an era of ballooning AI demand, efficiency is not a footnote. It is the whole ballgame.

More efficient inference has a pleasant downstream effect: it makes AI cheaper to run at scale, which tends to make it more accessible. When the cost of answering a question drops, more people and more applications can afford to ask.

Nine Months From Design to Tape-Out

The engineering story here is just as striking as the chip itself. Jalapeño went from initial design to tape-out, the moment a chip design is finalized and sent to be manufactured, in roughly nine months. For a large, reticle-sized ASIC, meaning a chip about as big as fabrication tools allow, that is an unusually fast cycle. High-performance silicon of this class often takes considerably longer.

A Milestone in Engineering Velocity

How did they move so quickly? OpenAI used its own models to help accelerate the design work. I find this genuinely fascinating, because it is a small, concrete example of a positive feedback loop: AI helping to build the very hardware that will one day run AI more efficiently. It is the kind of milestone that says as much about *how fast* we can now iterate as it does about the product.

Jalapeño was co-developed with Broadcom, a company with deep custom-chip expertise, which is a sensible pairing. Designing frontier silicon is enormously hard, and combining OpenAI's model know-how with Broadcom's fabrication and design muscle plays to both partners' strengths.

What to Watch Next

Initial deployment is targeted for late 2026. As always, I would treat early performance figures as promising rather than final, real-world results at scale are the true test, but the direction is clear and constructive. Purpose-built inference chips point toward AI systems that are cheaper to operate and lighter on energy, and the remarkable design velocity suggests the field's hardware ambitions are only accelerating.

For anyone who cares about making powerful AI both sustainable and widely available, Jalapeño is a milestone worth cheering.

Sources: OpenAI (June 24, 2026); Tom's Hardware (June 24, 2026).