OpenAI Unveils Its First Chip, Built With Broadcom

OpenAI has designed its own chip. On Tuesday the company and its manufacturing partner Broadcom pulled the cover off Jalapeno, an inference accelerator that OpenAI is calling its first Intelligence Processor and the opening piece of a hardware platform the two firms plan to build together over several generations.

Inference is the unglamorous half of running an AI model. Training gets the headlines, but inference is what happens every time someone asks ChatGPT a question, and at OpenAI's scale the cost of answering billions of those questions is a business in its own right. Jalapeno is built for that job and little else. The company says early silicon delivers performance per watt well above the best parts available today, though those are OpenAI's own figures and independent testing will tell the fuller story.

Nine months from idea to tape-out

The detail that made engineers sit up was the timeline. OpenAI and Broadcom say they took Jalapeno from initial design to manufacturing tape-out in nine months, which they describe as one of the fastest development cycles ever for a high-performance custom chip. Tape-out is the moment a design is frozen and handed to the foundry, and reaching it in under a year for a part this complex is genuinely quick. It helps to have Broadcom handling the silicon implementation, the networking, and the connective tissue, with Celestica building the boards and racks.

OpenAI frames the work as building the full stack behind its products. The phrase is doing a lot of work. Right now OpenAI rents most of its compute, much of it Nvidia hardware sitting in other companies' data centers, and every token it generates pays a margin to someone else. A chip of its own, tuned to the exact shape of its models, is a way to claw some of that back and to stop being quite so dependent on a single supplier.

A pattern, not a coincidence

Jalapeno did not arrive in a vacuum. The same day, Qualcomm laid out its own plan to challenge Nvidia with a data center roadmap and a software acquisition. Google has run its own TPUs for years, Amazon has Trainium, and Anthropic already spreads its work across all three of those plus Nvidia. One by one, the biggest buyers of AI chips are deciding they would rather design their own.

The first Jalapeno units are due to be deployed by the end of 2026, with later generations to follow. OpenAI has not said it will stop buying from Nvidia, and given how fast its demand is growing, it cannot. But the message to its suppliers is hard to miss: the company that did so much to create the appetite for AI silicon would now like to make some of it too.