OpenAI and Broadcom unveiled Jalapeño on Wednesday, the first custom inference chip designed by the company that has spent the last four years as one of Nvidia’s largest customers. Broadcom CEO Hock Tan, speaking to Bloomberg, said the accelerator is running inference at roughly 50% the cost of typical AI graphics processing units.

The framing matters more than the silicon. Since 2022, OpenAI’s economic story has been inseparable from its GPU bill, and Nvidia’s market cap has been inseparable from buyers like OpenAI. Jalapeño doesn’t break that loop, Nvidia still handles pre-training, but it puts a number on the part OpenAI thinks it can take back: the compute-intensive work of serving ChatGPT and other models to users.

OpenAI says the chip went from initial design to manufacturing tape-out in nine months, which the company describes as possibly the fastest ASIC development cycle ever in high-performance advanced semiconductors. Greg Brockman, asked by CNBC’s David Faber how that was achieved, credited OpenAI’s own models: “The degree to which our models have been able to accelerate it was very surprising to us.”

That’s a quietly significant claim. The pitch for frontier AI inside enterprises has always run on testimonials. Here, the testimonial is the chip itself.

For Broadcom, the announcement extends a custom-silicon franchise that has already taken its shares up roughly 10% this year and almost sevenfold since the end of 2022. Celestica handles rack integration. Tan told CNBC the rollout begins with “small prototype development” late in 2026 before scaling, with a broader roadmap built around enabling the deployment of gigawatt scale data centers with Microsoft and other partners beginning in 2026.

The structural read is straightforward. Every hyperscaler-adjacent AI lab now reaches the same conclusion at roughly the same scale: at some level of inference volume, paying Nvidia margins becomes the binding constraint on the business model. Google reached it with TPUs years ago. Amazon reached it with Trainium and Inferentia. OpenAI has now reached it on a nine-month clock, with a co-design partner whose stock chart suggests the market understood the trade before the press release did.

Sources