OpenAI Developed Its First Custom AI Chip, Named “Jalapeño”

OpenAI and Broadcom have announced Jalapeño, an AI chip built by OpenAI to optimize its large language models. This is a company that has relied almost exclusively on Nvidia GPUs for its work, and now has its first homemade chip, which has been tailored from the ground up for the specific tasks it needs to perform to power ChatGPT, Codex, and all of its other products.

Jalapeño focuses on inference — when an AI model receives input from a user and produces an output. Imagine training an AI model is something you do once in a while, maybe every few months. But providing service to your users is something you have to do billions of times a day, every second someone loads up ChatGPT or hits your API. You could use Nvidia GPUs to do this, but those are general purpose chips that have to do a lot of different things. If you have a chip that’s specifically tailored to your models, you can do the same thing more quickly, efficiently, and cheaply.

1The real stakes: breaking free from Nvidia

Today, OpenAI has to pay Nvidia rates for its chips, get them on Nvidia’s timeline, and fit within Nvidia’s product plans. During periods of high demand for AI, OpenAI has to wait its turn, just like everyone else. With custom chips, OpenAI gets exactly what it wants, builds it using Broadcom and TSMC’s manufacturing processes, and doesn’t have to pay a markup to a middleman.

This is just one piece of a larger deal between OpenAI and Broadcom, which involves building and installing 10 gigawatts of custom AI chips designed by OpenAI, starting in the second half of 2026 and continuing until 2029. Broadcom has the semiconductor knowhow and network equipment, and OpenAI knows best what its models need from a hardware perspective. Ten gigawatts is an enormous amount of power — the equivalent of a few big nuclear power plants — and highlights just how much infrastructure is required to make modern AI products work.

2What this means for AI costs — and for you

Over the last couple years, the cost of providing AI services has dropped dramatically thanks to software optimization and intense competition. Next, we’ll see a step change in infrastructure efficiency as companies start to deploy custom inference chips, which are optimized for specific AI models. Without having to pay a Nvidia markup, OpenAI can offer its services cheaper to end users when it runs its own hardware.

As a user of ChatGPT or an API developer, you won’t notice Jalapeño — it will be used in data centers far away from you. But one of the reasons why AI costs dropped so much over the last couple years, from expensive and scarce to cheap and ubiquitous, was because of infrastructure efficiencies. The next round of those efficiencies come from custom chips, which we haven’t seen yet but will when they become widely deployed.