OpenAI built its own AI chip and called it Jalapeno

OpenAI and Broadcom have unveiled Jalapeno, OpenAI's first custom-designed AI chip. It's an inference part, the kind that runs models rather than trains them, and the rollout is staged: small prototype data-centre deployments by the end of 2026, a production ramp through 2027, and full scale in the first half of 2028.

The scale being targeted is enormous. OpenAI and Broadcom have committed to deploying OpenAI-designed accelerators at 10 gigawatt scale, working with Microsoft and other partners through 2029. For context, that's the kind of power draw you measure against small countries, not server rooms.

Why build your own chip?

Two reasons, and they both come down to leverage. The first is cost. Nvidia's data-centre GPUs are the most profitable product in tech right now, and every model OpenAI runs pays the Nvidia tax. A custom inference chip tuned for OpenAI's own models can cut that bill. The second is supply. When you're bottlenecked on how many chips you can buy, designing your own gives you a second pipeline that nobody else controls.

OpenAI isn't alone here. Google has run its own TPUs for years, Amazon has Trainium and Inferentia, Meta has its MTIA line. The big AI players are all quietly becoming chip companies, because at this scale the maths only works if you own more of the stack.

The honest caveat

Custom silicon is hard, and first attempts rarely beat Nvidia on raw performance. Jalapeno's value isn't being the fastest chip in the world; it's being a chip OpenAI controls. If it shaves a meaningful slice off inference costs and loosens the Nvidia chokehold even slightly, it's done its job. The timeline also means this changes nothing for anyone in 2026. It's a 2027 and 2028 story.