OpenAI has done something it has never done before: put its name on a piece of silicon. On Wednesday it unveiled Jalapeño, its first custom AI chip, built in partnership with Broadcom — a move that says as much about the economics of running AI as it does about the technology.

What it actually is

Jalapeño is an inference chip. That word matters. AI has two expensive phases: training (teaching a model, a one-off mammoth compute job) and inference (actually running the finished model to answer your questions, which happens billions of times a day, forever). Jalapeño is built for the second — it's the chip that would serve ChatGPT's answers, not the one that trains the next model.

OpenAI says it designed the chip itself, "from scratch," around its own understanding of how large language models behave, with Broadcom handling the silicon implementation and manufacturing and a third firm, Celestica, building the boards and racks around it. So the honest description isn't "OpenAI built a chip" so much as "OpenAI designed a chip and Broadcom is building it" — a custom accelerator, made to order.

The headline boast is the speed of development: from initial design to manufacturing "tape-out" in just nine months, which OpenAI and Broadcom claim may be the fastest such cycle ever for a high-performance chip — helped, it says, by using its own AI models in the design process. It also claims "performance per watt substantially better than current state-of-the-art." Both of those are OpenAI's own figures, unverified by outsiders, so file them as a confident pitch rather than an established fact. The chip isn't in service yet either: initial deployment is slated for the end of 2026.

Why a software company is designing hardware

The interesting question isn't the chip — it's the motive. Why does OpenAI, a software lab, want its own silicon?

Two reasons, both about control. The first is cost and supply. Nearly every AI company runs on Nvidia's chips, which are expensive and perennially in short supply. A custom chip tuned narrowly for OpenAI's own models, doing only inference, can in principle be cheaper and more efficient per query than a general-purpose Nvidia GPU — and when you serve as many queries as OpenAI does, even a small per-query saving compounds enormously.

The second is independence. Relying on a single supplier for the thing your entire business depends on is a strategic weakness, and OpenAI has talked openly about wanting to "build the full stack" — models, software and now the hardware underneath. Jalapeño is a step toward not being entirely at Nvidia's mercy on price and availability.

The Nvidia question — with a caveat

Plenty of the coverage frames this as "a strike at Nvidia," and there's truth in it: every chip OpenAI runs on its own silicon is a chip it doesn't buy from Nvidia. But it's worth keeping the scale honest. This is one inference chip, not yet deployed, that complements rather than replaces the enormous fleet of Nvidia hardware OpenAI still relies on — including for training, which Jalapeño doesn't touch. OpenAI is also far from alone here: Google has run its own TPUs for years, Amazon has its Trainium and Inferentia chips, and Meta is building its own too. Custom inference silicon is becoming table stakes for anyone operating at this scale, not a surprise raid.

The takeaway

Jalapeño is a real and significant move — a sign that the biggest AI companies increasingly see designing their own chips as essential to controlling cost and supply, not a side project. But read the announcement for what it is: a first-generation, inference-only, not-yet-shipping chip with performance claims that only OpenAI has seen so far. The direction of travel is unmistakable — AI's giants want to own the whole stack down to the silicon. Whether Jalapeño delivers on its bold numbers is something only the end of 2026, and someone other than OpenAI, will be able to confirm.