OpenAI built its own AI chip. The target is Nvidia.

OpenAI just showed off Jalapeño, its first home-grown AI chip, built with Broadcom. It is designed for inference, not training. And it is the clearest sign yet that the company most dependent on Nvidia wants a way out.

OpenAI has a chip now. On Wednesday it unveiled Jalapeño, the first piece of silicon it has designed itself. It is a pointed answer to a question that has hung over the company for years. What happens when the world’s biggest buyer of AI compute decides it no longer wants to rent all of it from Nvidia?

The chip was built with Broadcom, Axios first reported. OpenAI did the core design. Broadcom brought the connectivity and networking know-how, plus its Tomahawk switching silicon. A third partner, Celestica, handled the boards and racks.

OpenAI is already running the first samples in its labs. There they answer Codex queries and run workloads for a model it calls GPT-5.3-Codex-Spark.

OpenAI is not trying to replace Nvidia overnight. It is trying to stop being a captive customer. That is the real story here.

What Jalapeño actually is

Jalapeño is built for inference: the day-to-day work of answering user queries. It is not built for training new models. OpenAI calls it an “Intelligence Processor”. It stresses that this is a blank-slate design, not a general-purpose accelerator bent to fit the job. The pitch is efficiency.

Early testing, the company says, shows performance per watt “substantially better than current state-of-the-art”. Thermal behaviour also came in better than expected.

Those are OpenAI’s own numbers, and a full technical report is still months away. For now, the claim to watch is narrow but real. A chip tuned for one job can beat a flexible one at that job. Inference is where AI meets actual users. So even small gains in cost and speed compound fast across hundreds of millions of daily queries.

The plan is to put Jalapeño to work later this year. Broadcom expects the first chips in commercial use at Microsoft and other partners by the end of 2026, though OpenAI says real volume arrives next year. The longer goal is bigger still. OpenAI wants its custom chips powering 10 gigawatts of compute by 2029, roughly the output of ten nuclear reactors.

A nine-month sprint, designed partly by AI

The timeline is the flex. OpenAI and Broadcom say they took Jalapeño from first design to manufacturing tape-out in nine months, which they believe is the fastest such cycle ever for an advanced, high-performance chip. Tape-outs at this level usually take far longer.

Part of how they did it carries a neat twist. OpenAI used its own models to speed up parts of the chip design.

The same systems people query through ChatGPT helped build the hardware that will soon run them. If AI can genuinely help engineers design better chips faster, that lowers the cost of compute for everyone, which is exactly the kind of self-reinforcing loop OpenAI likes to talk about. It also helps explain the recent rush of startups using AI to design chips.

Why build your own chip at all

The motive is control as much as cost. “This gives OpenAI full stack control,” said Richard Ho, who leads the company’s hardware programme. OpenAI now designs the model, the software, the serving systems and, increasingly, the chip underneath. Owning each layer lets it tune the whole thing toward one goal: cheaper, faster intelligence.

Broadcom chief executive Hock Tan put the case more bluntly. “At the end of the day, you cannot, should not rely on some other third-party GPU to do it for you, because it’s such a key part,” he said. The not-so-subtle target is Nvidia, whose chips have powered almost all of OpenAI’s training and inference to date, and whose ecosystem lock-in is precisely what large customers now want to loosen.

OpenAI joins a crowded club

OpenAI is late to a party its biggest rivals threw years ago. Google has its TPUs, Amazon its Trainium and Graviton lines, and Microsoft its Maia accelerators. Each pairs custom silicon with Nvidia chips rather than replacing them outright. Anthropic is exploring its own chips too.

The logic is shared across all of them: at this scale, designing your own silicon is cheaper than paying Nvidia’s margins forever.

One name keeps recurring on the other side of these deals. Broadcom now sits behind a striking share of the industry’s custom accelerators, from Google’s to Jalapeño, and recently struck a huge compute pact with Anthropic and Google.

The company has quietly become the kingmaker of the post-Nvidia chip scramble, supplying the connectivity and manufacturing muscle that the AI labs lack.

OpenAI has already started spreading its bets. Beyond Nvidia, it recently began using Cerebras chips for inference, part of a wider challenge to Nvidia in inference specifically, where rivals see the best chance to break the grip. Jalapeño turns that diversification into something OpenAI owns outright.

The case for caution

A first chip is not a finished strategy. Jalapeño handles inference, not training, and training is where Nvidia’s lead is hardest to challenge. OpenAI concedes Nvidia remains a key partner there. So this is diversification at the edges, not a divorce.

The performance claims also rest on OpenAI’s own early testing, with the detailed report yet to come. Vendor benchmarks at launch deserve a raised eyebrow until independent numbers land. Building chips is slow, capital-hungry and unforgiving work, and a nine-month tape-out is a long way from reliable production at gigawatt scale.

None of that makes the move wrong. It makes it a first step.

The open question is whether OpenAI can really build its way out from under Nvidia, or whether owning a slice of its own silicon simply gives it leverage at the negotiating table. Either way, the company that buys more AI compute than almost anyone has decided it would rather make some of it.

The rest of the industry will be watching how well Jalapeño holds up under real load.