“For years, OpenAI has been building some of the world’s most talked-about AI models on top of other companies’ hardware. Now it is starting to build a piece of that hardware for itself.”
OpenAI has unveiled its first custom AI chip, a new processor called Jalapeño that was built with Broadcom and designed to handle one of the most expensive parts of running tools like ChatGPT: answering user requests in real time.
The announcement is a big one for OpenAI, but not because it suddenly means the company no longer needs Nvidia. It doesn’t. At least not yet.
What it does mean is that OpenAI is trying to take more control of the machinery underneath its products, at a moment when the cost of serving AI has become almost as important as the models themselves.
Jalapeño is not a general-purpose chip. It was built specifically for inference, the stage where an already-trained AI model responds to prompts, writes code, answers questions or generates text. That matters because inference is where AI companies can burn through vast amounts of money very quickly. Every ChatGPT response, every coding request, every API call has to run somewhere, and that somewhere is usually an expensive cluster of chips.
So OpenAI is trying to change that equation.
The company says Jalapeño was designed around the workloads it knows best, including the systems behind its coding tools and language models. Early testing, according to OpenAI and Broadcom, shows the chip can deliver better performance per watt than current top alternatives, which is another way of saying it may be able to do the same kind of work with less energy and, eventually, lower cost.
That is the business logic behind this move. The strategic logic is even bigger.
For a long time, Nvidia has been the company everyone in AI has had to queue up behind. If you wanted the best chips to train and run advanced models, you were probably buying Nvidia hardware, paying Nvidia prices and dealing with Nvidia’s supply constraints. OpenAI has been one of the companies most exposed to that reality because its computing needs have exploded alongside ChatGPT’s growth. Building its own inference chip does not end that dependence overnight, but it is a clear attempt to reduce it over time.
SEE ALSO: OpenAI Brings in Industry Heavyweights as IPO Preparations Gather Pace
The chip was built with Broadcom, which handled the implementation side of the design, while Celestica contributed system and board integration. OpenAI says the chip went from design to production in about nine months, helped in part by OpenAI’s own models. That is an aggressive timeline by chip industry standards, and it also tells you how urgent this project has become for the company.
There is also something revealing in the name of the chip itself.
OpenAI is calling Jalapeño its first “Intelligence Processor,” language that feels deliberate. The company is trying to frame this as more than a side hardware experiment. It wants the chip to be seen as part of a broader platform strategy, where OpenAI is not just making models and apps, but also shaping the infrastructure underneath them: chips, memory systems, networking, deployment systems, and the rest of the stack that turns an AI model into a usable product.
That full-stack ambition is becoming harder to miss.
OpenAI is already building data center partnerships, agentic products like Codex, and more advanced models that demand enormous amounts of compute. Once you are operating at that scale, buying off-the-shelf hardware forever starts to look less like convenience and more like a vulnerability. Google, Amazon and Meta have all come to the same conclusion in their own ways, which is why each of them has spent years building custom silicon for parts of their AI operations. OpenAI is arriving later to that race, but it is clearly trying to catch up fast.
Still, there are limits to what this announcement actually changes right now.
Jalapeño is focused on inference, not the heaviest training jobs used to build giant frontier models from scratch. Those training workloads still lean heavily on Nvidia’s GPUs, and OpenAI is not pretending otherwise. The company is essentially starting where it thinks it can get the clearest cost and efficiency gains first, then building outward from there. In other words, this is not OpenAI declaring independence from Nvidia. It is OpenAI quietly starting to build an exit ramp.
And maybe that is the most important part of the story.
For all the attention around flashy models and product launches, the AI race is increasingly being decided by more basic questions: who can get enough chips, who can afford to run them, and who controls the infrastructure rather than renting it from somebody else. OpenAI has now made its move on that board.
Jalapeño may not be the chip that changes everything overnight. But it is the clearest sign yet that OpenAI no longer wants to be just a customer in the AI hardware boom. It wants to help build the factory too.
