OpenAI has made a chip. Jalapeño is the first processor the ChatGPT maker has designed itself, built with Broadcom to run the models behind ChatGPT and Codex. Broadcom sent its CEO Hock Tan to hand the wafer to Sam Altman and Greg Brockman in person—and with that, OpenAI joins the small but ever-growing group of AI firms—Google, Amazon, Microsoft and Meta—that have decided buying chips off the shelf isn’t good enough and started making their own.
OpenAI’s Jalapeño is built for AI’s boring, expensive problem
Jalapeño isn’t built to train models. It’s built for inference, the unglamorous job of actually answering your questions once the model already exists. Training gets the headlines; inference gets the bill. Every reply ChatGPT spits out, every task Codex chews through, runs on it—and at OpenAI’s scale that’s a fantastic amount of compute, all day, every day. Build a chip that does only that, and does it well, and the economics start to shift.The pitch isn’t really about speed, though. It’s about electricity. OpenAI says Jalapeño wrings far more performance out of each watt than the chips you can buy today. Power is becoming as much of a constraint as chips themselves. Data centres can only pull so much from the grid. So, a chip that does more with less means more compute from the same power budget. Although, OpenAI hasn’t put numbers to the claim yet, and a full technical report is still months out. Broadcom’s Hock Tan says the chip is as good as Nvidia’s Blackwell, though he didn’t get specific on power.It also came together fast. Nine months from design to manufacturing, which OpenAI reckons might be a record for a chip this complex. To pull it off, the company used its own AI models to help with the design—a slightly dizzying thought, the models helping build the chip that will go on to run them. Samples are already running in OpenAI’s labs on its GPT-5.3-Codex-Spark model, apparently cooler than anyone expected. Broadcom did the silicon and networking, Canada’s Celestica is assembling the servers, and TSMC, as ever, makes the actual thing.
Why bother? Nvidia, mostly
The reason OpenAI—or anyone else—wants its own chip is Nvidia. OpenAI has been one of Nvidia’s biggest customers for years, and Nvidia’s GPUs are both costly and hard to get in the quantities AI labs now need. A homegrown chip means lower bills and less reliance on a single supplier. It’s the same calculation that pushed everyone else down this road.And nearly everyone has gone down it. Google’s TPUs have powered company’s AI efforts for years and are now on their eighth generation is running the Gemini. Amazon’s Trainium chips have sold out so fast the company is weighing whether to sell them to outside buyers. Microsoft has its own inference accelerator, the Maia 200, which it began deploying earlier this year. Meta is building four new MTIA chips with Broadcom too, though they haven’t shipped yet. Even Anthropic is exploring a design of its own, and Broadcom may well be its partner of choice too.Notice the recurring name—Broadcom is in nearly every one of these stories. It’s become the firm you call when you want a custom chip but don’t fancy building a silicon team from scratch, and its shares duly ticked up on the news. Tan doesn’t hedge about it: lead in AI, he says, and you can’t keep borrowing someone else’s GPU.Jalapeño only handles inference for now, though OpenAI hints training chips could follow. Broadcom expects the first ones live at Microsoft and elsewhere by year’s end, with real volume in 2027. The eventual goal is the ambitious bit—OpenAI wants its own chips powering 10 gigawatts of compute by 2029. Whether it gets there is another question. But it’s no longer just renting.