OpenAI Is Building Its Own Chip. 'Jalapeño' Is a Shot Across Nvidia's Bow

The most expensive word in AI right now is “Nvidia.” OpenAI, the company running ChatGPT, pours billions of dollars a year into buying and renting GPUs. Now it’s doing something about it. OpenAI is partnering with Broadcom to design its own inference chip, reportedly codenamed Jalapeño.

A quick note before we dig in. This isn’t a story the community has chewed over yet — there’s been little real discussion on Reddit or Hacker News in the past month. So instead of reaction-mongering, let’s focus on why this move matters and how the pieces fit together.

Why an Inference Chip, Specifically

One concept first. AI chips do two broad jobs. The first is training — the brute-force process of making a model smart from scratch. It demands staggering amounts of compute, and here Nvidia’s top-end GPUs are essentially in a league of their own.

The second is inference — the trained model actually answering your questions. Every time you type something into ChatGPT, that’s inference at work.

OpenAI is going after the second one, and the logic is simple. You train a model once, in big expensive bursts. But inference repeats endlessly, scaling with every user. When you have hundreds of millions of users, inference costs dwarf training costs. The place where OpenAI is really bleeding money is inference.

Shave even a few percent off that recurring bill with custom silicon, and the absolute savings run into the hundreds of millions. That’s why Jalapeño is built for inference.

Why Broadcom

OpenAI isn’t building chip fabs. Designing silicon and manufacturing it are completely different games, and that’s where Broadcom comes in.

Broadcom is a stranger to most consumers, but in the industry it’s a heavyweight in custom silicon (ASICs) — chips built to do exactly what a single customer needs, and nothing else. It already played a deep role in developing Google’s homegrown AI chip, the TPU.

OpenAI knows better than anyone what models it runs and how it runs them. Broadcom knows how to turn those requirements into actual silicon. Together they’re aiming straight at the inefficiency of buying expensive general-purpose GPUs and using only a slice of what they offer.

Nvidia’s GPUs are jack-of-all-trades chips. They do everything well, which means OpenAI pays for capabilities it never touches. A custom chip is the opposite: it packs in only what OpenAI needs and squeezes out maximum efficiency.

The Real Challenge to Nvidia’s Monopoly

So does Nvidia wobble? Not in the short term. Its position in the training market is rock solid. Even with Jalapeño in hand, OpenAI will keep buying Nvidia GPUs in bulk for a long while.

The real shift is somewhere else: leverage.

Until now, AI companies have been the supplicants. Chips were scarce, so they lined up and paid whatever Nvidia asked. But once Nvidia’s biggest customer can say “we have our own chip too,” the conversation changes. The bargaining weight tilts, even slightly, toward the buyer.

And this isn’t just an OpenAI story. Google already runs TPUs, Amazon has Trainium and Inferentia, and Meta is building its own silicon. One by one, the giants are trying to escape Nvidia dependence. Jalapeño is the signal that OpenAI has joined the chorus.

The Mountains Still to Climb

This isn’t all upside. Building your own chip is a money pit. It takes years from design to mass production, and a single failure burns enormous sums.

More importantly, Nvidia’s real weapon isn’t the chip — it’s CUDA, the software ecosystem developers have built on for years. Walking away from that and standing up a fresh software stack for custom hardware is brutally hard. Making the chip is only half the battle.

So the idea of Jalapeño replacing Nvidia is unrealistic. The realistic picture is coexistence: Nvidia for training and the gnarly work, custom silicon for the high-volume, repetitive inference. A division of labor, not a coup.

The Takeaway

Jalapeño isn’t just product news. It’s a sign that the power structure of the AI era is starting to rearrange itself. It’s also the moment OpenAI’s identity widens — from a company that buys chips to one that makes them.

In the end the question is this: will the future of AI belong to the company that owns the model, or the one that owns the silicon running it? OpenAI now wants both. The only thing left to see is whether the bet pays off.