OpenAI and Broadcom unveiled Jalapeño on June 24, 2026, OpenAI’s first custom-built AI chip. It’s an inference processor, meaning it’s designed to run already-trained models when you send a prompt, not to train them from scratch. OpenAI designed it from the ground up in roughly nine months, with help from its own AI models. Broadcom handled the silicon and networking, while Celestica handles boards and rack systems. Early lab testing reportedly shows performance-per-watt substantially better than current state-of-the-art. Deployment is targeted for late 2026, with a real ramp through 2027 and 2028.
The reason this matters connects directly to OpenAI’s finances. Inference, the cost of serving models to hundreds of millions of users, is one of the biggest line items behind the company’s losses. A chip tuned to run OpenAI’s own models more cheaply per query is a direct attempt to fix that math, query by query. It also reduces dependence on Nvidia’s expensive GPUs, joining Google and Amazon in building custom silicon. The catch: it’s an ASIC, efficient but inflexible, deployment is more than a year out, and no independent benchmarks exist yet.
Best for anyone tracking AI economics, OpenAI’s IPO, or the Nvidia-dependency story. Not ideal for readers wanting a consumer product, because you’ll never hold this chip, you’ll just feel it in faster, cheaper answers eventually.
Last week the leaked numbers said OpenAI lost around $38.5 billion in a year.
The single biggest reason wasn’t training fancy new models. It was the boring, relentless cost of running the ones it already has, every single time someone hits enter.
This week, OpenAI’s answer arrived. It’s a custom chip, and it’s named after a pepper.
What OpenAI Actually Announced
OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom-designed AI processor. Broadcom’s CEO Hock Tan handed a physical sample to Sam Altman and Greg Brockman on Wednesday. That’s the kind of staged moment that signals a company wants you to take its hardware ambitions seriously.
The key word is inference. There are two phases in an AI model’s life. Training is the expensive, one-time process of building the model by chewing through enormous datasets. Inference is everything after that: the model answering your prompt, writing your code, reading your document, responding in a conversation. Training happens once. Inference happens billions of times. Jalapeño is built only for the second one, the part that runs every time a user does anything.
OpenAI designed the chip from scratch around its own models and serving systems. Broadcom contributed the silicon implementation and networking. Celestica handles the boards and rack-level production. It’s an ASIC, an application-specific chip, which means it’s tuned tightly for one job rather than being a flexible general-purpose processor. Engineering samples are already running real workloads in OpenAI’s labs, including a coding model, at production power and frequency. Deployment is targeted for late 2026. Tan told CNBC the real ramp comes in 2027, and “full tilt” in the first half of 2028.
Why a Chip Is the Answer to a Financial Problem
Here’s the part that connects to everything. We covered OpenAI’s $38.5 billion loss last week, and the structural problem underneath it is that costs grow alongside usage. The more people use ChatGPT, the more it costs to answer them. Cost of revenue is the literal expense of serving the models. It nearly tripled in a year. Growth doesn’t relieve that pressure. It adds to it.
Inference is the factory floor of that cost. Every ChatGPT answer, every Codex task, every API call runs on expensive hardware drawing expensive power. So far that hardware has mostly meant Nvidia GPUs. They’re powerful, general-purpose, and priced like the scarcest thing in tech, because they are. When you’re serving models at OpenAI’s scale, even a small cut in the cost-per-query compounds into enormous savings.
That’s what Jalapeño is for. A chip designed for exactly OpenAI’s workloads can hit better performance-per-watt than a general-purpose GPU. Better performance-per-watt means more answers per dollar of electricity. That bends the cost curve behind the $38.5 billion loss in the right direction. The chip isn’t a vanity project. It’s a margin play, and the timing, weeks before a rumored IPO, is not an accident.
The Shot at Nvidia
There’s a second target here, and its name is Nvidia.
Since the AI boom started, OpenAI has been one of the largest buyers of Nvidia’s GPUs on earth. That dependence is a strategic weakness. You don’t control your own costs when the most critical part of your stack comes from one supplier. That supplier sets the price. Building custom silicon is how you claw back leverage. OpenAI is now joining a path Google blazed with its TPU chips and Amazon followed with Trainium. The big AI players are all trying to escape the Nvidia tax.
Brockman framed the logic on OpenAI’s podcast. The company looked for specific workloads that general-purpose hardware underserves, then built something to accelerate exactly those. An ASIC like Jalapeño can be cheaper and more efficient than a GPU for its narrow job, at the cost of flexibility. It won’t replace Nvidia for training, where flexible high-end GPUs still rule. But for the repetitive, high-volume work of serving models, a specialized chip can win.
Broadcom, worth noting, is the quiet winner of this whole era. It helps frontier labs design custom chips. Its stock is up roughly sevenfold since late 2022, and it climbed again on the Jalapeño news. While everyone watches Nvidia, Broadcom has been arming Nvidia’s customers to build alternatives.
The AI-Designed-the-Chip Feedback Loop
One detail in the announcement is genuinely interesting and easy to miss. OpenAI used its own AI models to help design Jalapeño.
Brockman said the degree to which the company’s models accelerated the chip’s development “was very surprising to us.” The whole thing went from initial design to manufacturing tape-out in about nine months. That’s fast for custom silicon, and OpenAI credits its own AI with part of that speed. So you get a loop. AI helps design the chip that runs AI more efficiently, which funds better AI, which designs better chips. Whatever you make of the hype around it, that loop is a real dynamic, and it’s the kind of compounding advantage that’s hard for slower competitors to match.
It’s also a tidy proof point for OpenAI’s pitch. If your models are good enough to meaningfully speed up frontier chip design, that’s a more concrete capability claim than topping another benchmark nobody can name.
What To Actually Take From This
Strip away the codename and the staged handoff, and here’s the honest read.
The strategy is sound. Owning the inference layer is one of the few levers OpenAI has to fix the economics that produced last year’s enormous loss. Custom inference silicon is a real cost advantage at scale. The company is right to chase it. This is what a company trying to grow into its valuation rather than just spend toward it looks like.
But the caveats are real too. Jalapeño doesn’t deploy until late 2026, and doesn’t ramp meaningfully until 2027 and 2028. The cost relief is years out, not quarters. There are no independent benchmarks yet, only OpenAI’s own “substantially better performance-per-watt” claim, which deserves the same skepticism as any vendor benchmark. And an ASIC’s efficiency comes from inflexibility. If OpenAI’s model architectures shift dramatically, a chip tuned for today’s workloads could age awkwardly.
The bigger picture is the clearest part. OpenAI is trying to become a full-stack company, owning the model, the serving software, and now the silicon underneath. That’s the same vertical-integration move Google made and Amazon made. Nvidia is making it from the other direction. Everyone is racing to control the entire stack, because in a business where inference costs compound forever, whoever runs the models cheapest eventually wins.
OpenAI lost $38.5 billion and answered with a pepper. The name is a joke. The math behind it isn’t. Whether Jalapeño actually bends the cost curve won’t be clear until 2027. But the move tells you exactly where OpenAI thinks the war gets won. Not in the demos, but in the boring, brutal economics of running the thing billions of times a day.
