OpenAI wants to build more than AI models. Now it wants to build the chips that run them.
On Wednesday, OpenAI and Broadcom unveiled “Jalapeño,” the first custom artificial intelligence chip developed as part of a broader effort to create specialized infrastructure for the next generation of AI systems. The new processor is designed specifically for inference, the stage where AI models generate responses for users after training is complete.
OpenAI says Jalapeño was built from the ground up to handle the demands of large language models, including the systems powering ChatGPT, Codex, API services, and future AI agents.
According to OpenAI, engineering samples of the chip are already running machine learning workloads in testing environments, including the company's GPT-5.3-Codex-Spark model. The company said early testing suggests the processor delivers significantly better performance per watt than current state-of-the-art alternatives.
A push to build the full AI stack
Jalapeño represents more than a new piece of hardware for OpenAI.
The company has increasingly emphasized its ambition to control more layers of the AI ecosystem, from models and products to infrastructure. OpenAI said the chip is part of a long-term strategy to build "the full stack" behind its AI services, allowing tighter integration between software and hardware.
In a statement, OpenAI President and Co-Founder Greg Brockman said, "Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses, and can be used to solve more important problems."
The company argues that designing its own hardware allows it to optimize every layer of the system, from chip architecture and networking to deployment software and user-facing products, around the specific requirements of frontier AI models.
Built in just nine months
One of the most notable aspects of the project is the speed at which it was completed.
OpenAI and Broadcom said Jalapeño progressed from initial design to manufacturing tape-out in approximately nine months, a timeline the companies believe may represent the fastest ASIC development cycle achieved for a high-performance advanced semiconductor.
OpenAI said the same AI systems used by customers today helped engineers improve parts of the chip development process, offering a glimpse into how AI could increasingly contribute to the design of future computing hardware.
Reading between the lines
Strip away the branding, and Jalapeño says something specific about where OpenAI thinks the AI business is actually won or lost: not in training bigger models, but in the unglamorous cost of running them at scale, over and over, for hundreds of millions of users.
Training a model happens once. Inference happens every single time someone uses it. As ChatGPT and Codex usage keep climbing, that per-query cost adds up fast, and it's the part of the business OpenAI has the least control over today.
It's also a hedge against a very specific kind of risk: dependency. OpenAI is preparing for an IPO that could value the company in the trillion-dollar range, and investors backing that kind of valuation will want to see a path to real margins, not just usage growth.
A company that has to buy all its compute from one supplier, at that supplier's prices, has a much harder time controlling its own economics. Building proprietary silicon doesn't eliminate that exposure, but it gives OpenAI a second lever to pull.
What it means for the industry
For OpenAI's customers, businesses building on the API, developers using Codex, and everyday ChatGPT users, the promise is cheaper, faster, more available AI. Lower inference costs are exactly the kind of thing that doesn't show up as a flashy new feature but quietly makes products more responsive and more affordable to run at scale.
For Broadcom, this cements its position as the company hyperscalers call when they want custom silicon without building an in-house chip team from nothing. Broadcom already does this kind of design work for Google, and its stock is up about 10% so far this year and roughly sevenfold since the end of 2022.
According to CNBC, it climbed further following the announcement.
For Nvidia, it's another data point in a trend that's been building for a while. Google has its TPUs, Amazon has Trainium, and now OpenAI has Jalapeño. None of these companies is abandoning Nvidia GPUs outright; demanding work like pre-training will likely continue to rely on Nvidia hardware for the foreseeable future, but the list of customers willing to build alternatives keeps growing.
The caveats
It's worth being clear about what hasn't happened yet.
OpenAI has not published final performance benchmarks; the company says a detailed technical report is still months away. "Engineering samples" running in a lab are not the same as chips operating at scale in a live data center handling real user traffic.
First-generation custom chips have a track record of underdelivering on their early promises; even experienced chipmakers often need a few hardware revisions before a design meets its specs in production.
Timing is also looser than the announcement might suggest. Broadcom's CEO described 2026 as a year for small prototype runs, with meaningful scale not arriving until 2027 and full-tilt production not expected until early 2028. That's a multi-year runway before Jalapeño meaningfully changes OpenAI's cost structure, plenty of time for Nvidia, Google, or AMD to ship their own next-generation answers.
Also read: OpenAI’s Patch the Planet program offers funding and technical support for open-source security projects, including work on software supply chains, memory safety, and critical infrastructure.