OpenAI Built Its First Custom Chip — Named Jalapeño

OpenAI and Broadcom have unveiled Jalapeño, OpenAI's first custom chip designed specifically for running large AI models. This is a company that for years has been almost entirely dependent on Nvidia's hardware — and for the first time, it now has silicon of its own, co-engineered from the ground up for the exact workloads that power ChatGPT, Codex, and every product OpenAI ships.

The chip targets what's called inference — the moment when an AI model reads your question and generates a response. Think of it this way: training an AI model happens occasionally, maybe once every several months. But responding to user requests happens billions of times a day, every second someone opens ChatGPT or hits an API endpoint. Nvidia's graphics chips can handle this work, but they're general-purpose hardware built for many things. A chip designed specifically for one company's models does the same job faster, more efficiently, and at a lower cost per request.

The real stakes: breaking free from Nvidia

Every chip OpenAI uses today is purchased at Nvidia's prices, allocated on Nvidia's schedule, and constrained by Nvidia's product roadmap. When AI demand spikes, OpenAI waits in line like every other lab and startup. Custom silicon changes that equation: OpenAI designs exactly the chip it needs, manufactures it through Broadcom and TSMC's supply chains, and operates it without a margin going to a third-party vendor.

Jalapeño is part of a broader collaboration between the two companies: a joint plan to build and deploy 10 gigawatts of OpenAI-designed AI accelerators, with deployment targeted to start in the second half of 2026 and run through 2029. Broadcom brings semiconductor expertise and networking infrastructure; OpenAI brings deep knowledge of what its models actually demand at the hardware level. Ten gigawatts is a staggering figure — roughly equivalent to the entire power output of several large nuclear plants — and underscores how massive the infrastructure buildout behind today's AI products actually is.

What this means for AI costs — and for you

The price of running AI services has been falling sharply over the past two years, driven by software improvements and fierce competition between providers. Custom inference chips — optimized specifically for one company's models — represent the next wave of structural cost reduction. When OpenAI operates on hardware it designed itself, there's no Nvidia premium to absorb and pass along to users.

If you use ChatGPT or build with OpenAI's API, Jalapeño is invisible: it runs in data centers you never visit. But the AI price trajectory of the past two years — from expensive and capacity-constrained to cheap and broadly available — was built partly on efficiency improvements in infrastructure. Custom silicon is the infrastructure improvement that hasn't fully landed yet. When it scales up, it'll show up as lower prices, faster responses, or both.