OpenAI Builds Its Own Brain: Inside Jalapeño, the Custom Chip Made to Loosen Nvidia's Grip

The single biggest constraint on the artificial intelligence boom is not ideas or data but raw computing power, and for years that power has flowed almost entirely through one company's chips. Now OpenAI, one of the hungriest consumers of that hardware, has decided it can no longer afford to simply rent someone else's silicon, and has built a chip of its own.

Meet Jalapeño

Unveiled on June 24, 2026, Jalapeño is OpenAI's first custom processor, a piece of hardware the company describes as an Intelligence Processor built specifically to run artificial intelligence models. It was developed in partnership with the chip specialist Broadcom, marking OpenAI's most serious move yet into the physical layer of computing.

The division of labor is telling. OpenAI designed the accelerators at the heart of the chip, while Broadcom handled the physical implementation and supplied its Tomahawk networking technology to tie everything together, and the manufacturing itself was carried out by TSMC, the Taiwanese giant that fabricates the world's most advanced chips.

Nine Months From Idea to Silicon

Perhaps the most striking part of the story is not the chip but the speed at which it appeared. OpenAI says Jalapeño went from initial design all the way to manufacturing tape-out in just nine months, which it believes is the fastest development cycle ever achieved for a high-performance advanced semiconductor.

That pace is remarkable in an industry where custom chips typically take years to move from concept to production. With the Broadcom partnership only announced in October 2025, the fact that a finished design reached tape-out so quickly signals just how aggressively OpenAI is willing to move to secure its own supply of computing power.

Built for Inference, Not Training

Crucially, Jalapeño is aimed at inference rather than training, meaning it is designed to run models that have already been built rather than to create them from scratch. Inference is the workload that happens every single time a user sends a prompt, and at OpenAI's scale it represents an enormous and relentless recurring cost.

By owning the chip that handles this work, OpenAI hopes to significantly reduce those operational expenses while gaining far greater control over its own hardware stack, rather than being locked into whatever the broader market can supply and whatever price it happens to charge at any given moment.

The Performance Claims

On raw capability, OpenAI's numbers are bold. The company says its first homegrown inference chip can do up to 1.9 times more AI work per watt than the systems it compared against, a meaningful gain in a business where electricity is one of the largest and fastest-growing costs of all.

Alongside that efficiency, OpenAI reports that Jalapeño delivered between 1.5 and 1.9 times more work per watt and between 1.7 and 3.6 times lower end-to-end latency across three large open models. As with any figures released by the chip's own maker, independent testing will be needed before the claims can be taken as settled.

Loosening Nvidia's Grip

The deeper significance is strategic. The AI industry has been overwhelmingly reliant on a single dominant supplier of accelerators, and by designing its own chip OpenAI is taking a decisive step toward loosening that grip, joining a wider movement among tech giants to bring critical silicon in-house.

Jalapeño is designed for initial deployment by the end of 2026, with its role expanding in the years that follow, which means the true test of the project lies just ahead. If it performs in the real world as promised, it could reshape not only OpenAI's costs but the balance of power across the entire business of building artificial intelligence.