OpenAI Built Its Own Chip to Depend Less on Nvidia

OpenAI no longer wants to rent all its computing power from Nvidia. With chip partner Broadcom, it unveiled its first custom AI chip, code-named Jalapeño, built to run AI models and aimed squarely at loosening Nvidia's grip. The plan is to start deploying it by the end of 2026, after a nine-month development push.

The chip is built for inference, the everyday work. Training a model is the one-time heavy lift, but inference, actually running the model for millions of users, is the constant, growing cost. Jalapeño is designed for that job, and a chip tuned for one task can be cheaper and more efficient than a general-purpose Nvidia processor doing the same work. The target is OpenAI's own largest expense.

The strategic point is independence. OpenAI pays enormous sums for Nvidia hardware and competes with every other lab for the same limited supply, so designing its own chip reduces both the cost and the dependence. It also gives OpenAI leverage in negotiations and a hedge if Nvidia capacity stays scarce. Owning the silicon is owning your own destiny in AI.

It also pulls OpenAI deeper into hardware alongside everyone else. Google, Amazon and Meta already design their own AI chips, so OpenAI joining them confirms that the biggest AI buyers all want to escape paying the Nvidia premium. Broadcom, which helps design and build these custom chips, is quietly becoming one of the most important companies in AI by serving exactly this demand.

The honest caveat is that this does not dethrone Nvidia. A custom inference chip handles a slice of the workload, while Nvidia still dominates training and the broader market, and building competitive silicon is hard and slow. Jalapeño is not shipping until late 2026, and first chips often underwhelm. This is OpenAI reducing its exposure, not replacing its main supplier.

So OpenAI is moving from pure software into the silicon that runs it, chasing lower costs and less dependence on one vendor. A custom inference chip, a Broadcom partnership, and a late-2026 target. The AI giants increasingly want to own their whole stack. Watch whether Jalapeño ships on time and how much of OpenAI's compute it actually carries.