Nvidia's hold on AI chips just faced its first challenge

OpenAI just showed the first real evidence its own chip can compete with Nvidia's best. At a hardware conference Tuesday, the company ran its new chip, called Jalapeño, through speed tests. It beat Nvidia's current top AI chips, handling more requests per watt and responding faster.

OpenAI didn't build this chip alone. Broadcom, a chipmaker best known for networking gear, manufactures it to OpenAI's specifications. That partnership was announced back in June. Tuesday was the first time outsiders saw real performance numbers, not just promises. The test has a wrinkle worth noting. OpenAI ran it together with SemiAnalysis, the same firm whose benchmark measured the results, inside OpenAI's own lab. That's not the same as Nvidia, or an unrelated third party, independently confirming the numbers.

Keep reading — what to watch

Will Jalapeño actually change how OpenAI runs its business? Not yet. The chip isn't in any real data center. Deployment doesn't start until the end of 2026, and full-scale use isn't expected until 2028. Until then, OpenAI still relies on Nvidia hardware to run its products.