OpenAI has published the first benchmarks for its own AI chip, and the numbers are pointed squarely at Nvidia.
According to Tom's Hardware, OpenAI presented the results at the Hot Chips conference on Tuesday, claiming its in-house accelerator — codenamed Jalapeño and co-developed with Broadcom — beats Nvidia's GB300. The headline figures: up to 1.9x more throughput per kilowatt and 3.6x lower latency. Jalapeño is a 700W part; the Nvidia flagship it is measured against draws 1,400W.
Anadolu Ajansı reported the same claim in broader terms, saying OpenAI's first custom AI chip surpasses Nvidia systems in key tests. Moomoo framed the result as a breakthrough on two fronts specifically: response speed and energy efficiency versus the GB300.
A word on what these numbers are and aren't. Throughput per kilowatt is a measure of work done per unit of electricity, not raw speed — a chip can win on efficiency while a rival still finishes more total work per rack. Latency is the delay before a model starts responding, which is what makes a chatbot feel snappy or sluggish. And all of these figures come from OpenAI's own presentation; none of the three sources describe independent verification.
The context that makes this matter is power. Modern AI runs in data centers where electricity, cooling, and grid capacity are the hard ceilings, not the number of chips a company can buy. A part that does more per watt lets an operator serve more users inside the same power envelope — which is exactly the constraint OpenAI has been running into as it scales.
It also matters commercially. Nvidia's GPUs are the default hardware for training and running large AI models, and OpenAI has been one of the biggest buyers of that hardware. Designing a competitive alternative with Broadcom gives OpenAI leverage over its costs and its supply.
Why it matters: if OpenAI's numbers hold up outside its own slides, the company most dependent on Nvidia just showed it can build its way out of that dependence.