OpenAI Ships Its First In-House Inference Chip: Jalapeño

OpenAI unveiled Jalapeño — a proprietary chip designed specifically for running (inferencing) AI models, not training them. It is the company's first silicon solution: until now, OpenAI was entirely dependent on third-party GPUs, primarily Nvidia's.

1Benchmark Results

According to an independent InferenceX test by analytics firm SemiAnalysis, Jalapeño outperforms all current market solutions on two metrics simultaneously: tokens per user (response speed) and tokens per watt (energy efficiency). OpenAI VP of Hardware Richard Ho called it "the best of both worlds" — low latency and high throughput at the same time.

2What This Changes for the Industry

Building a proprietary chip is a strategic move: OpenAI reduces its dependence on Nvidia and gains control over token costs. If Jalapeño genuinely delivers better performance per watt than existing GPUs, it creates direct downward pressure on industry pricing. After Google TPU and Amazon Trainium, another major player has entered the custom AI silicon space.

For API users and developers, Jalapeño means potential token price reductions from OpenAI — the chip can serve more requests on the same power draw. Competition in AI hardware is no longer Nvidia's monopoly: the next two to three years in this segment look set to be eventful.