OpenAI Says Jalapeño Beats Nvidia—but Its Real Test Is Still Ahead

OpenAI Says Jalapeño Beats Nvidia—but Its Real Test Is Still Ahead
OpenAI is pitching Jalapeño as a powerful new bargaining chip in the AI compute race: faster responses, lower energy use and less reliance on Nvidia. The harder question is whether benchmark gains can survive the jump from controlled tests to vast, real-world deployment.

The effort began with Jalapeño’s announcement last October, developed with Broadcom and aided by OpenAI’s own models. On Tuesday, at the Hot Chips conference, the company released its first detailed results from SemiAnalysis’ public InferenceX benchmark. Richard Ho, OpenAI’s hardware chief, called them a “very, very significant performance advance over state of the art.”

OpenAI’s own account frames the chip as more than a one-off accelerator. Tested on GPT-OSS 120B, DeepSeek R1 and Kimi K2.5, it said Jalapeño delivered 1.5 to 1.9 times more work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than comparison systems. The company argues that designing chips, memory, networking and serving software together lets it keep model state local, reducing the data movement that slows inference.

That full-stack argument is also strategic. OpenAI says first-party silicon gives it greater control over “the economics of serving” its models while retaining a broad supplier portfolio that includes Microsoft, Nvidia, AMD, AWS and others. In other words, Jalapeño is meant to add leverage—not replace every outside partner.

The limitations are substantial. The comparison was against currently available Nvidia Blackwell hardware, while rival systems will advance before Jalapeño reaches broad deployment. Nor is the chip built to train frontier models. Ho stressed that OpenAI will still need Nvidia, AMD and other providers: “We’re going to need a lot of compute.”

OpenAI says it plans an initial deployment by year-end, with later generations already in development. The company’s ambition is clear: turn inference efficiency into cheaper, more responsive AI. Its proof point, however, will come when Jalapeño moves from benchmark charts into the infrastructure serving millions of users.

Continue reading https://foxvector.com/stories/01a039f8-6639-271d-7208-0394030690cf