OpenAI has demonstrated its first custom AI chip, codenamed Jalapeño, and the early benchmark numbers are turning heads across the semiconductor industry.
The chip was developed in collaboration with Broadcom, according to the International Business Times, and is designed for inference — the work of actually running an AI model to answer your questions, as opposed to training it in the first place.
According to the chip-industry research firm SemiAnalysis, whose analysis is cited across the coverage, Jalapeño is outperforming comparable silicon from Nvidia, AMD and Google. SemiAnalysis published its findings under the headline "OpenAI's Jalapeño: Better Than Nvidia Blackwell." The-decoder.com reports the chip beat both Nvidia's Blackwell and its forthcoming Rubin generation in inference benchmarks, while the International Business Times specifically cites tests against Nvidia's GB300.
On the numbers: a report published by Glitchwire says OpenAI's chip delivers up to 1.9x more throughput per watt and 3.6x lower latency than Nvidia's flagship Blackwell.
Some outlets are more measured. Seeking Alpha's framing is that Jalapeño tops Blackwell "in some inference tests" — a narrower claim than a blanket win. Yahoo Finance describes the results as claims OpenAI is making about its own hardware, a reminder that these figures originate with an interested party rather than a neutral referee.
CNBC framed the obvious follow-up question directly: OpenAI says its Broadcom custom chip is a winner — what does that mean for Nvidia?
That is the crux. Nvidia's enormous valuation rests on the assumption that anyone building serious AI must buy its GPUs. OpenAI is one of the largest buyers of AI compute on earth, and a credible in-house alternative — even one confined to inference — chips away at that assumption.
Why it matters: if OpenAI's biggest customers can build their own silicon, the economics of the entire AI boom, and the pricing you eventually pay for AI services, start to shift.