OpenAI's in-house artificial intelligence (AI) chip "Jalapeño" outperformed Nvidia's lineup in performance tests.
Richard Ho, OpenAI's head of hardware, said in an interview with Bloomberg on the 25th (local time) that the in-house Jalapeño beat Nvidia's top-performing "GB300" on metrics such as AI workload handled per unit of power and response speed.
He added that the company plans to deploy Jalapeño to run its AI models within the year.
Ho led Google's tensor processing unit (TPU) project before moving to OpenAI in 2023. Regarding Jalapeño, he said, "It's a really good chip. It will significantly reduce (OpenAI's) expense."
OpenAI developed Jalapeño in partnership with Broadcom, which supplies custom chips. The two companies, which released their partnership last year, emphasized in June that Jalapeño was completed at an unusually rapid pace when they unveiled it.
Ho explained that unlike typical AI chips that show strength in either processing capability or response speed, Jalapeño has both. OpenAI will decide which AI models to run on Jalapeño, allowing customers to use the chip for their goals such as expense reduction or performance improvement.
He added that Jalapeño can deliver high performance with a relatively low 700 watts (W) of power, which could help cut power expenses that make up a large share of operating costs at large-scale data centers.
IT trade outlet SemiAnalysis reported that Jalapeño is equipped with HBM4 memory offering 15.4 TB/s of bandwidth per package, and the supplier is presumed to be Samsung Electronics.
OpenAI said on the company blog that it publicly tested the new chip with its small open-source models as well as third-party models from DeepSeek and MoonshotAI, and showed a bigger performance edge on Moonshot's "Kimi" model.
However, Jalapeño has not undergone comparative testing against Nvidia's next-generation chip "Vera Rubin," which has just begun shipping.
In addition, Jalapeño was not designed for training AI models, where Nvidia is strong, and is specialized for the "inference" stage, in which already trained models respond to prompts and process tasks.
Ho said OpenAI will not immediately replace the chip technology from Cerebras, which it currently uses for relatively small models. "We have such massive computing demand that we have contracts with many suppliers, and this situation will continue for the time being," he said.
He added, "Nvidia is a really good partner, and we will need a lot of Nvidia going forward."
OpenAI is also accelerating development of follow-up in-house chips. Ho said the company is shortening development time by using its own AI models and noted that the second-generation chip is already well along in development and is expected to complete "tape-out," the final design stage, within the next few months.
OpenAI has also begun conceptual design of a third-generation chip to lower AI infrastructure expenses.