OpenAI and Broadcom plan custom AI inference chip for data centers
The companies say Jalapeño is built for large language model inference and should reach data centers by year-end.
By Hana Yoshida · Markets Reporter
2 min read
OpenAI and Broadcom have announced Jalapeño, a custom chip aimed at running large language model inference in data centers. The project matters because OpenAI is trying to gain more control over the computing systems behind products such as ChatGPT and Codex while demand for AI capacity remains tight.
In a company announcement cited by Ars Technica, OpenAI and Broadcom described Jalapeño as the first generation of a longer custom-silicon program. Both companies said they expect to improve the chip line over time rather than treat this release as a one-off product.
A chip built around OpenAI’s model plans
Broadcom said Jalapeño is an application-specific integrated circuit designed from the ground up for large language model inference. The company said its work drew on detailed discussions with OpenAI researchers and on OpenAI’s roadmap for future models and products.
According to Broadcom, the chip’s design and production took nine months. The companies are positioning Jalapeño as a more tailored option than the processors now used for inference systems in data centers, Ars Technica reported.
OpenAI said early tests indicate Jalapeño will offer substantially better performance per watt than current top systems. The company also said it has not completed its measurements and plans to release a detailed technical report in the coming months.
Part of a broader push for custom silicon
Ars Technica reported that OpenAI wants to own more of the technology stack that supports its models and services. That goal could reduce its reliance on outside suppliers such as Nvidia, while potentially improving efficiency or performance through tighter coordination between models, software and hardware.
The announcement comes as AI companies compete for limited data center capacity, according to Ars Technica. Custom chips have become one way for model developers and large cloud operators to try to extract more usable computing power from available infrastructure.
Broadcom already sells chips to companies building compute infrastructure, Ars Technica reported. The company has also expanded its business in custom chips for hyperscalers and teams working on frontier AI models during the recent AI boom.
OpenAI and Broadcom said Jalapeño chips are scheduled to be deployed in data centers by the end of this year. They did not release full benchmark figures or detailed technical specifications with the announcement.
This story draws on original reporting from Ars Technica.