OPENAI AND BROADCOM ANNOUNCE JALAPENO CHIP FOR AI INFERENCE

OpenAI and Broadcom have announced a custom silicon chip called Jalapeño, designed specifically for large language model inference in data centres. The Application-Specific Integrated Circuit took nine months to develop and was created based on detailed insights from conversations between Broadcom engineers and OpenAI researchers. Both companies state that Jalapeño will be deployed in data centres by the end of this year, marking the first generation of a long-term project to refine the technology over time.

Broadcom designed the chip from scratch using information from OpenAI's roadmap for future models and products. OpenAI claims that early testing indicates Jalapeño will deliver substantially better performance per watt than current state-of-the-art inference systems, though the company stated it has not completed performance measurement and will present a detailed technical report in the coming months. The chip represents a more specialised alternative to the general-purpose hardware currently used in existing data centre inference systems.

OpenAI's development of custom silicon reflects the company's broader strategy to own the full technology stack behind its models and services, potentially reducing dependence on external suppliers such as Nvidia. Custom chips have become increasingly relevant as AI companies compete for limited data centre capacity during the current period of rapid compute demand. Broadcom, already an established chipmaker for data centre infrastructure, has recently expanded its business to provide custom processors to hyperscalers and teams building frontier AI models.