OpenAI details the economics behind its Jalapeño chip

OpenAI is developing a custom chip, called Jalapeño, with Broadcom to reduce the rising cost of running large language models. The company says the project is aimed at lowering dependence on thirdparty hardware and improving efficiency for inference, the process of serving model responses at scale. The article says OpenAI spent about $8.4 billion last year on keeping ChatGPT responsive and expects that figure to rise to roughly $14 billion this year as usage grows. It also notes that OpenAI has committed about $1.4 trillion to computing over the next eight years, while currently generating around $25 billion in annual revenue. The chip is designed as a dedicated inference processor, with OpenAI defining the architecture and Broadcom handling silicon engineering and networking integration. Manufacturing will be done by TSMC, while Celestica will build board and rack systems. OpenAI says early samples are already testing frontier workloads, and deployment in data centers is scheduled to begin by the end of 2026. The move reflects a broader industry shift toward custom silicon as major AI companies try to control infrastructure costs and improve performance. OpenAI says the project is part of a longterm strategy to make compute more efficient and support future model growth.