Jalapeño, OpenAI’s first specially created AI chip, was launched on Wednesday. Developed in partnership with Broadcom, it was created especially for the inference workloads that drive ChatGPT and the company’s wider range of products.
Richard Ho, OpenAI’s hardware head, told Reuters that the Jalapeño CPU is “crafted to operate swiftly and effectively” with the LLMs powering the company’s AI apps. “We think it will do well in all future LLM versions,” Ho stated. According to OpenAI, the chip’s prototypes are already operating in its labs and meeting desired power and performance standards.
According to Hock Tan, CEO of Broadcom, the Jalapeño chip’s performance is comparable to that of Google’s Tensor Processing Units and Nvidia Blackwell chips. As the first stage of a multi-generation chip development project, OpenAI anticipates deploying Jalapeño at a gigawatt scale through data center partners, including Microsoft, starting later this year.
A Prolonged Journey to Silicon
The device is the result of a collaboration that began in October 2024, when Reuters disclosed that OpenAI was collaborating with Broadcom and TSMC to develop its first internal inference-focused chip. As the design approached tape-out in early 2025, OpenAI’s hardware team, headed by Ho, a former Google TPU engineer, expanded from about 20 to 40 members.
In October 2025, the firms publicly announced a 10-gigawatt custom accelerator agreement. Deployments are scheduled to start in the second half of 2026 and end by the end of 2029. The chip uses a systolic array architecture with high-bandwidth memory and is made using TSMC’s 3-nanometer technology.
Financing Difficulties and the Competitive Environment
There have been difficulties along the way to production. The Information revealed in May that Broadcom required Microsoft to buy about 40% of the chips before agreeing to finance production, creating a $18 billion financing barrier for the project’s first phase. The initiative was internally nicknamed “Project Nexus.” The disclosure on Wednesday indicates that those talks have progressed.
Google, Amazon, and Meta have all created proprietary chips to control rising AI compute costs, and OpenAI’s shift to bespoke hardware reflects these efforts. With agreements with Nvidia, AMD, Broadcom, and Amazon, OpenAI has committed to roughly 33 gigawatts of capacity and presently operates more than 2 gigawatts. OpenAI’s most direct move to lessen its reliance on Nvidia’s GPUs, which have cost the company billions of dollars a year, is the Jalapeño chip.
OpenAI stated that the chip was “purpose-built for the LLM workloads powering ChatGPT, Codex, the API, and future agentic products” on its official X account.