OpenAI has revealed its first custom-built inference processor, a chip named Jalapeño developed in partnership with Broadcom, marking the company’s first concrete step toward reducing its heavy reliance on Nvidia’s graphics processing units for running its AI models.
The chip was announced Wednesday and is specifically designed for inference, the process by which pre-built AI models respond to user commands in real time. OpenAI said its own AI models assisted in the development of Jalapeño, and that early testing shows significantly better performance per watt compared to current state-of-the-art alternatives. The chip is still in testing and has not yet been deployed at scale.
The Broadcom partnership was officially announced in October, but OpenAI’s ambitions to develop custom silicon have been an open secret in the industry for some time. Google and Amazon have both built their own AI accelerator chips, purpose-built silicon designed to speed up machine learning workloads, and OpenAI is now following a similar path to give itself more control over the economics and performance of its infrastructure.
OpenAI president Greg Brockman explained the thinking behind the chip in a company podcast episode released shortly after the Broadcom partnership was announced. “We have a deep understanding of the workload,” Brockman said. “We’ve really been looking for specific workloads that are underserved, and asking how can we build something that will be able to accelerate what’s possible?”
The company was explicit about how Jalapeño fits into its broader ambitions. Rather than simply building models or products, OpenAI is now designing the infrastructure beneath them, including chip architecture, memory systems, networking, and deployment systems. “Because OpenAI operates across the stack, each layer can be optimized around the same goal: making its models faster, more reliable, and more affordable for users,” the company said in its announcement. The chip’s emphasis on low operating costs for real-time coding models suggests Nvidia hardware will likely still handle more computationally intensive tasks like pre-training, but even marginal reductions in inference costs could meaningfully improve OpenAI’s financial position as it scales.
The announcement arrives as OpenAI continues to prepare for its own public market debut, with Anthropic, its primary competitor in the frontier AI space, having already filed confidentially for an IPO.
