Broadcom and OpenAI collaborate to build LLM accelerator architecture designed for gigawatt-scale deployment

OpenAI and Broadcom have unveiled Jalapeño, OpenAI's first Intelligence Processor designed to accelerate large language model inference, marking a significant expansion of OpenAI's strategy to control the full stack of AI development. Progressing from initial design to production in just nine months—a timeline substantially faster than typical advanced semiconductor development cycles—the chip's accelerated development was aided by OpenAI's own models optimizing aspects of the design process. The companies plan to deploy the accelerator at gigawatt scale with data center partners, signaling substantial investment and anticipated demand for the new technology.

This unveiling represents a decisive move by OpenAI beyond model development and into custom hardware design. Rather than adapting existing chip architectures, Jalapeño was designed from the ground up specifically for LLM inference. Richard Ho, who leads OpenAI's hardware program, explained that the chip was "designed from the ground up for LLM inference using detailed insights from our close collaboration with OpenAI researchers." This blank-slate approach prioritizes minimizing data movement and balancing compute resources, aiming for realized performance closer to theoretical peaks. Early testing indicates that Jalapeño delivers improved performance per watt compared to current accelerators—a critical metric for the energy-intensive task of running large language models.

The collaboration extends beyond manufacturing. Broadcom's silicon implementation and Tomahawk networking technologies are central to achieving scale, while Celestica contributes system integration capabilities. "Our collaboration with OpenAI represents a fundamental commitment to scaling the physical infrastructure required for the next decade of AI," stated Hock Tan, President and CEO of Broadcom. Greg Brockman, President and Co-Founder of OpenAI, added that "Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, and more affordable for people and businesses."

Engineering samples are already demonstrating promising results, with early testing indicating performance per watt significantly exceeding current accelerators while running production workloads at target frequency and power, including GPT-5, 3-Codex, and Spark. This vertically integrated approach—extending OpenAI's control from model development to silicon design—demonstrates a new paradigm where AI assists in creating the hardware that powers it, potentially lowering compute costs industry-wide.