OpenAI and Broadcom unveil Jalapeño chip for AI inference systems

OpenAI and Broadcom unveiled Jalapeño on June 24 as OpenAI’s first Intelligence Processor and the first AI accelerator in a multi-generation compute platform, a move that pushes the ChatGPT maker deeper into the chip business it once bought from others. The processor was designed from scratch around OpenAI’s models, kernels, serving systems and product needs, and Broadcom President and CEO Hock Tan and Broadcom President Charlie Kawwas delivered the chip to OpenAI CEO Sam Altman and President Greg Brockman. Jalapeño reached engineering samples in nine months and is already running machine-learning workloads in the lab, including GPT-5.3-Codex-Spark.

The design is aimed at the economics of inference, the compute that answers user prompts and powers products like ChatGPT. Early testing shows Jalapeño should deliver performance per watt substantially better than current state-of-the-art, while the architecture cuts data movement and balances compute, memory and networking so realized utilization moves closer to theoretical peak performance. Broadcom’s platform is built for deployment at gigawatt scale with data center partners, and the stack includes Broadcom’s Tomahawk networking silicon as well as Celestica’s chip implementation, board and rack integration, and scalable production work.

The chip sits inside OpenAI and Broadcom’s larger October 13, 2025 plan for 10 gigawatts of OpenAI-designed AI accelerators and network systems, with Broadcom targeting deployments to start in the second half of 2026 and finish by the end of 2029. The companies had been working together for 18 months before going public with that plan.

OpenAI and other labs are struggling to secure enough computing horsepower for the latest chatbots and coding tools. Jalapeño is part of OpenAI’s strategy to build the full stack behind its models and products rather than depend entirely on outside suppliers. Tan compared the chip with Nvidia’s Blackwell line and Alphabet’s Google tensor processing units, while OpenAI hardware chief Richard Ho said it was built to work efficiently with current and future large language models. A detailed technical report on performance will follow in the coming months.