OpenAI and Broadcom have unveiled Jalapeño, an inference chip built to run large language models without wasting so much juice.
OpenAI and Broadcom said Jalapeño is OpenAI’s first Intelligence Processor and is meant to be the first AI accelerator in a multi-generation compute platform built by the two outfits.
The companies claim that early testing shows Jalapeño will deliver substantially better performance per watt than today’s best systems. The chip has been built from scratch for current and future large language models across the industry.
OpenAI said the design moved from concept to production in nine months, helped by its own models. Jalapeño expands OpenAI’s full-stack effort from products and models into chips.
The accelerator is planned for gigawatt-scale deployment with data centre partners across several generations.
Broadcom president and chief executive Hock Tan and Broadcom president Charlie Kawwas delivered Jalapeño to OpenAI chief executive Sam Altman and OpenAI president Greg Brockman.
OpenAI designed the chip around what it says is its deep understanding of LLM fundamentals. That includes its model roadmap, kernels, serving systems and product needs.
Broadcom and Celestica helped industrialise the platform through chip implementation, board design, rack integration, networking and production systems.
OpenAI said Jalapeño is flexible enough to work with all LLMs, based on its view of inference needs across the industry. Engineering samples are running machine learning workloads in the lab at production target frequency and power. Those workloads include GPT-5.3-Codex-Spark.
OpenAI said it is still measuring final performance. A detailed technical report will appear in the coming months, which is where the marketing mist may clear a bit. The architecture is meant to cut data movement and balance compute, memory and networking resources.
OpenAI said this should push realised use closer to theoretical peak performance. Broadcom’s silicon implementation and networking technologies, including Tomahawk networking silicon, are meant to drag the thing into large-scale production.
OpenAI president and co-founder Greg Brockman said: “Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses, and can be used to solve more important problems. By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access.”
OpenAI hardware programme lead Richard Ho said: “Jalapeño was designed from the ground up for LLM inference using detailed insights from our close collaboration with OpenAI researchers. We optimised the architecture around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models. Based on early testing, Jalapeño will efficiently execute our most important workloads close to the hardware’s theoretical limits.”
“By co-developing our industry-leading silicon directly with OpenAI, we are enabling the deployment of gigawatt-scale data centres with Microsoft and other partners beginning in 2026,” he said.
