OpenAI introduced Jalapeño, its first custom-designed AI chip built specifically for large language model (LLM) inference workloads. The chip highlights a major shift in OpenAI’s infrastructure strategy. The tech giant aims to reduce reliance on NVIDIA’s highly sought-after GPUs and gain greater control over AI computing resources. Jalapeno has been developed with Broadcom, a semiconductor and infrastructure software provider.
Speaking about the new launch, Greg Brockman, President and Co-Founder of OpenAI, said, “Jalapeno is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI that is faster, more reliable, more affordable for people and businesses, and can be used to solve more important problems.”
While AI accelerators usually support many different workloads, Jalapeño is designed specifically for inference. The inference allows AI models to take inputs and produce outputs. Examples of the applications of inference include ChatGPT and coding assistants. According to OpenAI, this chip has been built from scratch to serve upcoming versions of LLMs.
OpenAI has already tested the processor internally with its GPT-5.3-Codex-Spark model. Preliminary reports suggest the chip is delivering the expected performance and efficiency in the laboratories of OpenAI.
The launch reflects a broader trend among leading AI companies to develop proprietary silicon. With demand for AI processing power surpassing the available supply, businesses are seeking alternatives to the GPUs from NVIDIA, which dominate the industry. OpenAI joins companies such as Google, Amazon, Microsoft, and Meta in building custom chips tailored to their AI workloads.
Looking ahead, Broadcom CEO Hock Tan said, “Jalapeño delivers performance comparable to NVIDIA’s Blackwell processors and Google’s Tensor Processing Units for inference tasks. The chip is expected to help OpenAI lower operating costs while improving efficiency at scale.”
OpenAI also claimed that Jalapeño was the first step in a multi-generation chip roadmap. Development took around nine months, and the chip is expected to be installed throughout the OpenAI network by the end of 2026. The manufacturing of the processor will be done by TSMC, while the computer servers will be designed by Celestica.