OpenAI has unveiled its first custom AI accelerator, called Jalape\u00f1o, marking the company\u2019s move into chip design as it looks to reduce the cost and improve the efficiency of running large language models (LLMs).<\/p>\n
Developed in partnership with Broadcom and Celestica, the chip is designed specifically for AI inference \u2014 the process of generating responses from trained AI models. OpenAI said early testing shows Jalape\u00f1o delivers significantly better performance per watt than current state-of-the-art AI accelerators, though detailed benchmarks will be released later.<\/p>\n
The announcement expands OpenAI\u2019s efforts to control more of the infrastructure behind its products. In addition to building models and applications such as ChatGPT and Codex, the company is now designing the hardware that powers them. Engineering samples of Jalape\u00f1o are already running machine learning workloads in the lab, including GPT-5.3-Codex-Spark, at production target frequency and power levels, according to the company.<\/p>\n
Custom chip, faster AI<\/p>\n
Unlike general-purpose AI accelerators adapted for multiple workloads, Jalape\u00f1o was built specifically for LLM inference. OpenAI said the architecture was designed around the compute, memory, networking, and serving requirements of modern AI models.<\/p>\n
The company claims the chip reduces data movement while balancing compute, memory, and networking resources to improve hardware utilization. Broadcom contributed silicon implementation and networking technologies, including its Tomahawk networking platform.<\/p>\n
\u201cJalape\u00f1o 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,\u201d said Greg Brockman, President and Co-Founder of OpenAI.<\/p>\n
Richard Ho, who leads OpenAI\u2019s hardware program, said the accelerator was optimized around the workloads most important for frontier AI systems.<\/p>\n
\u201cBased on early testing, Jalape\u00f1o will efficiently execute our most important workloads close to the hardware\u2019s theoretical limits,\u201d Ho said. The chip is also intended to support future LLMs across the broader AI industry, not just OpenAI\u2019s own models.<\/p>\n
Nine-month design sprint<\/p>\n
According to the companies, Jalape\u00f1o was developed from initial design to manufacturing tape-out in just nine months. OpenAI described the effort as potentially the fastest ASIC development cycle achieved for a high-performance advanced semiconductor.<\/p>\n
The development process involved extensive software-hardware co-design between OpenAI and Broadcom engineers. OpenAI also said its own AI models<\/a> were used to accelerate portions of the chip design and optimization workflow.<\/p>\n \u201cOur collaboration with OpenAI represents a fundamental commitment to scaling the physical infrastructure required for the next decade of AI,\u201d said Hock Tan, President and CEO of Broadcom.<\/p>\n The companies plan to deploy the accelerator at gigawatt-scale data centers<\/a> beginning in 2026. Jalape\u00f1o is the first product in what OpenAI describes as a multi-generation compute platform that will combine OpenAI-designed accelerators with Broadcom networking and connectivity technologies and Celestica\u2019s system integration expertise.<\/p>\n