OpenAI Launches Jalapeño AI Chip With Broadcom to Boost AI Speed and Efficiency

OpenAI has announced its first custom AI chip, called Jalapeño, marking a major step in the company's effort to design not only artificial intelligence models but also the hardware used to run them. The chip has been developed in collaboration with Broadcom and is designed primarily to improve AI inference performance, reduce latency and increase energy efficiency.

AI inference refers to the process through which a trained artificial intelligence model processes a request and generates a response. Every time a user interacts with an AI chatbot, coding assistant or other AI-powered service, inference hardware is used to process the request.

OpenAI said Jalapeño has been specifically designed around the requirements of its AI models. The company worked with Broadcom on the chip, while Celestica contributed to the boards, rack systems and production hardware.

According to OpenAI, the custom AI chip was developed by co-designing the chip, software, memory, networking and serving systems to work together with its models. This approach is aimed at improving performance and efficiency across the entire AI infrastructure stack.

The company tested Jalapeño using models including GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T. OpenAI said the chip delivered between 1.5 and 1.9 times more AI work per watt compared with the systems used for comparison.

Jalapeño also recorded between 1.7 and 3.6 times lower end-to-end latency across the tested models. For highly interactive AI workloads, OpenAI said performance reached up to 4.1 times higher.

The improved efficiency could allow AI systems to process more requests while consuming less electricity and reducing the response time experienced by users. These gains could become increasingly important as companies scale AI services to handle millions of requests.

The launch also represents OpenAI's broader strategy to gain greater control over the infrastructure powering its AI models. Rather than relying entirely on third-party hardware, the company is working to optimise its chips, software, memory, networking and serving infrastructure together.

OpenAI said Jalapeño moved from initial design to manufacturing tape-out within nine months, with its own AI models assisting engineers during parts of the chip design and optimisation process.

However, the new chip is not expected to replace Nvidia hardware immediately. OpenAI plans to deploy Jalapeño in limited volumes by the end of 2026 before expanding deployment through 2027. The company also plans to continue working with other hardware partners, including Nvidia, as part of its broader computing strategy.