OpenAI and Broadcom Launch Jalapeño, Revolutionary AI Chip

OpenAI and Broadcom Launch Jalapeño, a Revolutionary Inference Chip

OpenAI and Broadcom recently unveiled a major innovation in the field of artificial intelligence with the launch of Jalapeño. This new accelerator, designated as OpenAI's first Intelligence Processor, is specifically designed to optimize the inference of large language models (LLMs). This initiative is part of a strategic collaboration between OpenAI and Broadcom to develop a multi-generational computing platform aimed at making artificial intelligence faster, more reliable, and accessible to a broader audience.

During the official presentation, OpenAI CEO Sam Altman and President Greg Brockman received Jalapeño from Hock Tan, President and CEO of Broadcom, and Charlie Kawwas, President of the company. This event marks a crucial milestone in OpenAI's strategy to build the entire infrastructure necessary for its models and products.

A Custom Design for LLM Needs

The design of Jalapeño was entirely carried out by OpenAI, based on a deep understanding of LLM requirements. This design was guided by OpenAI's roadmap concerning models, kernels, service systems, and product needs. In collaboration with Broadcom and Celestica, OpenAI worked on the industrialization of this platform, integrating chips, card and rack systems, as well as a high-performance network and scalable production systems. Jalapeño is designed to be flexible enough to work with all LLMs, leveraging OpenAI's insights into current and future inference needs in the AI industry.

Engineering samples of Jalapeño are currently being tested in labs, executing machine learning workloads at targeted frequencies and power levels for production, including the GPT-5.3-Codex-Spark model. While final performance evaluations are still underway, initial results indicate that Jalapeño could offer a significantly higher yield per watt compared to current technologies. A detailed technical report on these performances is expected in the coming months. The chip's architecture is designed to minimize data movement and balance computing, memory, and network resources to achieve real-world usage close to theoretical maximum performance. Broadcom's silicon implementation, along with its networking technologies, including Tomahawk network silicon, plays a key role in scaling the platform.

An Inference Platform Tailored for the Future

Jalapeño stands out for its innovative design, specifically dedicated to the inference of modern LLMs. Unlike general-purpose accelerators, Jalapeño is designed to meet the specific requirements of the systems used daily by OpenAI, such as ChatGPT, Codex, the API, and future agentic products. The goal is to combine the power and throughput of current AI accelerators with latency comparable to the fastest specialized inference systems, making Jalapeño particularly suited for large-scale interactive LLM products.

In developing Jalapeño, OpenAI is not only creating cutting-edge models or products based on these models; it is also designing the infrastructure that supports them. This includes chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and user experience. By optimizing each layer of this infrastructure, OpenAI aims to make its models faster, more reliable, and more affordable for users.

Accelerated Development in Nine Months

The development of Jalapeño was accomplished in record time, just nine months from initial design to manufacturing. This custom AI accelerator program represents what is considered the fastest ASIC development cycle ever achieved in the field of advanced high-performance semiconductors. This speed is the result of close collaboration between OpenAI's engineering teams and Broadcom's implementation expertise, as well as the use of OpenAI models to accelerate certain parts of the design and optimization process.

The same models used by OpenAI users contribute to enhancing the infrastructure needed to run future models. If artificial intelligence can help engineers design better chips more quickly, it could reduce the cost of computing across the industry and promote democratic access to advanced AI.

Towards a Multi-Generational Platform

Jalapeño represents the first step towards a multi-generational computing platform, designed for initial deployment by the end of 2026 and intended to expand in the coming years. This platform combines accelerators designed by OpenAI with Broadcom's silicon implementation, networking and connectivity technologies, and Celestica's expertise in cards, racks, and systems.

Making Advanced AI Accessible to All

The ultimate goal of this project is to make inference, where AI interacts with users, more efficient. Every improvement in cost, speed, and reliability can translate into faster responses from ChatGPT, smoother Codex task execution, a less expensive API product to develop, or more reliable access even during peak demand.

Democratizing AI means making advanced models available, reliable, and affordable enough for more people to benefit from them daily. Jalapeño helps OpenAI transform a larger part of its infrastructure into useful intelligence for students, developers, small businesses, researchers, enterprises, and anyone looking to learn, create, or solve complex problems.