OpenAI and Broadcom Unveil ‘Jalapeño’ AI Chip to Power Next-Generation Inference

Custom processor marks OpenAI’s move toward owning more of its AI infrastructure, reducing dependence on third-party accelerators while improving performance and efficiency.

OpenAI has taken a major step in its infrastructure strategy with the introduction of Jalapeño, its first custom AI processor developed in collaboration with Broadcom. The chip is designed specifically for AI inference—the process of running large language models in real-world applications such as ChatGPT, Codex, APIs, and future AI agents.

The announcement reflects a broader trend among leading AI companies seeking greater control over the hardware that powers their models. While Nvidia remains the dominant supplier of AI accelerators, organizations such as Google, Amazon, Microsoft, and Meta have increasingly invested in custom silicon to optimize performance, reduce costs, and secure access to critical computing resources. OpenAI’s entry into this space signals its intention to build a more vertically integrated AI platform.

According to OpenAI and Broadcom, Jalapeño was designed from the ground up for large language model inference rather than as a general-purpose AI accelerator. The architecture has been optimized around the compute, memory, networking, and data-movement requirements of modern AI models. OpenAI says early testing indicates the chip delivers significantly improved performance per watt compared with current state-of-the-art solutions, although detailed benchmarks have yet to be released.

The project also highlights the growing importance of custom application-specific integrated circuits (ASICs) in the AI era. As demand for AI services continues to surge, the cost of operating large-scale models has become a key challenge. By designing its own silicon, OpenAI aims to improve efficiency, lower infrastructure expenses, and scale AI services more economically.

Broadcom played a crucial role in the chip’s development, leveraging its expertise in custom semiconductor design and networking technologies. The collaboration also includes infrastructure partner Celestica, which is helping develop supporting systems for large-scale deployment. OpenAI said Jalapeño is the first chip in a multi-generation roadmap that will support future AI workloads and increasingly sophisticated models.

The companies plan to deploy Jalapeño at scale beginning later this year and into 2027, with the technology expected to support data-center infrastructure operated alongside strategic partners. OpenAI views the chip as part of a broader “full-stack” approach that spans models, software, infrastructure, and now custom hardware.

For the semiconductor industry, the launch underscores a growing shift toward purpose-built AI silicon. As AI workloads continue to expand, custom chips such as Jalapeño could become a key differentiator for companies seeking higher efficiency, lower operating costs, and greater control over their AI ecosystems.