OpenAI has announced 'Jalapeño,' an LLM-optimized inference chip developed in collaboration with Broadcom.



OpenAI has announced its first custom processor, ' Jalapeño ,' which it designed and manufactured in collaboration with semiconductor company Broadcom. Jalapeño is a chip specifically designed for OpenAI's inference system and was developed in just nine months from design to product, leveraging OpenAI's AI capabilities.

OpenAI and Broadcom unveil LLM-optimized inference chip | OpenAI
https://openai.com/index/openai-broadcom-jalapeno-inference-chip/



OpenAI, Broadcom debut custom Jalapeño chip for AI inference - SiliconANGLE
https://siliconangle.com/2026/06/24/openai-broadcom-debut-custom-jalapeno-chip-llm-inference/

In October 2025, OpenAI entered into a strategic partnership worth hundreds of billions of yen with Broadcom, with the goal of 'independently developing a 10-gigawatt class chip.'

OpenAI signs strategic partnership with Broadcom to develop its own AI processing chip, planning to produce a 10-gigawatt class chip with a contract worth hundreds of billions of yen - GIGAZINE



The newly announced Jalapeño is not a repurposed design for general-purpose AI workloads, but a newly designed chip based on modern LLM inference. In addition to Broadcom, Celestica, which is responsible for boards, racks, and system integration, also participated in the development. Broadcom is responsible for semiconductor packaging and Tomahawk network silicon, supporting large-scale production and data center deployment.

OpenAI stated, 'Jalapeño is an accelerator designed based on OpenAI's vision for the future of LLM inference and is the first AI accelerator in the multi-generation computing platform that both companies are jointly building. This platform aims to make advanced AI faster, more reliable, and more accessible to more people.'



As of the time of writing, OpenAI has revealed very few details about the design of Jalapeño. However, OpenAI has stated that the underlying architecture 'reduces data movement,' suggesting that it may be designed to reduce 'data movement between logic circuits and off-chip memory,' which is one of the major performance bottlenecks in inference clusters.

According to OpenAI, the prototype chip is currently running machine learning workloads, including GPT-5.3-Codex-Spark, and its final performance under target frequency and power consumption conditions is being measured. OpenAI reports that 'at the time of writing, although still in the experimental stage, it has shown performance per watt that significantly surpasses state-of-the-art products.'

Greg Brockmann, co-founder and president of OpenAI, commented, 'The world is moving towards a computing-driven economy. Jalapeño is part of our long-term full-stack infrastructure strategy to enrich computing resources, which will result in AI being faster, more reliable, more affordable for individuals and businesses, and enabling them to solve more critical problems. By designing more of the stack in-house, we can continue to deliver advanced intelligence more efficiently and make advanced AI more widely available.'



Richard Ho, who leads OpenAI's hardware program, added, 'Jalapeño was designed from the ground up for LLM inference, leveraging detailed insights gained from close collaboration with OpenAI researchers. We optimized the architecture around the kernel, memory movement, networking, and serving patterns that are most important for state-of-the-art AI models. Initial test results suggest that Jalapeño will efficiently run our most critical workloads at a level close to the theoretical limits of the hardware.'

Broadcom CEO Hock Tan said, 'Our collaboration with OpenAI represents our fundamental commitment to scaling the physical infrastructure needed for AI over the next decade. This is just the beginning of a multi-generational roadmap. By directly co-developing our industry-leading silicon with OpenAI, we will enable the deployment of gigawatt-scale data centers with Microsoft and other partners starting in 2026.'

OpenAI plans to have an initial deployment of Jalapeño by the end of 2026, aiming for gigawatt-scale deployment with data center partners such as Microsoft. OpenAI also described Jalapeño as 'the first step in a multi-generational computing platform,' suggesting plans to develop additional inference processors in the future.

Related Posts:

in AI,   Hardware, Posted by log1i_yk