OpenAI and Broadcom Launch LLM Dedicated Chip Jalapeño

On June 24, 2026, OpenAI and Broadcom jointly released Jalapeño, OpenAI’s first “Intelligence Processor,” in San Francisco and Palo Alto. This accelerator is designed entirely around OpenAI’s vision for future big language model inference, and is also the first AI accelerator in a multi generation computing platform jointly built by both parties. Its aim is to make advanced AI faster, more reliable, and accessible to more people.

On that day, OpenAI CEO Sam Altman and President Greg Brockman received the Jalapeño chip from Hock Tan, President and CEO of Broadcom, and Charlie Kawwas, President of Semiconductor Solutions. This delivery marks an important step forward for OpenAI in its “full stack” strategy for building models and products.

Specially designed for LLM inference, not a universal accelerator

Jalapeño is a “blank paper” design specifically designed for modern large language model inference, rather than a general-purpose accelerator modified from early AI tasks. Its design is based on OpenAI’s accumulated system experience in running ChatGPT, Codex, APIs, and future intelligent agent products every day, while also taking into account the current and future needs of various LLMs in the industry. The goal is to combine the powerful computing power and throughput of today’s leading AI accelerators with the low latency of the fastest dedicated inference system, making Jalape ñ o highly suitable for large-scale interactive LLM products.

OpenAI hardware project leader Richard Ho said, “Jalapeño designed LLM inference from scratch, and we worked closely with OpenAI researchers to gain very detailed insights. We have optimized the architecture around the most critical core, data mobility, network, and server modes of cutting-edge AI models. According to early testing, Jalapeño will be able to perform our most important task loads with efficiency close to the theoretical limits of hardware.”

Nine months of streaming, AI model accelerates self design

Jalapeño took only nine months from initial design to chip fabrication, and this custom AI accelerator project may be one of the fastest ASIC development cycles in the high-performance advanced semiconductor field. This speed is due to the deep software and hardware collaboration of the OpenAI engineering team, Broadcom’s expertise in silicon implementation, and the use of OpenAI models to accelerate some design and optimization processes.

In other words, the same model that users use on a daily basis is helping to improve the operational infrastructure of future models. If AI can help engineers design better chips faster, it can reduce the computing costs of the entire industry and promote the popularization of advanced AI.

Early testing: Significant performance lead per watt, detailed report to be released several months later

Although OpenAI is still measuring the final performance, early tests have shown that Jalapeño’s per watt performance will be significantly better than current state-of-the-art products. A detailed technical performance report will be released in the coming months. This architecture reduces data movement and balances computing, memory, and network resources, making actual utilization closer to theoretical peak performance. Broadcom’s silicon implementation and network technology (including Tomahawk network chips) help the platform achieve large-scale production.

Expand OpenAI full stack platform, from products to models to chips

The release of Jalapeño further expands OpenAI’s full stack platform – from the product layer, model layer, and now extending to the chip layer. OpenAI President and co-founder Greg Brockman said, “The world is moving towards a computing powered economy. Jalapeño is part of our long-term full stack infrastructure strategy, which aims to make computing resources more abundant, making AI faster, more reliable, more affordable for individuals and businesses, and able to solve more important problems. By designing more stack layers ourselves, we can provide more intelligence more efficiently and continue to drive advanced AI towards wider applications.”

Multi generation platform planning, starting gigawatt level deployment by the end of 2026

Jalapeño is just the first step in a multi generational computing platform. The platform is planned to begin initial deployment by the end of 2026 and continue to expand in the coming years. It will combine OpenAI designed accelerators, Broadcom’s silicon implementation and network connectivity technology, as well as Celestica’s board, rack, and system integration capabilities.

Hock Tan, President and CEO of Broadcom, stated, “Our collaboration with OpenAI represents a fundamental commitment from both parties to expanding the physical infrastructure required for AI in the next decade. This is just the beginning of a multi generational roadmap. By collaborating with OpenAI to develop our industry-leading silicon chips, we are working with Microsoft and other partners to deploy gigawatt level data centers starting from 2026.”

The engineering sample has been loaded with ML, including GPT-5.3-Codex-Spark

At present, engineering samples of Jalapeño chips have been tested in the laboratory to produce machine learning loads at target frequencies and power consumption, including the GPT-5.3-Codex-Spark model. Although the final performance data is still being measured, the early results are encouraging.

OpenAI stated that Jalapeño not only serves its own model, but its architecture also has sufficient flexibility to adapt to various LLMs in the industry, thanks to OpenAI’s profound insight into current and future AI model inference needs. By reducing data handling and optimizing resource scheduling, this chip can fully unleash its hardware potential and provide higher practical utilization for large-scale inference tasks.

This release also includes an official OpenAI blog post, providing more technical details. OpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity. Broadcom is a technology leading enterprise that designs, develops, and supplies semiconductor and infrastructure software for complex critical tasks worldwide, headquartered in Palo Alto, California.

With the emergence of Jalapeño, the AI industry has taken an important step in the field of specialized inference hardware, and the nine month rapid development cycle has also demonstrated the enormous potential of AI assisted chip design. More performance data and deployment plans will be announced in the coming months.