OpenAI and Broadcom announced Jalapeño, OpenAI’s first intelligence processor designed to accelerate large-scale language model (LLM) inference. This represents a significant expansion of OpenAI’s strategy to control the full stack of AI development. The chip went from initial design to production in just nine months, and OpenAI’s unique model enabled rapid turnaround. The companies plan to work with data center partners to deploy accelerators at gigawatt scale, signaling significant investment and expected demand for the new technology. “Jalapeño is part of our long-term full-stack infrastructure strategy to make computing richer, resulting in faster, more reliable, and more affordable AI for people and businesses,” said Greg Brockman, president and co-founder of OpenAI.
Jalapeño chip: LLM-optimized inference architecture
The Jalapeño chip announcement marks OpenAI’s step beyond model development into custom hardware design, demonstrating a full-stack strategy for controlling the entire AI infrastructure. Developed in nine months from initial design to tape-out manufacturing, this first intelligence processor represents a fast and advanced semiconductor development cycle. This accelerated timeline was aided by the application of OpenAI’s proprietary models to optimize aspects of the chip design process and demonstrated the potential feedback loops in which AI improves the tools used to create it. Early tests show that Jalapeño offers improved performance per watt compared to current accelerators, an important metric for the energy-intensive task of running large language models.
The chip’s architecture prioritizes minimizing data movement and balancing computing resources, aiming for realized performance close to its theoretical peak. Broadcom’s silicon implementation and Tomahawk networking technology are central to achieving this at scale. Richard Ho, who leads OpenAI’s hardware program, explained that the chip is “designed from the ground up for LLM inference using detailed insights gained from close collaboration with OpenAI researchers” and is optimized for the specific needs of frontier AI models. This is not an adaptation of existing hardware, but rather a clean-sheet design focused on the demands of modern LLM inference, and promises to reduce latency and increase efficiency for interactive AI products.
Jalapeño was designed from the ground up for LLM inference using detailed insights gained from close collaboration with OpenAI researchers.
Richard Ho, head of OpenAI’s hardware program
9-month tapeout with OpenAI model
The artificial intelligence hardware market is currently dominated by established players who are adapting existing chip architectures to suit the demands of large-scale language models. OpenAI breaks this down with a vertically integrated approach that extends control from model development to silicon design. This strategy culminated in the rapid production of OpenAI’s first intelligence processor, Jalapeño, which was delivered to company executives in nine months. This is a schedule that significantly shortens typical advanced semiconductor development cycles. This joint development effort leveraged Broadcom’s silicon implementation expertise and Celestica’s system integration capabilities, but was fundamentally driven by OpenAI’s deep understanding of LLM requirements.
The world is transitioning to a computing-powered economy.
Greg Brockman, OpenAI President and Co-Founder
Broadcom silicon enables gigawatt-scale deployments
Broadcom is playing a pivotal role in moving OpenAI’s AI strategy into physical infrastructure, as evidenced by the rapid development and impending deployment of the Jalapeño chip. This collaboration is not just about manufacturing. It is about building a consistent vertically integrated system from model design to silicon implementation. This speed points to a new paradigm in which AI can help create the hardware that powers it, potentially reducing computing costs across industries. The scale of this effort is significant. OpenAI plans to deploy Jalapeño in collaboration with data center partners, indicating significant investment and expectations for strong demand.
Broadcom’s contributions extend beyond chip manufacturing to networking technologies such as Tomahawk silicon, which is essential for large-scale production and data transfer. “Our collaboration with OpenAI represents a fundamental commitment to the physical infrastructure expansion required for the next decade of AI,” said Hock Tan, President and CEO of Broadcom, emphasizing the long-term nature of the partnership. OpenAI’s approach focuses on optimizing every layer of the stack, from chip architecture to service systems, with the goal of creating AI. This architecture prioritizes minimizing data movement and maximizing resource utilization, aiming for performance levels approaching theoretical peaks and ultimately broader access to advanced AI capabilities.
Our collaboration with OpenAI represents a foundational commitment to the physical infrastructure expansion needed for the next decade of AI.
Mr. Hock Tan, President and CEO of Broadcom
Full-stack integration improves AI efficiency and access
OpenAI’s announcement of the Jalapeño Intelligence Processor marks a decisive move towards vertically integrated AI infrastructure, with implications for both performance and accessibility. The company is now directly involved in chip manufacturing, a strategy aimed at optimizing the entire AI pipeline, from algorithms to hardware, beyond model design. Developed in partnership with Broadcom and Celestica, this full-stack approach aims to address the limitations of existing accelerator technologies and improve efficiency. Engineering samples are already showing promising results, with early tests showing significantly higher performance per watt than current accelerators while running ML workloads at production target frequencies and powers such as GPT‑5, 3‑Codex, and Spark.
