How a custom-built, AI-accelerated silicon project is setting the stage for a cheaper, faster, and gigawatt-scale future.
- Purpose-Built for LLMs: Designed from scratch for modern AI inference, Jalapeño promises substantially better performance per watt than current state-of-the-art processors.
- Record-Breaking Development: The chip went from initial design to manufacturing tape-out in just nine months, with OpenAI’s own models actively helping to accelerate the engineering process.
- Gigawatt-Scale Deployment: Built alongside Broadcom and Celestica, this first-generation accelerator marks OpenAI’s shift to a full-stack infrastructure strategy designed to democratize AI access.
The artificial intelligence industry has reached a pivotal bottleneck: the sheer physical infrastructure required to power the next generation of intelligence. Today, OpenAI and Broadcom (NASDAQ: AVGO) took a massive step toward shattering that bottleneck by unveiling Jalapeño, OpenAI’s inaugural Intelligence Processor. Delivered directly to OpenAI CEO Sam Altman and President Greg Brockman by Broadcom’s executive leadership—President and CEO Hock Tan and President Charlie Kawwas—this new accelerator represents a tectonic shift in the AI landscape. OpenAI is no longer just a software and model company; it is officially a full-stack technology powerhouse.
A Blank Slate for Modern AI
Unlike many of today’s leading AI accelerators, which were adapted from earlier, generalized workloads, Jalapeño is a “blank-slate” design. It was engineered from the ground up with a singular focus: running Large Language Models (LLMs) as efficiently as physically possible. Drawing on OpenAI’s deep, proprietary understanding of LLM fundamentals, kernels, and serving systems, the architecture drastically reduces data movement. By perfectly balancing compute, memory, and networking resources, the chip achieves a realized utilization that sits unprecedentedly close to its theoretical peak performance.
“Jalapeño was designed from the ground up for LLM inference using detailed insights from our close collaboration with OpenAI researchers,” noted Richard Ho, head of OpenAI’s hardware program. “Based on early testing, Jalapeño will efficiently execute our most important workloads close to the hardware’s theoretical limits.”
Engineering samples are already hard at work in the lab, running at their production target frequency and powering massive ML workloads, including the advanced GPT‑5.3‑Codex‑Spark. While final performance metrics are still being officially measured ahead of a forthcoming technical report, early tests indicate that Jalapeño will deliver a staggering leap in performance per watt compared to the current state-of-the-art.
The Flywheel of Full-Stack Innovation
The creation of Jalapeño highlights a powerful operational flywheel for OpenAI. By controlling the entire stack—from the underlying chip architecture and memory systems to the models and user interfaces—OpenAI can optimize every single layer for a unified goal. Better infrastructure yields greater compute efficiency. This efficiency allows for better training and serving, which powers more capable models. Those models translate into superior products for developers and everyday users, generating the revenue needed to reinvest in the next generation of hardware.
Greg Brockman, President and Co-Founder of OpenAI, summarized this vision: “The world is moving to a compute-powered economy. Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable… and can be used to solve more important problems.”
AI Building AI: A Nine-Month Miracle
Perhaps the most fascinating aspect of Jalapeño’s creation is the speed at which it was born. Co-developed from initial concept to manufacturing tape-out in a mere nine months, the project represents what is believed to be the fastest ASIC development cycle ever achieved in high-performance semiconductors.
This unprecedented speed was made possible by deeply integrating OpenAI’s own models into the workflow. The very same AI models that serve millions of users daily were utilized to accelerate parts of the chip’s design and optimization process. It is a profound meta-milestone for the tech industry: AI is now effectively lowering the cost and time required to build the hardware that runs AI.
Scaling to the Gigawatt Era
Jalapeño is not a standalone experiment; it is the first step in a multi-generation compute platform. To bring this vision to life, OpenAI has partnered with industry titans. Broadcom brings its unparalleled silicon implementation and networking technologies, including the robust Tomahawk networking silicon, while Celestica provides critical expertise in board, rack, and system integration.
“This is just the beginning of a multi-generation roadmap,” stated Broadcom CEO Hock Tan. “By co-developing our industry-leading silicon directly with OpenAI, we are enabling the deployment of gigawatt scale data centers with Microsoft and other partners beginning in 2026.”
Democratizing Intelligence
From a broader perspective, the introduction of Jalapeño is about much more than benchmark scores and tape-out records; it is about human access. Inference—the process of an AI model generating a response—is the exact point where artificial intelligence touches people’s lives.
Every drop of efficiency squeezed out of Jalapeño translates to tangible benefits for the end user: a more instantaneous response from ChatGPT, a complex coding task completed by Codex without a wait, or an API that is suddenly cheap enough for a small startup to build upon. By driving down the cost of compute and scaling up reliability, OpenAI and its partners are actively democratizing advanced intelligence—ensuring that the transformative power of AI is affordable, dependable, and accessible to students, creators, and enterprises worldwide.
