OpenAI and Broadcom Unveil Jalapeño AI Inference Chip

Nine months. That's the record time it took OpenAI, in collaboration with Broadcom (NASDAQ: AVGO), to bring "Jalapeño," its first Intelligence Processor, from design to production. Unveiled today, this accelerator is specifically engineered to optimize large language model (LLM) inference, marking a pivotal step in OpenAI's strategy to expand its full-stack platform.

"Jalapeño" stands as the inaugural AI accelerator within a multi-generation compute platform that both companies are jointly building. Its core mission is clear: to render advanced artificial intelligence faster, more dependable, and broadly accessible to a wider audience. This chip embodies an architectural vision deeply rooted in OpenAI's understanding of future LLM inference needs and its internal model roadmap.

The delivery of "Jalapeño" to OpenAI CEO Sam Altman and President Greg Brockman by Broadcom President and CEO Hock Tan and President Charlie Kawwas highlights OpenAI's ambition to construct the entire technological stack underpinning its models and products. This involves extending beyond software to delve into hardware, designing the core infrastructure from the ground up.

OpenAI meticulously designed the chip, leveraging its profound insights into LLM fundamentals, informed by its roadmap of models, kernels, serving systems, and product requirements. Partners Broadcom and Celestica played crucial roles in industrializing the platform, contributing expertise in chip implementation, board and rack system integration, high-performance networking, and scalable production systems.

While OpenAI is still finalizing performance measurements, initial testing indicates that "Jalapeño" will deliver substantially better performance per watt compared to current state-of-the-art solutions. Engineering samples of the chip are already running machine learning workloads in the lab at production target frequency and power, including the GPT-5.3-Codex-Spark model. A comprehensive technical report detailing its performance is expected in the coming months.

"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 for people and businesses, and can be used to solve more important problems," stated Greg Brockman, President and Co-Founder of OpenAI.

The "Jalapeño" architecture minimizes data movement and meticulously balances compute, memory, and networking resources. This approach aims to achieve realized utilization much closer to theoretical peak performance. Broadcom's silicon implementation and networking technologies, including its Tomahawk networking silicon, are instrumental in scaling this platform to large-scale production.

"Jalapeño" is not a repurposed general-purpose accelerator; it's a blank-slate design specifically for modern LLM inference. It's engineered to combine the raw power and throughput of today's leading AI accelerators with latency closer to the fastest specialized inference systems. This makes it exceptionally well-suited for interactive, large-scale LLM products such as ChatGPT, Codex, and the API.

OpenAI's full-stack advantage means that every layer, from chip architecture to the end-user product experience, is optimized with a singular goal: making its models faster, more reliable, and more affordable for users. This creates a powerful "flywheel effect": superior infrastructure drives compute efficiency, which in turn enables better training and serving, ultimately leading to more capable AI models.

The ultimate purpose of this endeavor is straightforward: inference is the point where AI directly impacts people. Every improvement in cost, speed, and reliability can manifest as a quicker ChatGPT response, a Codex task that executes more steps with less waiting, a more economical API product to build, or more dependable access during peak demand. "Jalapeño" empowers OpenAI to transform more of its infrastructure into valuable intelligence for students, developers, small businesses, and researchers alike.

This chip represents the initial phase in a multi-generation compute platform slated for initial deployment by the end of 2026. It is designed to be deployed at gigawatt scale with data center partners, including Microsoft, as confirmed by Broadcom's Hock Tan. This collaboration underscores a fundamental commitment to scaling the physical infrastructure essential for the next decade of AI.