OpenAI and Broadcom on June 24 introduced Jalapeño, their first custom AI chip built for large language model inference. The company said the ASIC moved from design to tape-out in just 9 months, with deployment scheduled to begin by the end of 2026. The launch extends OpenAI’s reach from models and products down into the chip layer.
A custom ASIC built specifically for LLM inference
Jalapeño was designed from scratch for modern LLM inference rather than adapted from a general-purpose accelerator. At the announcement, Broadcom president and CEO Hock Tan handed a sample chip to OpenAI CEO Sam Altman and president Greg Brockman, making the hardware partnership official in public view.
According to OpenAI, engineering samples have already run machine learning workloads in the lab at target frequency and power levels, including GPT-5.3-Codex-Spark. Early test data showed performance per watt above the current top end of the market. Richard Ho, who leads OpenAI’s hardware program, said the team focused on memory movement, networking, and serving patterns that matter most for advanced AI models, aiming to push real-world utilization closer to theoretical peak output.
Part of a broader effort to diversify beyond Nvidia
Since the generative AI boom began in 2022, OpenAI has been one of Nvidia’s major GPU buyers. Rising compute demand has pushed the company to add more hardware options. Alongside Jalapeño, OpenAI has also been working with AWS on Trainium, as well as AMD and Cerebras, to reduce reliance on a single supplier and manage costs.
Broadcom handled manufacturing and network integration for Jalapeño, including Tomahawk networking silicon, while Celestica took charge of board and rack system integration. OpenAI said the chip will begin deployment in late 2026 and will scale through multiple generations with data center partners including Microsoft, targeting gigawatt-scale expansion.