OpenAI and Broadcom build a chip to rival Nvidia’s Blackwell

OpenAI has unveiled its first custom AI chip, called Jalapeño, co-designed with Broadcom for AI inference workloads. Broadcom CEO Hock Tan claims it matches Nvidia's Blackwell GPUs and Google's TPUs in performance. The chip is designed to work efficiently with large language models and is already running in OpenAI's labs with the GPT-5.3-Codex-Spark model. OpenAI plans to deploy Jalapeño by end of 2025 as part of a multi-generation chip roadmap, with Celestica building the server systems. TSMC will manufacture the chips, which took about nine months to design. High-bandwidth memory shortages are currently pressuring Broadcom's profit margins on custom AI chips.