Woofun AI reports that OpenAI has unveiled Jalapeño, its first custom accelerator chip engineered specifically for Large Language Model inference. The company handled architecture and algorithm design while partnering with Broadcom and Celestica for mass production. The initiative targets performance enhancements for ChatGPT, Codex, and API services alongside reduced computational expenses.

Leveraging AI-assisted design, the project completed tapeout in nine months, establishing a record for ASIC development speed. The chip employs algorithm-hardware co-design to optimize core processing, data movement, and network architecture. Engineering samples have successfully executed GPT-5.3-Codex-Spark workloads at target specifications, demonstrating superior energy efficiency compared to current leading devices. Broadcom manages silicon implementation and network connectivity via the Tomahawk chip, while Celestica handles system integration. Scheduled for large-scale deployment in gigawatt-level data centers by late 2026 in partnership with Microsoft, Jalapeño serves as the inaugural product in a multi-generation computing roadmap aimed at expanding full-stack capabilities.