OpenAI's First Chip, Jalapeño, Unveiled: 9 Months to Tapeout, Planned Gigawatt-scale Deployment by the End of 2026

According to Countervail Beating monitoring, OpenAI has launched its first custom accelerator chip named Jalapeño, specifically designed for Large Language Models (LLMs) inference. OpenAI was responsible for the chip's architecture and algorithm design, and collaborated with Broadcom and Celestica to drive industrial-scale production. Jalapeño is aimed at directly boosting the performance of ChatGPT, Codex, API interfaces, and future intelligent systems while reducing computational costs.

Benefiting from OpenAI's cutting-edge AI model-assisted design, Jalapeño went from initial concept to tape-out in just 9 months, setting a record for the fastest development of an Application-Specific Integrated Circuit (ASIC). The chip features algorithm-hardware co-design, focusing on the bespoke core for large language models, data movement, and network architecture restructuring, achieving near-maximum hardware utilization efficiency. The initial engineering samples have successfully run workloads such as GPT-5.3-Codex-Spark at the target frequency and power consumption in the lab, with early tests showing a significantly higher energy efficiency ratio compared to existing top-tier computational devices.

In the technology division of the industrial chain, Broadcom is mainly responsible for Jalapeño's silicon implementation and network connectivity technology, integrating the Tomahawk chip within it; board, rack, and system integration are supported by Celestica. As the first product in a multi-generation computing platform roadmap, Jalapeño is scheduled to begin its initial large-scale deployment in gigawatt-level hyperscale data centers in late 2026 in collaboration with partners such as Microsoft, aiming to expand the full-stack platform capabilities and reduce inference costs.