OpenAI Unveils Jalapeño Chip for LLM Inference, Aiming for Gigawatt-Level Deployment by 2026

According to Beating Monitor, OpenAI has launched its first acceleration chip, Jalapeño (also known as the "intelligent processor"), specifically designed for large language model (LLM) inference. OpenAI led the chip’s architecture and algorithm design, collaborating with Broadcom and Celestica to drive industrial-scale manufacturing. Jalapeño is engineered to directly enhance the speed and reduce the computational cost of ChatGPT, Codex, API services, and future agent-based products. Leveraging OpenAI’s cutting-edge AI models in its co-design process, Jalapeño achieved tape-out in just nine months from initial concept—setting a new record for the fastest development of a high-end ASIC. The chip employs algorithm-hardware co-design, rearchitecting around a dedicated LLM core, data movement, and networking infrastructure to achieve utilization rates approaching theoretical hardware limits. Initial engineering samples have successfully run workloads such as GPT-5.3-Codex-Spark at target frequencies and power levels in lab environments, demonstrating significantly superior energy efficiency compared to existing top-tier computing hardware. In terms of industrial division of labor, Broadcom is primarily responsible for Jalapeño’s silicon implementation and networking technologies, integrating the Tomahawk chip; Celestica provides support for board, rack, and system-level integration. As the first product in OpenAI’s multi-generation computing platform roadmap, Jalapeño is scheduled for initial large-scale deployment by the end of 2026 in gigawatt-scale data centers operated in partnership with Microsoft and others, to expand full-stack platform capabilities and reduce inference costs.