OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom Intelligence Processor designed specifically for large language model inference, completing the design-to-tape-out cycle in just nine months. Engineering samples are already running production workloads in the lab, including GPT-5.3-Codex-Spark, and early tests show substantially better performance per watt than current state-of-the-art hardware. The chip was delivered to OpenAI CEO Sam Altman and Broadcom CEO Hock Tan on June 24.
OpenAI designed the chip from scratch around LLM fundamentals and serving patterns, while Broadcom contributed silicon implementation, networking technologies including Tomahawk switching silicon, and Celestica handled boards and rack systems. The architecture reduces data movement and balances compute, memory, and networking to achieve utilization closer to theoretical peak performance. In a multi-year agreement announced in October 2025, the companies plan to deploy 10 gigawatts of custom AI accelerators with Microsoft and other partners, with initial deployment targeting the end of 2026 and expansion through 2029.
For OpenAI, owning inference hardware design means controlling both the physics and economics of query serving. Custom ASICs are less flexible than GPUs but can be tuned to specific kernels and patterns, potentially lowering cost per inference token significantly. The nine-month development cycle—accelerated by using OpenAI's own models to help design and optimize parts of the silicon—signals that the barrier to custom chip entry has fallen enough for frontier AI labs to build in-house. Broadcom's ASIC business, which already fabricates custom chips for other hyperscalers, gains recurring revenue and intellectual positioning as the workshop for AI companies seeking leverage over hardware costs.