The Jalapeño, labeled an Intelligence Processor, shifts OpenAI’s strategy from software-centric to a vertically integrated model. While Nvidia currently captures a 75% profit margin on its high-end chips, OpenAI operates on a much thinner 33% margin. With $1.4 trillion earmarked for computing power over the next eight years, the company is betting that proprietary hardware can resolve the data-movement bottlenecks that plague general-purpose accelerators.
Inside Jalapeño: OpenAI’s Pivot to Custom Silicon
Facing a projected $14 billion annual bill for server operations, OpenAI is moving to break its reliance on third-party hardware. The company has unveiled the Jalapeño chip, a custom-designed processor built in partnership with Broadcom, aimed at slashing the massive costs of running large language models at scale.
Richard Ho, who leads the hardware program, stated that the architecture is optimized specifically for LLM inference, balancing compute and networking resources to push utilization toward theoretical peaks. The design integrates Broadcom’s Tomahawk networking silicon, while TSMC oversees manufacturing in Taiwan. To accelerate a development cycle that usually takes years, OpenAI engineers utilized their own language models to automate portions of the silicon design, reaching the manufacturing tape-out stage in just nine months. Initial deployment is scheduled for late 2026, with plans to scale alongside Microsoft to support future gigawatt-scale data centers.
Comments (0)
No comments yet. Be the first!