OpenAI and Broadcom Unveil Custom Jalapeño AI Inference Chip

OpenAI has announced a collaborative project with Broadcom Inc. to develop a custom silicon processor dubbed Jalapeño, specifically engineered for artificial intelligence inference. Unlike Nvidia's general-purpose graphics cards which manage both training and inference, the Jalapeño chip focuses exclusively on performing inference, the process by which AI models generate responses to user queries. Early performance benchmarks suggest the new architecture achieves higher efficiency per watt compared to current industry standards.

The chip's architecture prioritizes the reduction of data movement between its logic circuits and memory, addressing a primary performance bottleneck for AI clusters. To further optimize throughput, OpenAI intends to integrate Broadcom's Tomahawk networking technologies, including the Tomahawk 6 chip series capable of processing 1.6 terabits of traffic per second. This hardware is intended to power specialized server racks developed in partnership with Celestia Inc., which are slated for deployment by the end of 2026.

By controlling its hardware stack, OpenAI gains potential operational independence from traditional suppliers and creates an opportunity to offer on-premises AI solutions to enterprise clients. This vertical integration strategy mirrors shifts seen in other large-scale AI enterprises and arrives at a critical juncture for the company as it prepares for a potential public offering. Future iterations of this compute platform may extend into model training capabilities, further challenging incumbent hardware leaders.