Nvidia Vera Rubin NVL72 Rack at Hot Chips 2026 (www.servethehome.com)

At Hot Chips 2026, NVIDIA unveiled its Rubin GPU, a critical component of the Vera Rubin platform, designed to meet the demands of agentic AI—a complex workload that integrates observation, reasoning, and action across various computing environments. The Vera Rubin architecture represents a complete stack, combining seven chips and five racks, including the Rubin GPU, Vera CPU, and NVLink 6, emphasizing a redesign of not just silicon but the entire infrastructure needed to support advanced AI functionalities. This platform is set to enhance token throughput significantly, allowing for dynamic sequences and higher interaction levels while optimizing performance metrics that exceed traditional benchmarks. Significantly, the Rubin GPU leverages innovative features such as adaptive sparsity and a cable-free, liquid-cooled design that improves efficiency and serviceability. By focusing on parameters like tokens per watt and time to first token instead of just raw computing power, NVIDIA aims to redefine operational performance for AI applications. The infrastructure can accommodate larger context lengths in AI workloads, enabling effective processing up to 400k tokens—a marked shift from the previous limitations. This strategic focus on maximizing revenue capability underscores NVIDIA’s intent to position the Rubin as not merely a chip but as part of a comprehensive and scalable AI factory solution, setting it apart in the competitive landscape against alternatives like AMD's Helios.