Vera Rubin Architecture: Redefining the Economics of AI Factories

In the high-stakes world of AI infrastructure, the economics are driven by one metric: tokens per dollar. NVIDIA’s latest MLPerf Inference v6.1 results show that the Vera Rubin NVL72 architecture is setting a new ceiling for what is possible. By optimizing the entire stack—from the H200 GPUs to the NVLink interconnects—NVIDIA has demonstrated that system-level performance is now the primary lever for scaling AI factories.

The Vera Rubin debut highlights a shift in semiconductor design. It’s no longer just about the transistor count on a single die; it’s about how those dies communicate across a massive, liquid-cooled rack. The NVL72 architecture treats the entire rack as a single, giant GPU, drastically reducing the latency that usually plagues large-scale AI inference. This is crucial for real-time Physical AI applications, where delays in processing can result in mechanical errors.

Furthermore, the MLPerf results emphasize the importance of continuous software optimization. NVIDIA’s TensorRT-LLM software has allowed the hardware to achieve performance gains far beyond what was possible at launch. As the industry moves toward more complex 'agentic' AI, the ability for semiconductors to handle massive throughput with minimal energy consumption will be the deciding factor in who wins the race to AGI. The Rubin architecture is the first clear look at the hardware that will power the next decade of autonomous intelligence.

Source: NVIDIA Blog