NVIDIA used the AIartificial intelligence Infra Summit to present a clear shift in how it wants AIartificial intelligence infrastructure to be measured: not by peak throughput alone, but by how many validated tokens a system can produce per watt or per megawatt. The message was tied to the company’s upcoming Vera Rubin platform and to DSX, the rack-level concept NVIDIA is promoting for AIartificial intelligence factories.
The platform description spans the full stack. NVIDIA points to Vera Rubin systems, Dynamo inference software, NeMo libraries and networking components such as NVLink for scale-up computing, Spectrum-X Ethernet and ConnectX SuperNICs for scale-out connectivity, plus BlueField-based storage and security functions. The company says AIartificial intelligence factory systems now need to be codesigned from silicon to grid, because power has become the limiting factor for large-scale inference.
To support that argument, NVIDIA cited benchmark and deployment data. In MLPerf Inference v6.1, a Vera Rubin NVL72 preview submission reportedly reached up to 3.7x higher throughput than GB300 NVL72. NVIDIA also said a 288-GPUgraphics processing unitA GPU is a parallel graphics processor widely used for AI acceleration. GB300 NVL72 submission achieved 99% scaling efficiency across four racks. On the software side, the company said optimizations alone lifted its MLPerf Inference v6.1 results by up to 1.6x versus v6.0.
Power management was another major theme. With DSX MaxLPS, NVIDIA says it can continuously monitor power across GPUgraphics processing unitA GPU is a parallel graphics processor widely used for AI acceleration.s and racks, shift unused capacity to where it is needed, and improve performance within a fixed power budget. In a result from Lambda, 19 nodes fit into the same power envelope normally used for 16 full-power nodes, while token throughput rose by 24% and performance per watt improved by 23%. For next-generation Vera Rubin NVL72 deployments, NVIDIA says the same approach can enable up to 40% more GPUgraphics processing unitA GPU is a parallel graphics processor widely used for AI acceleration. capacity within the same megawatt budget.
The company also described grid-aware operation through DSX Flex, including load reduction in response to demand signals from Silicon Valley Power. In that setup, lower-priority AIartificial intelligence jobs can be throttled back and later resumed according to a workload hierarchy. NVIDIA coupled those rack- and grid-level claims with newer agentic-AIartificial intelligence results, saying its Vera Rubin NVL72 systems continue to improve through architecture, interconnect and software changes while the company positions energy efficiency as a central design target.