Speaking at the AI Infra Summit, NVIDIA Vice President Ian Buck presented new full-stack infrastructure updates focused on optimizing energy efficiency for agentic AI workloads. Key advancements centered around the NVIDIA Vera Rubin architecture, NVLink scaling, and the NVIDIA DSX software platform, which measures system efficiency in validated agentic tokens per megawatt.
As part of the disclosures, cloud provider Lambda published initial benchmark results testing NVIDIA DSX MaxLPS on Blackwell server architectures. Using a testbed of 19 nodes, DSX MaxLPS dynamically allocated power across mixed training and inference jobs, delivering up to 1.4x more tokens per megawatt by optimizing unallocated thermal overhead.
Additionally, NVIDIA highlighted a joint project with Emerald AI and Silicon Valley Power demonstrating dynamic grid integration. Utilizing DSX Flex software, the facility automatically throttled non-critical workloads in response to grid demand signals without interrupting priority training tasks, illustrating how data centers can manage power draw interactively with local utilities.
Why it matters
Shifts infrastructure performance metrics from peak FLOPS to energy efficiency (tokens per megawatt).
Enables data center operators to increase rack density and reclaim unused power dynamically via software.
Helps facility operators mitigate grid constraints by participating in automated utility demand-response programs.
Source: blogs.nvidia.com
