Lambda's recent test of the Nvidia DSX MaxLPS power optimization platform allowed the company to run 19 compute nodes using the energy envelope of a 16 node baseline. This configuration delivered a 24% increase in token throughput and a 23% gain in performance per watt for Blackwell servers. The software optimization enables 1.4x more tokens per megawatt compared to standard configurations.

Ian Buck introduced the technology at the AI Infra Summit 2026, where Nvidia also showcased DSX Flex for grid responsive power management. This tool autonomously adjusts AI workloads based on real time electricity grid signals to protect critical jobs. These efficiency tools address power constraints that have become the primary bottleneck for competitive AI factory deployments.