NVIDIA Says DSX MaxLPS Can Fit 40% More GPUs Into the Same Power Budget

Static power planning reserves enough power for every node to reach its specified peak simultaneously, even though AI workloads rarely draw constant power, and unused capacity inside one per-node reservation cannot be applied to another node. [1] NVIDIA says DSX MaxLPS dynamically allocates power across participating resources through policy-governed power sharing, enabling customers to deploy up to 40% more GPUs within the same approved power budget. [1]

NVIDIA and Nscale evaluated the approach with Kimi K2.5 workloads on NVIDIA GB300 NVL72 systems at Nscale’s data center on the Verne campus in Keflavík, Iceland, which is powered entirely by renewable energy. [1] The evaluation measures power-performance trade-offs, explains controls for maintaining electrical limits, and presents a repeatable validation method that operators can use before deploying the technology at scale. [1]

DSX MaxLPS addresses the gap between power reserved for simultaneous GPU peaks and the lower, variable consumption observed during normal workloads. The central comparison is static power planning, which is straightforward but can strand capacity because unused power in one node’s reservation cannot be reassigned to another. NVIDIA and Nscale’s Kimi K2.5 test gives operators a validation pattern rather than only a capacity claim. What the sources don't address: how the reported gain varies across other models, facilities, latency targets, or power-distribution constraints.