NVIDIA DSX MaxLPS Scales 140 GPUs on Renewable Power

NVIDIA and Nscale ran a joint evaluation on 140 GPUs at a data center in Keflavík, Iceland, testing whether AI factories can be built with less electrical headroom than standard practice demands. The site, part of Nscale's Verne campus and powered entirely by renewable energy, hosted NVIDIA GB300 NVL72 systems running Kimi K2.5 workloads under a system called DSX MaxLPS. The premise: most data centers provision power for a scenario that almost never happens, every GPU hitting peak draw simultaneously, and that buffer sits mostly unused during normal operation.

DSX MaxLPS attacks that waste with policy-governed power sharing, moving power dynamically across GPUs instead of locking in fixed allocations. NVIDIA says the approach lets customers pack in up to 40% more GPUs within the same approved power budget, without breaching any electrical limit along the chain from utility feed down to individual chip. The Nscale test set out to measure what that tradeoff actually looks like in practice, tracking performance against power draw across real training and inference patterns, and to build a validation method other operators could run before committing hardware at scale.