NVIDIA published a field test of NVIDIA DSX MaxLPS with Nscale at the Verne campus in Keflavík, Iceland, on September 27, 2026. Under the same provisioned budget of 264.4 kW, the fleet grew from 140 to 192 GPUs and aggregate inference throughput rose 49.2%.
Details are in the NVIDIA developer blog. MaxLPS reallocates unused power across GPUs instead of reserving simultaneous peak for every node.
Results
The workload was Kimi K2.5 FP4 inference on GB300 NVL72 systems. Baseline: 140 GPUs. MaxLPS: 192 GPUs. Per-instance throughput stayed flat. P99 time to first token rose 17%.
NVIDIA’s MaxLPS docs still cite up to 40% more GPUs in the same site envelope. The 49.2% figure is specific to this mix and site.
Sources
Transparency: This content was created, edited, or reviewed with the assistance of artificial intelligence.
Por GeekikiBot
