Nscale Optimizes AI Efficiency and Grid Flexibility

Nscale builds and operates GPU infrastructure for AI workloads across Europe, where data center power is a constrained and costly resource. At its Iceland AI factory, Nscale ran large-scale inference across a fleet of NVIDIA Blackwell Ultra NVL72 racks, but like most operators, it provisioned power the traditional way: reserving enough for every node to reach its rated peak simultaneously.

AI inference doesn’t work that way. Between prefill and decode phases, memory-bound operations, network activity, and idle intervals, actual GPU draw fluctuates constantly. At any given moment, significant portions of provisioned power went unused locked inside individual node reservations, unavailable to nodes that could have used it. The aggregate facility stayed below its power limit while GPU capacity sat idle.

Static alternatives offered limited relief. Lowering per-node limits with a conservative fixed ceiling kept the facility within budget, but constrained nodes when demand spiked. Nscale needed a way to make provisioned headroom available to nodes that needed it—in real time, within strict policy boundaries, without adding infrastructure.

Nscale deployed NVIDIA DSX MaxLPS, a suite of chip, system, thermal, and software technologies designed to maximize AI factory output within existing land, power, and cooling constraints. Dynamic Power Software, the software control layer within DSX MaxLPS, manages power allocation across a defined group of nodes, shifting available headroom to where it’s needed while enforcing operator-set limits at every level of the infrastructure hierarchy.

Operating an AI factory in Iceland also introduced a second dimension: grid obligations. As a major power consumer, Nscale needed the ability to respond to utility demand response events—reducing GPU power within the participating test cluster. 

NVIDIA DSX Flex addresses the grid side of the power management problem. It is a grid-aware power orchestration layer that translates utility demand response signals into coordinated GPU power reductions across the facility, automatically and within operator-defined policy boundaries. When a demand response event is called, DSX Flex interprets the grid signals and coordinates with DSX MaxLPS to reduce power limits for participating GPUs across the fleet. These limits remain in place for the duration of the event, then return to normal when it ends—all without operator intervention or interrupting running workloads.

Nscale ran Kimi K2.5 with FP4 precision on NVIDIA GB300 NVL72 racks across both high-throughput and low-latency inference instances, providing a workload mix with distinct power and service profiles, which is the kind of complementary load that exposes usable headroom. NVIDIA Dynamo and NVIDIA TensorRT™ LLM handled serving and inference optimization across the managed fleet. Operator-defined policies set node limits, group budgets, reserve requirements, and escalation behavior—ensuring the system remained within Nscale’s approved infrastructure boundaries at all times.

The evaluation compared a static 140-GPU baseline two 52-GPU high-throughput instances and one 36-GPU low-latency instance against a DSX MaxLPS configuration that added a third high-throughput instance, bringing the managed fleet to 192 GPUs. 

In a demand-response test at Nscale, NVIDIA DSX Flex coordinated with DSX MaxLPS to deliver approximately 25% GPU power reduction, sustained for 30 minutes, with GPU power recovering after the event.

The results separated fleet-level gains from application-level behavior:

• Aggregate inference throughput increased 49.2% as the third high-throughput instance came online.
• Throughput per provisioned watt increased by 49.2%.
• Per-instance high-throughput output held flat with DSX MaxLPS, confirming the new instance added capacity without degrading incumbent workloads.
• Mean measured GPU power rose from 97.0 kW to 131.8 kW and total measured power rose from 166.2 kW to 198.9 kW as the additional instance ran—the method recovered stranded provisioned headroom.
• DSX Flex enabled a 25% power reduction on demand, sustained for 30 minutes within predefined notice windows and ramp rate.

DSX MaxLPS and DSX Flex empowered Nscale to maximize inference capacity and deliver grid-grade flexibility within its existing facility footprint without compromising incumbent workload performance or service quality. 

Building on a Validated Foundation

• Nscale’s validation work established a repeatable framework for expanding DSX MaxLPS deployments: Define the managed boundary, establish a static baseline, introduce policies conservatively, add capacity incrementally, and test latency tails and fault scenarios before approving production operating limits.
• Nscale plans to extend MaxLPS evaluation to next-generation NVIDIA infrastructure as it expands its European AI factory footprint.
• Nscale is exploring opportunities to deploy DSX Flex across its European AI factory footprint, enabling faster access to additional power capacity to complement the compute density gains delivered by DSX MaxLPS.