How NVIDIA DSX MaxLPS Maximizes AI Factory Throughput and Efficiency

NVIDIA DSX MaxLPS is designed to address a central infrastructure problem: AI data centers often reserve enough electricity for every accelerator to reach peak power simultaneously, even when real workloads rarely demand that much. By sharing unused power among participating resources, the technology can help customers deploy up to 40% more GPUs within the same approved budget. The approach matters because it could expand useful compute capacity without requiring an immediate increase in utility service or physical plant.

This conservative provisioning strategy stems from the many electrical limits governing an AI factory, including utility connections, substations, distribution equipment, racks, nodes, and individual GPUs. Operators must keep every managed boundary below its approved threshold while still meeting application throughput and latency targets. Static planning simplifies those constraints, but it can strand power inside unused per-node reservations. That stranded capacity can leave additional accelerators offline despite the facility as a whole remaining below its limit.

The gap appears because AI workloads have variable power profiles. Training cycles through computation, communication, synchronization, and checkpointing, while inference moves among prefill, decode, memory-bound operations, networking, and idle periods. Request shapes and concurrency levels can also make two instances of the same model consume different amounts of power. DSX MaxLPS monitors actual consumption and reallocates available capacity among participating resources while preserving the operator’s aggregate budget and policy boundaries.

At its core, MaxLPS combines chip, system, thermal, and software controls within the limits of available land, power, and shell capacity. Dynamic Power Software manages a loop built around topology, telemetry, policy, allocation, and enforcement. Operators map infrastructure into resource groups, collect power measurements at multiple levels, define priorities and reserves, and raise participating GPU limits when other resources are using less than their allocations. The system then compares measured consumption with the group budget and changes allocations as conditions evolve.

NVIDIA and Nscale evaluated the approach with Kimi K2.5 workloads running on NVIDIA GB300 NVL72 systems at Nscale’s Verne campus in Keflavík, Iceland. The site operates entirely on renewable energy, adding a sustainability dimension to the infrastructure test. The evaluation used NVIDIA Blackwell Ultra GPUs, Kimi K2.5 in FP4, NVIDIA Dynamo, and NVIDIA TensorRT LLM. It also employed an 8,000-token input sequence and a 1,000-token output sequence while examining both control behavior and workload trade-offs.

For AI operators, the immediate benefit is potentially more useful output from an already approved power envelope rather than a larger power allocation. Nscale’s involvement provides a real deployment environment in which telemetry and policy controls can be tested, while NVIDIA contributes the integrated hardware and software stack. The method does not create additional electricity or remove electrical safeguards; instead, it coordinates existing headroom more efficiently. That distinction makes rigorous validation essential before operators apply the technique across larger fleets.

MaxLPS also offers a repeatable model for infrastructure planning: define resource groups, establish conservative policies, measure real workload behavior, and expand GPU deployment only when the control loop demonstrates that limits remain intact. Industry interest is likely to center on whether improved power utilization translates into predictable throughput without compromising latency, reliability, or thermal management. As AI factories grow, dynamic capacity management could become as important as adding faster accelerators. The next stage will be broader operational evidence showing how consistently the system handles changing inference demand, maintenance events, and facility-wide constraints.

Turn any website into clean JSON.

Ready-made scrapers & AI agents

Try Apify