40% more GPUs is a power-scheduling claim. Benchmark watts per token before buying racks.

NVIDIA says DSX MaxLPS can enable up to 40% more Rubin GPU capacity inside the same power budget, and its representative inference tests show 35–39% more racks at similar workload throughput. The headline sounds like a hardware-density gain. The more important change is operational: power allocation is becoming a scheduler input rather than a fixed rack assumption.

MaxLPS uses Dynamic Power Software to compare allocated power with actual consumption and move unused headroom across managed resource groups. NVIDIA says the site power envelope stays fixed while policies constrain where that reclaimed capacity can go. That turns watts into a shared production resource with its own control loop.

Static rack limits are becoming scheduling policy

Traditional planning reserves enough power for every rack to reach its specified maximum simultaneously. NVIDIA argues that real AI workloads move through compute bursts, memory-bound phases, synchronisation, checkpointing and idle gaps, so actual draw often sits below the reserved ceiling.

In NVIDIA's example, static provisioning strands 170 kW inside a 540 kW site budget, while dynamic allocation reclaims enough headroom for another rack. That does not mean every site gets the same gain. It means rack count increasingly depends on telemetry quality, workload mix and policy behaviour, not only nameplate power.

Once power becomes schedulable, capacity planning has to test the control loop as seriously as the GPU.

Benchmark watts against accepted work

NVIDIA reports roughly 1.5× performance per watt on GB200 NVL72 and 1.3–1.4× on Vera Rubin NVL72 in representative inference tests, while preserving similar throughput. Those figures are vendor results, not a universal forecast. Operators should reproduce the comparison with their own models, batch shapes, latency targets and failure modes.

The useful acceptance metric is not average rack power alone. Track completed tokens or jobs per watt alongside latency, service error rate, throttling events and policy violations. A configuration that lowers power but increases retries or misses an SLO has moved cost rather than removed it.

Treat the power controller as production infrastructure

Dynamic Power Software is currently in Developer Preview, according to NVIDIA. Before using its output for facility sizing or deployment commitments, validate how it behaves when telemetry is stale, a rack drops out, the available site budget changes or an emergency policy activates.

Separate facility limits from application policy and keep both observable. Define which resource groups may borrow headroom, set hard ceilings that the scheduler cannot cross, and preserve the reason for every allocation change. Then test static and managed runs on representative workloads before changing procurement or rack-position assumptions.

The practical shift is that AI capacity is no longer just a count of installed GPUs. In a power-constrained factory, the scheduler, workload profile and thermal design determine how much useful compute the same electrical envelope can safely deliver.

Written by the Devence Lab research team.