Power is emerging as a practical boundary for scaled AI infrastructure. A reported Nscale test points to a different way of viewing that boundary: power allocation may shape how much compute and inference capacity can be used inside a fixed envelope.
What the evidence says
NVIDIA Blog reports that Nscale tested NVIDIA DSX MaxLPS on NVIDIA GB300 NVL72 systems. In that test, the same provisioned power budget was associated with 37% more enabled GPUs and nearly 50% higher aggregate inference throughput.
Nscale is described as using the NVIDIA DSX Platform to align power allocation with workload requirements. The account also frames developer needs around compute near data, customizable open models that can be served at scale, and infrastructure designed with security in mind.
The report characterizes power as a constraint as AI factories grow, while describing demand from AI-native organizations as a force accelerating the European AI cloud ecosystem.
Operator implications
The operational signal is the reported connection between workload requirements, power allocation, enabled GPU count and aggregate inference throughput. If the result holds in relevant conditions, power orchestration could be evaluated as a capacity lever rather than treated solely as a facility constraint.
A useful evaluation would keep the power envelope fixed while comparing enabled GPUs and aggregate inference throughput across the workloads and configurations that matter to an operator. The reported percentages should remain specific to the stated test unless further validation supports broader use.
The account also presents data proximity, model customization, scalable serving and security-by-design as connected infrastructure considerations. That framing may help operators assess these requirements together when examining a power-constrained deployment.
Limits and open questions
This single primary account is not independently confirmed. The account does not establish the test methodology, workload mix, duration, comparison baseline, configuration choices or repeatability. It also does not establish whether the reported gains apply to other systems, workloads, sites or production environments.
The account further does not establish how data proximity, open-model support or secure-by-design infrastructure were measured in practice. The reported test offers a capacity question to investigate, not a basis for generalizing its results.