NVIDIA DSX Cuts AI Data Center Power Needs by Up to 30%

Power availability is the single biggest constraint on expanding AI infrastructure right now, and DSX MaxLPS directly attacks that ceiling. NVIDIA's platform claims the same token output as conventional setups while consuming up to 30% less electricity, which means operators can squeeze more inference capacity out of every megawatt they have already contracted.

Nscale's test in Iceland makes the efficiency claim concrete: within a fixed power budget, it achieved 49% higher inference throughput with no additional power contracts, no new grid capacity, and no facility expansion.

The roster of cloud and infrastructure providers exploring DSX includes CoreWeave, Google Cloud, Lambda, Nebius, and Nscale, among others. That breadth matters because it signals the platform is being evaluated across meaningfully different deployment environments.

For investors tracking GPU demand, if operators can extract 49% more inference work from the same power envelope, the economic case for filling existing data centers with more NVIDIA hardware strengthens before a single new megawatt comes online.

NVIDIA's reference designs for DSX also use closed-loop liquid cooling and dry coolers to enable near-zero facility cooling water consumption, with savings cited at 54 million liters. That water efficiency angle adds a separate sustainability argument that increasingly matters to hyperscalers operating under regulatory and public scrutiny in Europe.