Nvidia says power‑sharing tool lifts AI factory output 49%

Nvidia said its new DSX MaxLPS system reallocates unused power across a data center in real time, letting facilities run up to 40% more GPUs within the same approved power budget. The company described the software in a blog post as reclaiming capacity that operators normally hold in reserve for a peak-demand moment that rarely occurs across an entire facility at once.

In a test at a partner's data center in Iceland, Nvidia said a cluster of GB300 NVL72 systems running Blackwell Ultra GPUs raised normalized aggregate throughput by 49.2%. Serving the Kimi K2.5 model on the same 264.4 kW power budget, throughput rose from 1.085 million to 1.618 million tokens per second, the company said. Efficiency rose from 4.10 to 6.12 tokens per second per provisioned watt, according to Nvidia.

Nvidia's own figures showed a tradeoff: 99th-percentile latency rose 17% under the power-sharing scheme. That could matter for workloads with strict response-time limits. The company has not said when MaxLPS will reach customers, and the results, drawn from its own blog post, have not been independently verified on hardware outside that single test.

The tool targets a persistent problem in AI data centers, according to Nvidia. Operators provision enough power for every GPU to hit peak draw at once, even though real workloads rarely do, leaving what the company called capacity left on the table. MaxLPS is aimed at Blackwell Ultra and GB300 NVL72 deployments, Nvidia said.