Nvidia and Nscale report 49.2% throughput gain from shared GPU power budget

Image: Primary

Published Sunday, September 27, 2026 · 10:06 PM EDT

Nvidia and Nscale report that dynamic power allocation raised aggregate AI inference throughput by 49.2% in a data center test without increasing its 264.4 kW provisioned power budget. Their comparison expanded the managed fleet from 140 to 192 GPUs by assigning unused power capacity to nodes that needed it. The test used Kimi K2.5 workloads on Nvidia GB300 NVL72 systems. Throughput for existing instances remained effectively unchanged at the reported precision, while the 99th-percentile time to first token rose 17% from a 15.7-second baseline. Nvidia says operators must test their own workload mix, telemetry and power limits before deploying the approach at scale.

Sources

NVIDIA Technical Blog

In this story

Nvidia Moonshot AI

Published by Tech & Business, a media brand covering technology and business. This story was sourced from NVIDIA Technical Blog and reviewed by the T&B editorial agent team.

Back to Newswire

Keep reading

Full wire

Capital AI

Nebuly raises $21m to expand AI return measurement business

Turin-based Nebuly has raised $21m in a Series A round led by Planven to expand its business measuring financial returns from enterprise AI.

AI Science

MIT technique improves climate AI predictions with less training data

MIT researchers developed a training method that helped AI climate models predict unfamiliar emissions scenarios using less data, MIT News reported.

AI Infrastructure

ElevenLabs announces custom model service and adds 15 MW through Nvidia deal

ElevenLabs announced a service to fine-tune AI models using businesses' customer conversations and expanded its Nvidia agreement with 15 MW of new computing power for large customers, The Next Web reported.

AI Infrastructure

AWS shares training checks to detect silent GPU errors on HyperPod

AWS published code for detecting silent hardware errors during distributed AI training on SageMaker HyperPod and tracing suspect GPUs to nodes for replacement. The checks compare repeated calculations and model weights to catch incorrect results that hardware monitoring can miss.

Policy AI

Commerce finalizes proposed controls on Chinese AI firms' remote chip access

The Commerce Department is finalizing a semiconductor control update intended to prevent Chinese AI firms from remotely accessing advanced AI chips through data centers, The Information reported, citing sources.

Robotics AI

Researchers open source Aspire framework for robots to learn new skills

Researchers including Nvidia's Jim Fan and Berkeley's Ken Goldberg open sourced Aspire this week, a framework that lets robots discover and use new skills, WIRED reports.