NVIDIA Unveils DSX MaxLPS to Boost AI Token Throughput by Up to 40% Per Megawatt
NVIDIA demonstrated its DSX MaxLPS power orchestration technology, which dynamically optimizes power allocation across AI factories to deliver up to a 40% increase in token throughput per megawatt. Integrated into Emerald AI's Conductor platform, early testing by cloud provider Lambda demonstrated a 24% boost in token throughput under a fixed power budget. Power grid capacity and energy availability have become the single biggest bottleneck scaling modern AI data centers. By enabling facilities to squeeze significantly more compute output out of fixed power envelopes without needing new electrical infrastructure, NVIDIA directly addresses the industry's critical energy constraint. In Lambda's real-world test using a 19-node cluster, DSX MaxLPS increased throughput from 4 million to 5 million tokens per second within the same power budget as a 16-node full-power deployment. Additionally, NVIDIA noted that sites equipped for Vera Rubin and Groq 3 LPX racks can house up to 40% more GPUs within existing site power limits.
## BACKGROUND
Modern AI models rely heavily on large-scale GPU clusters, where token throughput—the rate at which models process or generate text and data units—serves as the core metric for AI service capacity. As data centers scale up, power providers often cap the total wattage available to a facility, making energy efficiency (performance per watt) far more vital than peak hardware clock speeds.