How NVIDIA DSX Solves the AI Factory Power Crisis

Production Deployment of Grid Flexibility in Silicon Valley

On a sweltering August evening, Silicon Valley Power sent a signal to an AI factory in Santa Clara, California. Varun Sivaram and his Emerald AI team watched the event via Zoom from a San Francisco conference room alongside data center engineers and utility representatives, tracking power consumption drop from four megawatts to three megawatts. According to Varun Sivaram, watching the deployment across thousands of NVIDIA GPUs was tense. We were watching with bated breath, Sivaram stated, describing it as their first time deploying across thousands of NVIDIA GPUs. Mansi Shah, head of product at Emerald AI, added, This feels kind of like a SpaceX rocket launch. The facility runs Emerald AI’s Conductor platform as a participant in Silicon Valley Power’s Flexible Load Interconnect Program. When a grid constraint signal arrives, Conductor executes a predefined workload hierarchy that slows or reschedules lower-priority tasks while high-priority inference keeps running smoothly. Silicon Valley Power has since transmitted more than 200 demand signals to the facility, with the automated reduction succeeding every single time without requiring manual operator intervention.

Maximizing Token Throughput with NVIDIA DSX MaxLPS

While the Santa Clara facility proves grid participation works in commercial settings, cloud provider Lambda has released validation numbers for NVIDIA DSX MaxLPS, a software suite designed to optimize AI factory throughput per megawatt. Lambda tested the software on a five-rack, 19-node cluster running NVIDIA HGX B200 GPU Servers. According to Lambda’s findings released at the AI Infra Summit, running 19 nodes at an 85 percent power policy within the same overall power budget as 16 nodes operating at full power yielded a 24 percent increase in cluster-wide token throughput. Token production rose from roughly 4 million tokens per second to 5 million tokens per second, while performance per watt improved by 23 percent. With our proof of concept, we believe we’ve moved beyond the limitation of fixed power budgets, said Dave Ward, president of cloud services at Lambda. Ward noted that NVIDIA DSX MaxLPS reallocates power headroom dynamically between nodes based on whether they handle training or inference workloads, thereby reclaiming stranded capacity. NVIDIA projects that MaxLPS can enable up to 40 percent more GPU capacity for next-generation Vera Rubin NVL72 AI factories under suitable deployment conditions.

Whole-Factory Architecture and 800V DC Power Infrastructure

Addressing power limits requires optimizing the entire facility rather than isolated components, according to NVIDIA infrastructure presentations. Ian Buck, NVIDIA’s vice president of hyperscale and high-performance computing, emphasized AI factory efficiency as the core theme of his keynote at the AI Infra Summit. NVIDIA founder and CEO Jensen Huang has noted that a one-gigawatt factory will never become a two-gigawatt factory, highlighting the physical limits facing modern data center operators. To address these constraints, the NVIDIA DSX platform integrates simulation tools, open-source lifecycle management software, and reference designs spanning compute, networking, and liquid cooling. Furthermore, NVIDIA is incorporating an 800-volt direct current power architecture into its reference designs to reduce conversion complexity and support denser accelerated computing racks.