# NVIDIA's New Tech Boosts Energy Efficiency in AI Factories DSX MaxLPS maximizes AI output within fixed power budgets. - **Source:** [Nvidia Developer Blog](https://developer.nvidia.com/blog) - **Published:** 2026-08-21 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/7122b51d-625a-4114-8c1f-ac8651d27336 ## The Power Challenge in AI Factories AI factories now focus not on how many GPUs fit, but how much AI output each megawatt can produce. Traditional data center design wastes resources with static power allocation. NVIDIA's analysis shows only 60% of site power is usable for AI output due to inefficiencies in power distribution and cooling systems. ## Drawbacks of Static Provisioning Static rack provisioning allocates power based on theoretical peak demand, creating stranded capacity when workloads operate below these levels. Unused power in one rack can't be transferred to neighboring racks that could convert it into tokens. This results in significant efficiency losses across AI factories, impacting operational costs. ## The MaxLPS Solution NVIDIA DSX MaxLPS combines chip, thermal, system and software technologies. Based on Maximum Land Power Shell concept, it optimizes land, grid power and physical infrastructure constraints. Its Dynamic Power Software (DPS) enables real-time power redistribution from racks to GPUs, improving overall utilization. ## Dynamic Power Software (DPS) DPS models data center topology from facility level down to GPUs. Operators define resource groups, power budgets and governance policies. The system continuously compares allocated versus consumed power, redistributing unused capacity within managed groups. This maintains the site power envelope while increasing productivity. ## Operational Efficiency Gains DPS operates by collecting GPU, rack and group-level power telemetry. It reallocates power within policy constraints and validates compliance with group budgets. During grid events or emergencies, the system auto-adapts without manual intervention. This reduces operational overhead and improves response times to power fluctuations. ## Power Usage Comparison In a 100MW AI factory, 20% of grid power goes to facility overhead, 10% to rack losses. Another 10% is lost during failures, restarts and checkpoints. Traditional methods leave only 60MW for AI workloads. MaxLPS recovers significant portions of these losses through dynamic allocation and optimized thermal management. ## Workload Behavior Insights AI workloads transition through training, inference and post-processing phases, each with distinct power profiles. Static provisioning wastes resources by preparing for peak demands that rarely occur. MaxLPS dynamically matches power to actual needs, reallocating unused capacity to other units in real-time. ## Thermal Design Innovation MaxLPS operates with 45°C liquid cooling, enabling stable performance at higher temperatures than conventional data centers. This solution requires early site engagement and infrastructure optimization, but delivers 1.3-1.5x better performance per watt per GPU. The thermal design is critical for maintaining stability during dynamic power shifts. ## Capacity Increase Validated on NVIDIA Vera Rubin NVL72 and GB200 NVL72 systems, MaxLPS reduces per-rack power requirements. It enables up to 40% more GPU capacity within the same physical footprint. This accelerates ROI for existing facilities and potentially lowers costs for new builds, making AI infrastructure more accessible. ## Industry Impact AI factories have become industrial-scale systems where energy efficiency is the largest operational cost factor. Solutions like MaxLPS can reduce AI production costs, enabling more affordable AI services. For large-scale providers, this technology represents a crucial competitive advantage in an energy-constrained market. ## Future Perspective NVIDIA's innovation marks a significant step toward sustainable AI infrastructure. With increasing energy regulations and environmental concerns, dynamic power management is poised to become the future standard. The industry's adoption rate will directly influence AI's proliferation across sectors. ## Conclusions and Recommendations DSX MaxLPS offers a comprehensive solution for power efficiency in AI factories. Operators should focus on workload profiling, thermal design and early site engagement. This technology is particularly critical for large-scale AI service providers. As energy costs remain a major adoption barrier, MaxLPS could become an industry turning point for sustainable AI growth. ## Sources covering this story - https://developer.nvidia.com/blog/maximizing-ai-factory-performance-per-watt-with-nvidia-dsx-maxlps/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._