Nvidia made a significant announcement at Hot Chips 2026 regarding its approach to managing power consumption in data center GPU installations. The company unveiled details about its DSX MaxLPS technology, a comprehensive power management system designed to help facilities get more computing performance from fixed electrical budgets. For organizations planning large-scale GPU deployments, this development carries real implications for project costs and operational efficiency.

The Power Constraint Reality

Data centers face a hard ceiling on performance determined by the amount of electrical power available to the facility. Unlike individual chip performance metrics, which often dominate tech discussions, the total power delivered to a building and distributed across server racks represents the ultimate bottleneck for AI system productivity. Nvidia’s presentation centered on this critical constraint and demonstrated how intelligent power allocation could change the equation for facility operators.

The company showcased its next-generation Vera Rubin GPU architecture through the lens of a fixed 100 megawatt power budget scenario. Nvidia claims that by applying all of Rubin’s power management technologies together with DSX MaxLPS capabilities, operators could install approximately 40,000 of these GPUs, equivalent to roughly 40 Rubin DGX SuperPODs, within that power envelope. The company estimates such a configuration could deliver up to 2 zettaFLOPS for inference tasks and up to 1.4 zettaFLOPS for training workloads.

Static Provisioning: The Problem with Fixed Power Budgets

electrical power management facility
Photo by Fré Sonneveld

Historically, data center operators have approached power allocation conservatively. They would calculate worst-case peak power consumption for each rack and provision accordingly. This practice often results in substantial power waste because individual racks rarely operate at maximum draw simultaneously, and workload demands fluctuate throughout the day and across different applications.

In a concrete example from current-generation systems, a facility with a 540-kilowatt power budget might install four 135-kilowatt systems under traditional static provisioning. However, real-world usage patterns often leave approximately 170 kilowatts unused. This stranded capacity represents an entire additional rack that could run within the same electrical footprint if power could be dynamically reallocated based on actual demand rather than worst-case assumptions.

How DSX MaxLPS Changes the Game

DSX MaxLPS introduces intelligent, dynamic power management that operates continuously at the chip, rack, and multi-rack levels. The system monitors actual power consumption in real time and redistributes available capacity to wherever it is needed most at any given moment. When one workload requires additional resources while another operates below capacity, the technology shifts power allocation accordingly without manual intervention.

The approach includes workload-specific power profiles tailored to different computing scenarios, similar to the balanced and high-performance modes familiar to PC users. These profiles optimize behavior for inference, training, and memory or compute-bound applications. In a test case with current-generation Grace Blackwell hardware running a specific inference workload, Nvidia demonstrated that DSX MaxLPS reduced GPU power draw from an assumed 1400 watts to actual 1000 watts and rack power from 136 kilowatts to 101 kilowatts, with no reduction in delivered performance.

This efficiency gain means more racks can operate within the same facility budget, and organizations exploring GPU investments can stretch their power allocations further than traditional static approaches would allow.

Facility Lifecycle Flexibility

liquid cooling data center systems
Photo by Vladislav Osterman

An additional advantage emerges when facilities incorporate dynamic power provisioning from initial design. Early in a facility’s life cycle, when it operates primarily as a training center, higher power demands per GPU require conservative rack density. However, as workloads shift and older hardware transitions to inference tasks, power requirements drop substantially. Dynamic provisioning enables operators to install additional hardware in previously reserved space, generating more productive output from the same physical footprint without exceeding the facility’s electrical capacity.

Cooling and Power Efficiency

The Rubin NVL72 systems operate exclusively with liquid cooling and support inlet coolant temperatures of 45 degrees Celsius, significantly higher than previous liquid-cooled systems. This