NVIDIA AI Servers Run on 45°C Coolant in Efficiency Leap

What Happened

NVIDIA has introduced a data center cooling configuration for its latest AI servers that uses liquid coolant operating at 45 degrees Celsius, a temperature warmer than a typical hot tub. The company says the approach represents one of the most significant efficiency improvements in data center infrastructure to date, allowing facilities to reject heat without relying on energy-intensive chilling equipment.

Background

Data center cooling has become one of the primary operational cost drivers and energy constraints for large-scale AI infrastructure. Conventional cooling systems require chilled water or refrigerant loops that must be mechanically cooled to temperatures well below ambient, consuming significant electrical power in the process. As AI workloads have intensified, driven by the training and inference demands of large language models and other neural network architectures, the heat output of GPU clusters has grown correspondingly.

NVIDIA holds a dominant position in the AI accelerator market. Its GPU hardware, including the H100 and successor Blackwell series chips, is deployed across major cloud providers and enterprise data centers worldwide. Thermal management of these chips has been a persistent engineering challenge, as each accelerator can dissipate several hundred watts of heat under full load.

How the System Works

The 45-degree Celsius coolant threshold is significant because it exceeds what many standard data center cooling systems are designed to handle. Traditional facilities target supply water temperatures well below 30 degrees Celsius. By engineering servers that can operate effectively with warmer coolant, NVIDIA's design allows data centers to use what the industry calls "free cooling," in which outdoor air or cooling towers reject heat without mechanical refrigeration for a greater portion of the year and in a wider range of climates.

The approach uses direct liquid cooling routed to server components, a method that transfers heat more efficiently than air cooling alone. By raising the acceptable coolant temperature ceiling, the system reduces or eliminates the need for chillers, which are among the largest energy consumers in a data center facility.

Infrastructure and Deployment Implications

NVIDIA describes the 45-degree Celsius specification as applicable to its latest AI server platforms. Facilities adopting the configuration would need liquid cooling infrastructure capable of distributing warm-water loops to server racks, a departure from the air-cooled raised-floor layouts common in older data centers.

The practical effect for operators is a reduction in power usage effectiveness (PUE), the standard industry metric that measures total facility energy consumption against IT equipment energy consumption. Lower PUE values indicate less energy spent on overhead functions such as cooling. Data centers running large GPU clusters at scale can spend as much on cooling infrastructure as on the compute hardware itself, making efficiency improvements at this layer commercially material.

NVIDIA did not disclose specific PUE improvement figures in the wire report, nor did it name customer deployments or facilities currently using the configuration.

Industry Context

The push toward higher-temperature liquid cooling aligns with broader data center industry trends. The American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) has progressively raised recommended inlet temperature allowances for server equipment over the past decade. Several hyperscale operators, including facilities operated by Google and Meta, have publicly reported experiments with higher coolant temperatures and outside-air economization.

The timing of NVIDIA's announcement follows a period of intense industry focus on the total energy footprint of AI infrastructure. Governments in the United States, European Union, and several Asian markets have begun scrutinizing data center power consumption and water usage as AI deployment scales.

NVIDIA has not announced a specific commercial availability date for the full server platform incorporating the 45-degree Celsius cooling specification, and further technical details are expected to accompany product launch disclosures.