The International Energy Agency says data center cooling that reduces water use tends to consume more energy, a trade-off relevant to facilities designed for high-power artificial intelligence systems.
Uptime Institute’s 2026 survey asked operators to identify the most common rack density at their data centers. Among 679 respondents to that question, 28% selected 4 to 5 kW, while 3% said their typical rack density was above 50 kW.
The survey’s average of the most common densities exceeded 11 kW. Excluding a small number of high-density facilities brought the average to 7.8 kW, compared with 7.5 kW on the same basis in 2025. Uptime said the 4-to-5-kW band has remained the most frequently reported range for more than 10 years.
The NVIDIA DGX GB Rack Scale Systems User Guide lists an approximate 120 kW power draw for the DGX GB200 NVL72 system, which contains 72 GPUs. Its compute trays use liquid cooling, while its networking and storage components use air cooling.
A Delta Electronics-sponsored Business Reporter contribution published on The Independent’s website said AI racks commonly use 80 to 150 kW and that next-generation platforms may approach 250 kW. NVIDIA’s system is one product example and does not show how common such racks are across data centers.
The Open Compute Project’s technical paper describes a coolant distribution unit as equipment that carries heat from the IT cooling loop to the facility loop while keeping their liquids separate. The IEA reports that local communities have expressed concern about data centers’ impact on power prices, emissions and water.
Amazon Web Services (AWS) says its cooling design is intended to let it add liquid cooling where needed. Dave Klusas, AWS senior manager of data center cooling systems, said the design aims to “move just enough air through our servers to keep them from overheating.”
Uptime Institute found that 79% of respondents said management was at least somewhat concerned about digital-infrastructure costs, including high prices for power, staff and equipment, particularly for AI-related infrastructure and IT. The Open Compute Project’s Open Data Center for AI initiative is developing standardized reference designs covering AI data center power and cooling.



