Industry guidance emerged on preparing datacenters for the massive power consumption demands of AI workloads and GPU clusters.

As artificial intelligence takes off in enterprise settings, data center power usage is becoming a critical challenge. For data centers running typical enterprise applications, the average power consumption for a rack is around 7 kW, yet AI applications commonly use more than 30 kW per rack, according to data center organization AFCOM. This dramatic difference stems from AI's requirement for much higher processor utilization, particularly with power-hungry GPUs—Nvidia GPUs, for example, may run several orders of magnitude faster than a CPU but consume twice as much power per chip. The challenge is compounded by the fact that many data centers are already power constrained, and AI-oriented servers require greater processor density with more chips running very hot, which increases cooling demands alongside higher utilization rates.

Liquid cooling represents a significant solution for facilities pushing beyond traditional air-cooling limits. Fan cooling typically loses viability once a rack exceeds 15 kW, but water has 3,000 times the heat capacity of air, according to CoolIT Systems, a maker of enterprise liquid cooling products. John Sasser, senior vice president for data center operations at Sabey, a developer and operator of data centers, explains: "Liquid cooling is definitely a very good option for higher density loads. That removes the messy airflow issue. Water removes a lot more heat than air does, and you can direct it through pipes. A lot of HPC [high performance computing] is done with liquid cooling." While liquid cooling requires capital investment, Sasser notes it "might be a much more sensible solution for these efforts, especially if a company decides to move in the direction [of AI]."

Beyond hardware solutions, computational approaches can reduce power demands. Steve Conway, senior research vice president for Hyperion Research, notes that many workloads can operate at half or quarter precision rather than 64-bit double precision. "For some problems, half precision is fine," Conway says. "Run it at lower resolution, with less data. Or with less science in it." Double-precision calculations are primarily needed in scientific research at the molecular level, while AI training and inference on deep learning models typically don't require this level of precision—even Nvidia advocates for single- and half-precision calculations in deep neural networks.

Practical implementation requires strategic infrastructure planning. Doug Hollidge, a partner with Five 9s Digital, which builds and operates data centers, advises that new facilities allocate only a portion to higher power usage, noting "You're not going to put all of your facilities to higher density because there are other apps that have lower draw." The first step is assessing the energy supply to the building to ensure the power provider can increase capacity. David McCall, chief innovation officer with QTS, a builder of data centers, recommends spreading AI systems across racks rather than concentrating them, allowing a heterogeneous environment where most applications run at 8 to 10 kilowatts up to 15 kilowatts alongside denser AI workloads, bringing overall rack power closer to 12-15 kW—a range air cooling can manage. Hollidge emphasizes: "It's hard to give one-size-fits-all solution since all data centers are different," underscoring the importance of engineering assessment to determine which portions of a facility are best equipped for higher-density capabilities based on specific workload requirements.