Photo by Krzysztof Kowalik on Unsplash

Cooling Technologies for AI Data Centers

Taking the Plunge

Artificial intelligence (AI), machine learning, and neural networks have at least two things in common: They generate massive hype, and they massively ramp up energy demand in data centers. According to Statista, the global AI market reached approximately $244 billion in 2025 [1], and it is projected to grow by 26 percent annually through to 2031, with the US being the largest market at just south of $74 billion. Other sources cite slightly different figures, but the trend is clear: It is a steep upward climb. According to market forecasters, companies that fail to keep pace can expect to face a competitive disadvantage very soon. Sooner or later, many companies will therefore have to tackle the issue of how to integrate AI infrastructure meaningfully, particularly for providers of all orders of magnitude who are looking to offer their customers the required hardware or algorithms on a rent-to-use basis.

Power Hog GPU

AI hardware is forcing today's systems, most of which are air-cooled, to adapt. GPUs in particular are driving power consumption to unprecedented heights, so distinguishing between the thermal design point (TDP) and peak performance is important. The TDP describes the sustained load at which the system reliably remains below the maximum permissible temperature. In practice, however, power spikes regularly occur that lead to the maximum temperature being exceeded for extended periods.

The TDP values of current GPU models are now in the high three-digit and even four-digit watt range. If 168 GPUs are installed in a rack, and it is technically feasible, power consumption can reach several hundred kilowatts. For example, NVIDIA's Blackwell Ultra has a current draw of 1,400W, compared with 1,200W for its predecessor, whereas the H100, by comparison, got by with just 700W. AMD is no slouch either, with the MI355X's current