AI Data Centers Consume as Much as Entire Nations: This Technology Could Provide a Brake

One of the hottest topics when discussing artificial intelligence is undoubtedly energy consumption. Just these days, we have seen tensions grow in the United States due to distribution companies increasingly interested in meeting the demands of data centers, even at the expense of local citizens: emblematic cases include Lake Tahoe and the acquisition of Dominion Energy by NextEra. However, a group of researchers has reportedly managed to devise a solution to reduce, albeit marginally, the energy demands of data centers.

According to data reported by researchers at the University of Illinois Urbana-Champaign, by 2025, data centers are expected to consume 485 TWh of electricity, with about 30% of that energy used solely for cooling systems.

To tackle the problem, the team has developed a new technology based on 3D-printed pure copper cooling plates, designed with advanced mathematical algorithms. The results show over a 95% reduction in energy required for cooling compared to the traditional approaches used today in large AI-dedicated facilities.

The issue arises from the high energy density of modern AI accelerators. A single NVIDIA GB200 chip, for example, draws 1,200 watts, equivalent to about 28.8 kWh per day. Almost all of this energy is directly transformed into heat due to the Joule effect, an inevitable phenomenon during the operation of semiconductors.

The situation becomes even more complex in large AI clusters. The researchers cite the example of the Colossus 1 data center of xAI, a facility using approximately 220,000 GPUs with a total consumption of 300 MW. Without adequate cooling systems, the heat produced would reach levels incompatible with the operation of the hardware.

Traditional air systems use metal heatsinks and powerful fans to disperse heat generated by CPUs and GPUs. However, the new generations of AI accelerators have reached thermal densities that put pressure on these solutions. For this reason, many infrastructures are transitioning to direct-to-chip liquid cooling, a technology that utilizes metal plates with microchannels through which the coolant flows.

However, solutions currently available on the market prioritize production simplicity and manufacturing costs. The internal channels of cooling plates feature relatively simple geometries, often rectangular or cylindrical, made with aluminum or stainless steel alloys.

The paper published on EurekAlert explains that the U.S. research group decided instead to completely redesign the internal structure of the plates via a process of "topological optimization." The algorithms created much more complex configurations, with irregular and pointed surfaces designed to maximize heat exchange and reduce resistance to the passage of the coolant.

To achieve these geometries, the team used a technique called ECAM (Electrochemical Additive Manufacturing), an additive printing system capable of working with pure copper at details between 30 and 50 micrometers, smaller than the thickness of a human hair.

According to the results published by the researchers, the new cooling plates ensure up to a 32% improvement in cooling performance compared to conventional systems. At the same time, the pressure loss of the circuit decreases by up to 68%, a factor that significantly reduces the energy needed for pumping the liquid.

The estimates developed by the team show particularly interesting scenarios in large AI data centers. A 1 GW facility based on air cooling typically requires about 550 MW additional dedicated to cooling. With the new technology, the energy requirement would instead drop to about 11 MW.

This approach would allow for lowering the PUE (Power Usage Effectiveness) parameter to around 1.011, a value extremely close to the ideal theoretical limit of 1.0. Today, the most efficient hyperscale data centers generally operate between 1.1 and 1.3.

The researchers clarify that the data related to large facilities still derives from simulations and theoretical models. There are currently no operational implementations on a gigawatt scale. Nevertheless, the technology could offer very important advantages for the future of AI data centers, especially considering the continuous growth of energy demand related to artificial intelligence.