Global data center power consumption will reach as much as 1,050 TWh by the end of 2026. That is roughly the annual electricity use of Japan. The driver is not a secret. AI training and inference workloads, running on GPUs that pull three to five times the power of the CPUs they are replacing, have compressed a decade of capacity growth into two years.
The analysis in Data Center Dynamics' opinion brief lands on the same point every thermal engineer has been making for eighteen months. The legacy cooling architecture cannot hold. Traditional CRAC-and-hot-aisle layouts are topping out around 40 kW per rack. AI clusters are provisioning at 100 kW and climbing. The math does not work, and the permitting offices are starting to say so before the construction crews arrive.
The DCD piece cites test data showing that direct-to-chip liquid cooling, immersion, and two-phase cooling systems cut cooling-related power consumption by 50 to 60 percent compared to air. That is not a vendor claim. That is the number operators are reporting back from commissioned MCDU loops and tank deployments running for more than a year. The cooling share of total facility power typically runs 30 to 40 percent in an air-cooled facility. Cutting that in half is the difference between a 200 MW build that fits inside its utility allocation and one that does not.
This is why the cooling conversation has moved from the facilities team to the site selection team. When a regional utility offers a 150 MW interconnect instead of the 250 MW requested, an operator with direct-to-chip architecture can still put the same compute capacity on the floor. An operator with air cooling cannot. The cooling choice has become a capacity-planning decision.
The sustainability paragraph in the DCD piece is polite. The regulatory reality is sharper. Authorities in Nevada, Arizona, and parts of Spain and Ireland are either denying evaporative cooling permits outright or requiring hybrid architectures and recycled water sourcing as a precondition. The operators who assumed they could default to cooling towers are now three to six months behind on permitting.
Cooling method selection is a water permit decision before it is a thermal engineering decision. Dry coolers, air-side economization with trim chillers, and closed-loop liquid systems are moving from the bonus column to the baseline. The facilities that will break ground on schedule in 2027 are the ones that filed water-neutral designs in 2025.
Predictive thermal modeling with digital twins is one of the few technology categories where vendor marketing and operator reality have converged. A properly instrumented digital twin predicts rack-level thermal gradients, identifies the specific CDU pressure setpoints that minimize pump electrical load, and simulates failure modes before commissioning exposes them. Operators running twins on liquid-cooled halls are catching piping routing problems in the model instead of at startup.
The ROI is measured in deferred downtime and avoided cold plate damage. A single flow imbalance that throttles a rack of H100s for a week during production training can cost more than the entire twin implementation. This is why the hyperscalers have moved from piloting twins to requiring them as part of the design package.
The DCD piece closes on a line that belongs at the top of every operator memo. The cooling buildout is an infrastructure decision, not a technology decision, and the operators who treat it as such are the ones shipping AI capacity on schedule. Liquid cooling as the default, water-neutral facility design as the permitting path, and digital twins as the commissioning standard. Facilities designed under any other set of assumptions will slip to 2028 and 2029 while the operator explains to the board why the power allocation came in short.
The question for every operator with a project breaking ground in the next eighteen months is the same. Does the thermal architecture leave enough headroom for Rubin, and the generation after Rubin, without a second construction phase? If the answer is no, the build is already obsolete.