Data Center Power and Cooling Trends for the AI Era

Many data centers struggle to get enough power into the facility to run AI applications. Those that manage it face an enormous cooling problem — how exactly do you dissipate that much heat? Hence, both of these areas are front and center in data center planning and design.

Analyst firm Omdia recently delved into the trends shaping both areas in a report on AI data centers.

“Under the dual pressures of soaring power densities and gigawatt-scale expansion, 2026 will see data center power and cooling systems fundamentally redesigned to bridge the gap to gigawatt campuses,” said Omdia analyst Shen Wang in the report.

With this in mind, I’ve outlined some of the top trends impacting AI data center power and cooling.

1. Liquid cooling becomes the AI standard

Nvidia’s Rubin platform is designed with 100% liquid cooling in mind. The company is eliminating all fans and expanding liquid coverage to every key component. The pipes and cold plates needed for liquid cooling are being extended from GPUs to CPUs, switches, and even the optical modules used in networking.

This full-stack approach to liquid cooling has led the chip maker to focus on partnerships that can deliver integrated, high-performance liquid cooling systems rather than dealing with many different component manufacturers.

“Advanced AI infrastructure requires 100% liquid cooling,” said Wang. “Data centers are advised to transition critical, high-density areas to liquid cooling for better thermal management and energy savings while maintaining cost-effective air cooling in less demanding zones.”

2. Battery Storage

Battery energy storage systems (BESS) have graduated from being a nice-to-have to a necessity for AI data centers. This is largely because massive GPU clusters experience millisecond-level power surges (which can occur many times per second). Traditional uninterruptible power supply (UPS) systems and backup generators cannot handle this level of variability.

But BESS alone isn’t enough. Wang noted that supercapacitors are often included in data center blueprints to ensure the facility can handle extreme power swings, remain balanced, maintain stable voltage, and be a responsible grid citizen.

Advanced power electronics has become a breeding ground for innovation. New systems for AI data centers are being advanced by vendors such as Dimaag, ON.energy, Ramboll, Eaton, and Vertiv.

“Traditional data centers handle steady CPU workloads, whereas AI data centers driven by large GPU clusters experience millisecond power surges that require large-scale on-site energy, and they also must be able to offer services like peak shaving and frequency regulation,” said Wang.

More about data centers

3. High Voltage DC

High-voltage direct current (HVDC) is another trend rapidly becoming a necessity for AI to function as intended. By instituting a 800V DC power backbone, several challenges related to high-density racks can be overcome. They include high copper costs and the large amount of space that would otherwise be taken up by traditional AC cabling.

In addition, energy losses would be minimized by switching from AC to DC. Complex infrastructure, such as rectifiers and other equipment needed to convert current from AC to DC and back again, could also be eliminated.

That said, 800V DC raises severe safety concerns. This amount of voltage could fry someone. Wang believes the transition will be gradual. The market will gradually move from traditional AC systems to 400V and then 800V DC configurations over the next few years. Hyperscalers will be leading the charge. Expect to see HVDC prototypes being tested and a few initial deployments this year and some examples of scaled deployment in 2027.

Wang added that this switch from AC to DC will be facilitated by the introduction of solid-state transformers (SSTs), which offer the potential for higher efficiency, smaller footprints, and much lower weight.

“The rapid industry shift toward HVDC architectures creates the high-voltage foundation required for SST adoption,” said Wang. “We expect SST proof of concepts (POCs) to kick off in 2026–27 across hyperscale cloud service providers, with the rest of the cloud and other market segments following suit.”

Taken together, the trends point to a data center market where AI performance depends as much on electrical and thermal engineering as it does on chips. Operators that cannot modernize power delivery and cooling may find themselves constrained long before they run out of demand.

Also read: AI data centers could create new grid stability risks in Australia as compute demand rises across APAC.