Many information facilities wrestle to get sufficient energy into the ability to run AI purposes. Those who handle it face an unlimited cooling drawback — how precisely do you dissipate that a lot warmth? Therefore, each of those areas are entrance and heart in information heart planning and design.
Analyst firm Omdia lately delved into the traits shaping each areas in a report on AI information facilities.
“Below the twin pressures of hovering energy densities and gigawatt-scale growth, 2026 will see information heart energy and cooling programs essentially redesigned to bridge the hole to gigawatt campuses,” mentioned Omdia analyst Shen Wang within the report.
With this in thoughts, I’ve outlined a few of the high traits impacting AI information heart energy and cooling.
1. Liquid cooling turns into the AI customary
Nvidia’s Rubin platform is designed with 100% liquid cooling in thoughts. The corporate is eliminating all followers and increasing liquid protection to each key element. The pipes and chilly plates wanted for liquid cooling are being prolonged from GPUs to CPUs, switches, and even the optical modules utilized in networking.
This full-stack method to liquid cooling has led the chip maker to concentrate on partnerships that may ship built-in, high-performance liquid cooling programs quite than coping with many alternative element producers.
“Superior AI infrastructure requires 100% liquid cooling,” mentioned Wang. “Information facilities are suggested to transition essential, high-density areas to liquid cooling for higher thermal administration and vitality financial savings whereas sustaining cost-effective air cooling in much less demanding zones.”
2. Battery Storage
Battery vitality storage programs (BESS) have graduated from being a nice-to-have to a necessity for AI information facilities. That is largely as a result of large GPU clusters expertise millisecond-level energy surges (which might happen many instances per second). Conventional uninterruptible energy provide (UPS) programs and backup turbines can not deal with this stage of variability.
However BESS alone isn’t sufficient. Wang famous that supercapacitors are sometimes included in information heart blueprints to make sure the ability can deal with excessive energy swings, stay balanced, preserve secure voltage, and be a accountable grid citizen.
Superior energy electronics has turn out to be a breeding floor for innovation. New programs for AI information facilities are being superior by distributors akin to Dimaag, ON.energy, Ramboll, Eaton, and Vertiv.
“Conventional information facilities deal with regular CPU workloads, whereas AI information facilities pushed by giant GPU clusters expertise millisecond energy surges that require large-scale on-site vitality, and so they additionally should have the ability to provide companies like peak shaving and frequency regulation,” mentioned Wang.
Extra about information facilities
3. Excessive Voltage DC
Excessive-voltage direct present (HVDC) is one other development quickly turning into a necessity for AI to perform as supposed. By instituting a 800V DC energy spine, a number of challenges associated to high-density racks will be overcome. They embrace excessive copper prices and the big quantity of area that might in any other case be taken up by conventional AC cabling.
As well as, vitality losses could be minimized by switching from AC to DC. Complicated infrastructure, akin to rectifiers and different tools wanted to transform present from AC to DC and again once more, is also eradicated.
That mentioned, 800V DC raises extreme security issues. This quantity of voltage might fry somebody. Wang believes the transition can be gradual. The market will progressively transfer from conventional AC programs to 400V after which 800V DC configurations over the following few years. Hyperscalers can be main the cost. Anticipate to see HVDC prototypes being examined and some preliminary deployments this 12 months and a few examples of scaled deployment in 2027.
Wang added that this swap from AC to DC can be facilitated by the introduction of solid-state transformers (SSTs), which supply the potential for greater effectivity, smaller footprints, and far decrease weight.
“The fast business shift towards HVDC architectures creates the high-voltage basis required for SST adoption,” mentioned Wang. “We count on SST proof of ideas (POCs) to kick off in 2026–27 throughout hyperscale cloud service suppliers, with the remainder of the cloud and different market segments following go well with.”
Taken collectively, the traits level to an information heart market the place AI efficiency relies upon as a lot on electrical and thermal engineering because it does on chips. Operators that can’t modernize energy supply and cooling could discover themselves constrained lengthy earlier than they run out of demand.
Additionally learn: AI data centers could create new grid stability risks in Australia as compute demand rises throughout APAC.