Liquid cooling enters the mainstream in data centers

Soaring artificial intelligence workloads are pushing traditional air-cooling systems to the brink. To combat this challenge, both chipmakers and data center operators are turning to a more efficient solution: liquid cooling.

Data center operator Equinix led the charge back in 2023, deploying liquid cooling in 100+ data centers across 45 cities. Similarly, Digital Realty launched a high-density colocation offering powered by liquid cooling, handling workloads of up to 70 kilowatts (kW) per rack.

Chip giant Nvidia is also on board, with its next-gen servers specifically designed for liquid cooling to manage the heat generated by their central processing units (CPUs) and graphics processing units (GPUs).

With rack densities set to grow beyond 70kW, the only viable solution for cooling these high-performance servers currently is liquid cooling — primarily direct-to-chip or immersion cooling, according to Andrew Green, Regional Data Center Practice Lead, Work Dynamics, JLL.

“We’ve reached a point in the development of data centers where rack densities have gone beyond what’s possible with the physics of air cooling,” says Green. “Liquid cooling is an inevitable development to cope with increased AI workloads.”

For data center operators under pressure to improve their power usage effectiveness (PUE), the power efficiency gains achieved by liquid cooling offer a significant advantage. Less energy is required to cool the refrigerant, which translates directly to substantial cost savings.

Freeing up floor space for additional data hall capacity

“A liquid cooling installation eliminates the need for mechanical equipment for air cooling, which frees up valuable floor space to be converted into additional data hall capacity,” says Green.

Another advantage, in the case of an immersion bath cooling setup, is the lower floor-to-floor height requirement of four meters, similar to a grade A office space.

The structural loading trade-off

This major reduction, compared to the standard six meters needed for most existing data centers, comes with the trade-off of needing an increased structural floor loading. “If the same amount of compute is installed within the same floor area, the weight per square meter increases from 12 to 15 kilopascals (kPa) to at least 20 kPa in a liquid-cooling installation,” Green explains.

Despite the clear benefits of liquid cooling, a pivot in cooling technology is a major undertaking for operators, potentially requiring an overhaul of the existing infrastructure and design. The shift comes as AI creates unprecedented demand that traditional infrastructure struggles to accommodate.

One of the key design changes involves plumbing installations. “Traditionally, we’ve put a lot of effort into keeping water out of the data halls, be it pre-action sprinklers or water detection systems,” says Green.

But in direct-to-chip cooling, for instance, small-bore pipes are required to deliver the coolant directly to the chips for heat removal.

This reality means new, purpose-built AI data centers are better positioned to adopt liquid cooling more readily. While existing data centers with spare capacity can be partially upgraded, it’s unlikely that an existing facility can be upgraded to 100% AI due to infrastructure and power constraints.

Managing risk during live data center upgrades

“Any major change in a live data center environment is a time of high risk and has to be carefully managed,” says Green. “Operators need to take into account how, from a resiliency standpoint, those works may result in service disruptions in other data halls.”

The data center industry generally aims to be leading-edge, not bleeding edge, Green says. “Operators are often slow to adopt change and typically favor proven solutions.”

But advancements in AI and the wider technology industry are forcing data centers to adapt.

“Data center companies are figuring how best to do AI deployment, be it in smaller, older buildings, or in the design of their future facilities,” says Green. However, operators must balance innovation with the risk of premature obsolescence as technology evolves rapidly.

^{The content was originally published March 18, 2024 and updated in September 2026.}