Related Posts

Top 5 This Week

The race to cool AI: How modern data centers keep from overheating

Not all data centers are cooled the same way. In fact, the biggest shift happening in AI infrastructure right now may not be the chips themselves, but how operators prevent them from melting down.

Traditional cloud data centers were designed around rows of servers cooled with circulating air. But AI processors generate exponentially more heat than conventional computing equipment, pushing many facilities beyond what standard air conditioning systems can handle. This has triggered a race to redesign cooling systems that are faster, cheaper, more energy efficient and less water intensive.

Hereโ€™s a look at data center cooling through the ages. 

1. Traditional air cooling

The original data center model

This is the classic method that powered the internet boom of the 1990s and 2000s. Early generations of Google data centers built in the early 2000s relied heavily on this design, and it is still commonly used for traditional cloud computing, storage and business applications.

Cold air is pumped into a room through large HVAC-style systems called CRAC (Computer Room Air Conditioner) or CRAH (Computer Room Air Handler) units. Servers are arranged in โ€œhot aislesโ€ and โ€œcold aisles,โ€ where cool air enters from the front of racks and hot exhaust exits through the back. Raised floors often distribute chilled air underneath equipment.

Pros

  • Cheapest and most familiar system
  • Easier maintenance because servers remain dry
  • Proven technology with decades of operational history
  • Lower upfront construction costs

Cons

  • Struggles with high-density AI computing
  • Requires enormous airflow and electricity
  • Cooling becomes inefficient as racks get hotter
  • Larger physical footprint

2. Chilled water cooling

A more powerful version of air cooling

Instead of refrigerant-based air conditioning, chilled water circulates through heat exchangers that cool air before it reaches servers.

These systems are more efficient than conventional AC because water transfers heat more effectively than air. Many large-scale hyperscale data centers adopted chilled water systems during the 2010s. Several early hyperscale campuses from Google Cloud and Amazon Web Services used chilled water plants during the expansion of cloud computing in the 2010s. This model is still used for many cloud campuses with heavy workloads.

Pros

  • More efficient than conventional HVAC
  • Better for larger facilities
  • Can support denser computing loads

Cons

  • Higher construction cost
  • Often requires substantial water infrastructure
  • More plumbing and mechanical complexity

3. Evaporative cooling

The water-saving system that sometimes uses more water

Evaporative systems cool air using water evaporation, similar to how sweat cools skin. Warm air passes over water, and evaporation lowers the temperature before the air re-enters the building. This can dramatically reduce electricity consumption compared to refrigeration-based cooling. It is often called adiabatic cooling or swamp cooling. Many Western U.S. facilities from major hyperscalers used evaporative cooling because of hot, dry climates. The Google Mesa Data Center in Mesa, Arizona became one of the most discussed examples of data center water use in the desert Southwest because it planned to use evaporative cooling, which can significantly reduce electricity consumption compared to conventional refrigeration systems. Google reportedly secured water allocations tied to cooling operations, with local reporting indicating allowances of up to millions of gallons per day depending on project phases and demand. That sparked public concern over whether large industrial users should expand in drought-prone regions Critics raised concerns because Mesa relies heavily on water supplies tied to the stressed Colorado River system

The concern is not hypothetical. A major Georgia project operated by QTS drew scrutiny after millions of gallons of water were consumed during development amid drought conditions, raising questions about infrastructure strain and transparency. The company said its future operations would rely on a closed-loop system with minimal cooling water demand.

Many of the data centers have pledged to offset their water use. Amazon, for example, does not provide overall figures for the water its datacentres use worldwide. But it does claim that it will be โ€œwater positiveโ€ by 2030, offsetting its consumption by providing water to communities and ecosystems in areas of scarcity elsewhere. 

Pros

  • Highly energy efficient
  • Lower electricity bills
  • Effective in dry climates

Cons

  • Can consume enormous amounts of water
  • Less effective in humid climates like New Jersey
  • Community concern and water permitting challenges

4. Direct-to-chip liquid cooling

The AI-era favorite

Instead of cooling the room, liquid is piped directly to the hottest components โ€” usually CPUs and GPUs.

Tiny tubes carry coolant to processors, removing heat at the source. This is becoming one of the dominant approaches for AI facilities because modern GPUs generate extraordinary amounts of heat. Microsoft announced a new chip-level liquid cooling architecture for AI data centers in 2024โ€“2025 as part of a zero-water strategy. Amazon is redesigning next-generation AI facilities with liquid cooling to support high-power NVIDIA systems.

Pros

  • Far more efficient than air cooling
  • Supports much denser AI racks
  • Lower cooling energy demand
  • Reduces overall facility footprint

Cons

  • Expensive retrofits
  • More engineering complexity
  • Requires leak prevention systems
  • High upfront cost

5. Closed-loop cooling systems

The technology everyone is suddenly talking about

Closed-loop systems recycle the same cooling water repeatedly instead of constantly consuming new water. Rather than evaporating water through cooling towers, heat moves through sealed loops and heat exchangers. Some systems use outside air or dry coolers to reject heat without major water loss. The goal is to reduce freshwater demand. 

CoreWeave’s new data center in Kenilworth, NJ is using a closed-loop cooling system designed to require very little water during normal operations. According to project materials, the system will need approximately 310,000 gallons of water one time during construction. After startup, the cooling loop will be filled with a pre-mixed propylene glycol solution that continuously circulates through the system. 

Pros

  • Major reduction in water consumption
  • Less stress on municipal water systems
  • Better public acceptance
  • Easier permitting in water-sensitive areas

Cons

  • Higher upfront capital costs
  • More complex mechanical systems
  • May still require occasional replacement water
  • Less effective in some extreme heat scenarios

6. Immersion cooling

The futuristic option: dunking servers in liquid

Servers are submerged in a special nonconductive liquid that absorbs heat directly. Instead of cooling air, the liquid itself pulls heat away from electronics. Think of it as a giant mineral-oil bath for computers. There are two main approaches. Single-phase immersion cooling submerges servers in a non-conductive dielectric fluid that remains in liquid form throughout the cooling process. As the hardware generates heat, the surrounding fluid absorbs it and is circulated through a heat exchanger, where the heat is removed before the cooled fluid is returned to the tank. Because the fluid never changes state, single-phase systems are relatively simple to operate and maintain, making them the most common form of immersion cooling in commercial deployments.

Two-phase immersion cooling uses a specialized dielectric fluid with a low boiling point. When hot processors and other components heat the fluid, it boils and turns into vapor. The vapor rises to a cooled condenser above the servers, where it condenses back into liquid and falls into the tank to repeat the cycle. This phase-change process removes heat very efficiently and can support extremely high-density computing environments, but the systems are more complex and typically carry higher installation and operating costs than single-phase designs.

High-performance computing and crypto-mining operations pioneered immersion cooling. It is currently being used by Bitcoin miners in Texas, Wyoming, and Alberta. It is also used by the University of Teaxasโ€™ TACC Frontera, one of the worldโ€™s fastest academic supercomputers.

  Researchers say immersion systems can cut cooling energy dramatically while shrinking building size requirements. Major AI operators are testing it for next-generation GPU clusters.

Pros

  • Extremely energy efficient
  • Handles ultra-dense AI computing
  • Smaller building footprint
  • Potentially lower operating costs

Cons

  • Maintenance is harder
  • Specialized equipment required
  • Retrofitting older centers is expensive
  • Limited operating history

The next challenge for data centers 

As AI drives demand for ever-larger and more powerful data centers, cooling technology is becoming as important as the processors inside the buildings. The industry is rapidly moving beyond the air-cooled designs that powered the internet era toward systems that prioritize efficiency, water conservation and higher computing density. Whether through direct-to-chip liquid cooling, closed-loop systems or even immersion tanks, operators are searching for ways to support the next generation of AI while addressing growing concerns about energy use, water consumption and community impact. The result is a new arms race in infrastructure, one where the future of computing may depend as much on how facilities remove heat as on how fast their chips can process information.

Sources:

Amazon Web Services. AWS rolls out liquid cooling in data centers. About Amazon. https://www.aboutamazon.com/news/aws/aws-liquid-cooling-data-centers

Amazon Web Services. AWS announces new data center components to support AI innovation and further improve energy efficiency. Business Wire. https://www.businesswire.com/news/home/20241201928268/en/AWS-Announces-New-Data-Center-Components-to-Support-AI-Innovation-and-Further-Improve-Energy-Efficiency

LiquidStack. (n.d.). Immersion cooling technology overview. https://liquidstack.com

Microsoft. Sustainable by design: Next-generation datacenters consume zero water for cooling. Microsoft Cloud Blog. https://www.microsoft.com/en-us/microsoft-cloud/blog/2024/12/09/sustainable-by-design-next-generation-datacenters-consume-zero-water-for-cooling/

Data Center Dynamics. Microsoft’s upcoming data centers to use closed-loop, zero-water evaporation design. https://www.datacenterdynamics.com/en/news/microsofts-upcoming-data-centers-to-use-closed-loop-zero-water-evaporation-design/

TechCrunch. AWS bets on liquid cooling for its AI servers. https://techcrunch.com/2024/12/02/aws-bets-on-liquid-cooling-for-its-ai-servers/

Texas Advanced Computing Center. Frontera supercomputer. https://www.tacc.utexas.edu/systems/fronteraWired. Data center operators are trying to fix their water use problems. https://www.wired.com/story/data-center-operators-fix-water-use-problems

Sign up and get the latest infrastructure news in your inbox every week.

We donโ€™t spam! Read our privacy policy for more info.

Sign up and get the latest infrastructure news in your inbox every week.

We donโ€™t spam! Read our privacy policy for more info.

Popular Articles