The Evolution of Thermal Management Technologies

Air Cooling vs. Liquid Cooling: How Data Center Cooling Evolved for AI

 

Liquid cooling is often presented as a new solution for AI data centers. However, it has been used in high-performance computers for decades. Early water-cooled mainframes and immersion-cooled supercomputers were developed when their heat output exceeded what air cooling could effectively handle.

 

Air cooling later became the standard for most commercial data centers because it supported widely used server designs and familiar maintenance practices. Liquid cooling remained important in specialized high-performance systems.

 

The rapid growth of AI computing has brought liquid cooling back into wider use. Today, air cooling, immersion cooling, and direct-to-chip liquid cooling continue to serve different requirements. At the same time, rising chip heat loads are driving further development in both single-phase and two-phase cooling.

 

This article traces the history of data center cooling and explains how earlier technologies have shaped the cooling systems used for today’s AI infrastructure.

 

1960s to 2000s: Air Cooling Becomes the Mainstream Approach

 

The modern data center developed from the computer rooms used to house mainframe systems in the 1960s. Over the following decades, fans, heat sinks, and room air-conditioning systems became the most common way to remove heat from computing equipment.

 

Cost and ease of operation contributed to this adoption. For example, the air-cooled Amdahl 470V/6, introduced in the 1970s, used finned cooling towers and directed airflow over its processors. Its air-cooled design was less expensive than the water-cooled assemblies used in competing IBM systems.

 

As enterprise computing and standard rack servers expanded through the 1990s and 2000s, air cooling remained suitable for most data centers. However, liquid cooling was already being used in specialized computers with higher cooling requirements.

 

1960s to 1980s: Liquid Cooling Enters High-Performance Computing

 

Liquid cooling was already being developed for systems that generated more concentrated heat than air cooling could manage effectively. Cold plates had been used in aerospace and military electronics since the 1960s. By the 1980s, IBM was using liquid-cooled cold plates to remove heat from high-power processor modules.

 

Immersion cooling also appeared during this period. The Cray-2 supercomputer, released in 1985, immersed densely packed circuit boards in Fluorinert, a nonconductive cooling fluid. The heated fluid circulated through an external cooling system before returning to the computer.

 

These examples show that cold-plate and immersion cooling developed as separate approaches during the same period. Liquid cooling was therefore not a single stage that followed air cooling. It was already being used selectively wherever computing density created greater cooling demands.

 

2010s: Energy Efficiency Renewed Interest in Liquid Cooling

 

As data centers grew, cooling energy became a larger operational concern. Liquid cooling began attracting renewed attention because it could collect heat closer to high-power components and carry it away at a higher temperature than many air-cooling systems.

 

IBM’s Aquasar supercomputer, introduced in 2010, demonstrated this approach using warm-water cooling. Water entered the chip-level cooling system at up to approximately 60ºC. Because the processors operated at a higher temperature, heat still moved from chips into the water. The heated water left the systems at up to approximately 65ºC and could then be used for building heating. This design removed the need to produce chilled water before cooling the processors.

 

Aquasar showed that liquid cooling could do more than support high-power processors. By carrying heat away at a higher temperature, it could reduce cooling energy and make the recovered heat easier to reuse. These goals became increasingly relevant as data centers prepared for higher computing densities.

 

2020s: AI Brings Liquid Cooling into Rack-Scale Systems

 

During the 2020s, growing AI and HPC workloads increased the amount of heat generated within individual servers and racks. Liquid cooling began moving beyond specialized supercomputers and into a wider range of commercial data center systems.

 

Cold-plate and immersion cooling developed alongside one another during this period. The Open Compute Project now organizes cold plates, immersion systems, door heat exchangers, coolant distribution units, and heat reuse under its Cooling Environment Projects. This reflects an industry in which several cooling approaches are advancing alongside one another.

 

The shift became more visible with rack-scale AI platforms. NVIDIA introduced the GB200 NVL72 in 2024 as a rack-scale, liquid-cooled system connecting 72 Blackwell GPUs and 36 Grace CPUs. The significance of the GB200 NVL72 is not that it introduced liquid cooling for the first time. It’s that this product shows how a technology once reserved mainly for specialized high-performance computers is becoming part of commercially available AI infrastructure.

 

How Does CreAlien W2PC Continue the Development of Liquid Cooling

 

The development of liquid cooling now extends beyond bringing coolant closer to AI chips. It also involves improving how the coolant absorbs increasingly concentrated heat. CreAlien W2PC addresses this challenge through two-phase direct-to-chip cooling.

 

W2PC uses DI water in a sealed, negative-pressure loop. The lower pressure allows the water to begin boiling at approximately 45ºC. As the DI water passes through cold plates attached to CPUs and GPUs, part of it vaporizes and absorbs heat through phase change. The vapor then moves to the condensation side, releases the heat, and returns to liquid.

 

In a single-phase system, a higher heat load may require a higher coolant flow rate or a greater temperature rise. In W2PC, more DI water can vaporize to absorb the additional heat. This reduces the system’s dependence on increasing flow rate or coolant temperature rise.

 

The approximately 45ºC operating temperature also allows the collected heat to be rejected without first producing low-temperature chilled water. Under suitable site and outdoor conditions, this can support chiller-less cooling.

 

W2PC shows how an established two-phase cooling principle can be adapted to direct-to-chip cooling for current and future AI systems. Contact CreAlien to discuss the cooling requirements of your AI infrastructure.