AI Pushing Data Centers to Liquid Cooling

Why it matters

The power demands of new AI hardware are driving the shift, with NVIDIA's H100 GPU consuming 700W and the newer B200 requiring 1000W. A single rack of NVIDIA's GB200 servers now calls for approximately 140 kW of liquid cooling capacity, far exceeding the capabilities of legacy air-cooled systems. The data center liquid cooling market is forecast to expand from $4.8 billion in 2025 to $27.1 billion by 2035. This growth is a direct response to AI workloads that are pushing average rack densities from 15 kW to over 80 kW. While only 19% of data centers have currently implemented liquid cooling, 36% are planning to adopt it within the next two years. Two primary technologies are emerging: direct-to-chip cooling, where cold plates are mounted directly on processors, and full immersion cooling, which submerges entire servers in a non-conductive fluid. Microsoft is actively deploying direct-to-chip cold plate technology, which it found can reduce greenhouse gas emissions by around 15% and water consumption by 30% to 50% across a data center's lifecycle compared to air cooling. While upfront costs for liquid cooling can be higher, long-term Total Cost of Ownership (TCO) is often lower due to significant energy savings. One analysis of a 10 MW AI data center projected that using immersion cooling could reduce the 10-year TCO by 39%, a saving of over $110 million compared to air cooling. At higher densities, liquid cooling can even reduce initial capital expenditures by 10-14%. Cooling infrastructure accounts for 30-40% of a data center's total energy consumption. By improving Power Usage Effectiveness (PUE), liquid cooling directly addresses this overhead. Optimized liquid-cooled facilities can achieve a PUE of around 1.15, compared to 1.3-1.5 for conventional air-cooled sites. Major tech firms like Google and Amazon Web Services are investing heavily in liquid cooling research and development. This has spurred a specialized market of providers, including Vertiv, Schneider Electric, and CoolIT Systems. In a sign of market consolidation, power management company Eaton recently acquired liquid cooling specialist Boyd Thermal for $9.5 billion.

Key numbers

• A recent analysis finds that traditional air cooling hits a hard ceiling at approximately 41.3kW per rack.
• The power demands of new AI hardware are driving the shift, with NVIDIA's H100 GPU consuming 700W and the newer B200 requiring 1000W.
• A single rack of NVIDIA's GB200 servers now calls for approximately 140 kW of liquid cooling capacity, far exceeding the capabilities of legacy air-cooled systems.
• The data center liquid cooling market is forecast to expand from $4.8 billion in 2025 to $27.1 billion by 2035.

What happens next

• The data center liquid cooling market is forecast to expand from $4.8 billion in 2025 to $27.1 billion by 2035.
• While only 19% of data centers have currently implemented liquid cooling, 36% are planning to adopt it within the next two years.
• One analysis of a 10 MW AI data center projected that using immersion cooling could reduce the 10-year TCO by 39%, a saving of over $110 million compared to air cooling.

Sources

• analysis finds
• The power demands of
• A single rack of NVIDIA's
• This growth is a direct
• While only 19% of data
• Two primary technologies
• Microsoft is actively
• While upfront costs for
• One analysis of a 10
• At higher densities,
• Cooling infrastructure
• Optimized liquid-cooled
• Major tech firms like
• In a sign of market consolidation

Quick answers