The Turning Point in Data Center Cooling: From Air to Liquid
AI workload expansion is demanding a fundamental redesign of data center infrastructure. At the center of that shift is cooling.
Why Cooling, Why Now
The core issue is power density. In 2017, a standard server rack consumed around 15kW. In 2025, the NVIDIA GB200 NVL72 rack draws 132kW. The Rubin Ultra NVL576, expected in 2027, will reach 600kW — a 40x increase in a single decade.
|
Year |
Representative Hardware |
Power per Rack |
|
2017 |
Standard Server |
~15 kW |
|
2025 |
NVIDIA GB200 NVL72 |
132 kW |
|
2027 (est.) |
Rubin Ultra NVL576 |
600 kW |
Traditional air cooling maxes out at roughly 20–25kW per rack. AI server deployments have already blown past that ceiling. Cooling is no longer a facilities afterthought — it is now the primary variable determining whether high-density workloads can be supported at all.
The global data center cooling market is projected to grow from $18.8B in 2025 to $54.2B by 2034 (CAGR 12.6%), with the liquid cooling segment outpacing the overall market at 18.2% CAGR over the same period.
Three-Stage Evolution of Cooling Technology
① Direct Liquid Cooling (DLC) — Today’s Standard
Cooling plates are mounted directly onto CPUs and GPUs, with coolant circulated through them to extract heat at the source. Fans remain in operation, making this a hybrid architecture that is relatively straightforward to retrofit into existing facilities.
The central component is the CDU (Coolant Distribution Unit), which manages coolant flow and pressure across the data center. Leading CDUs now handle upward of 80kW per rack, with rear-door heat exchangers (RDHx) enabling parallel deployment alongside existing air cooling infrastructure.
DLC is the dominant cooling method for AI server environments in the 2025–2026 cycle.
② Immersion Cooling — Fast-Growing Segment
Servers are fully submerged in dielectric fluid — an electrically non-conductive coolant. Two configurations exist:
• Single-Phase: Fluid absorbs heat and circulates through an external heat exchanger. Simpler to operate.
• Two-Phase: Fluid vaporizes upon contact with heat, then condenses and returns — leveraging the latent heat of vaporization for superior efficiency.
Immersion cooling can bring PUE down to 1.05–1.15, compared to 1.4–1.8 for conventional air-cooled facilities. With no fans required, noise and vibration are eliminated, and rack densities exceeding 100kW become manageable.
|
PUE (Power Usage Effectiveness) Total data center power consumption ÷ IT equipment power consumption. A value of 1.0 is ideal; lower figures indicate less energy wasted on cooling, lighting, and other overhead. |
The immersion cooling market was valued at approximately $1.7B in 2025, with a projected CAGR of 19.8% through 2035.
③ Microfluidic Cooling — The Next Decade
Microchannels are etched directly into the silicon die, allowing coolant to flow through the chip itself. Theoretically the highest-efficiency cooling method possible, as heat is intercepted at the point of generation.
EPFL spinout Corintis leads this space, applying AI-driven bio-inspired channel design (modeled on leaf venation and butterfly wing structures). By 2025, the company had raised a cumulative $58M and reached a ~$400M valuation, with Intel CEO Lip-Bu Tan joining its board.
Commercial-scale deployment is projected for the 2030s. Key hurdles remain: dedicated silicon design requirements, microchannel clogging risk, and manufacturing complexity.
Waste Heat Recovery
Liquid cooling creates a practical pathway for capturing and repurposing the heat extracted from servers. Real-world deployments include:
• Finland: Google’s data center feeds waste heat into the local district heating network.
• Denmark: Meta’s facility delivers 165,000 MWh of recovered heat annually, supplying approximately 11,000 households.
• South Korea: Korea District Heating Corporation is actively developing a waste heat reuse program linked to colocation operations.
On the regulatory side, the EU’s revised Energy Efficiency Directive (EED) and Germany’s Energy Efficiency Act (EnEfG) will effectively mandate PUE ≤1.3 and waste heat recovery for new data centers from 2027.
Water Usage Efficiency (WUE)
The rapid expansion of AI data centers has brought evaporative cooling water consumption under scrutiny. Microsoft introduced a closed-loop, zero-evaporation cooling system in December 2025. Google has committed to eliminating evaporative water use in all new facilities by end of 2027 — a transition that reduces water consumption by 125 million liters per facility annually relative to evaporative designs.
Market Outlook
|
Segment |
2025 |
2035 |
CAGR |
|
Data Center Cooling (Total) |
$18.8B |
$54.2B |
12.6% |
|
Liquid Cooling |
$4.8B |
$27.1B |
18.2% |
|
Immersion Cooling |
$1.7B |
~$10B |
19.8% |
Discover More.
Join us as we take you through every step of the development of our state-of-the-art hyperscale data center in this blog. We loves to talk about what we do. Feel free to chat with our sales experts about data center, managed colocation, and connectivity.
Content is protected by copyright law and is owned by Daou Technology Inc
It is prohibited to modify or commercially use this content without prior consent.
Featured images via gettyimages.
References
1. Korea Data Economy Newspaper, GM Insights, Fortune Business Insights, TierPoint, IoT Analytics (2025–2026)