Liquid Cooling: The Non-Negotiable Future of Hyperscale AI
- Sanjay Kr Singh
- Nov 13, 2025
- 9 min read
Updated: Nov 18, 2025
About: Beyond the Fan: Why Liquid Cooling is the Future of Hyperscale Infrastructure
1) Sustainability & PUE: How Liquid Cooling Radically Changes the Power Usage Effectiveness (PUE) Metric and its Environmental Impact
The relentless demand for high-performance computing (HPC), AI, and machine learning is pushing modern data centers to their thermal breaking point. Traditional air-cooling systems, the industry’s decades-long workhorse, are now struggling to manage the immense heat density generated by next-generation CPUs and GPUs. The future of hyperscale infrastructure demands a paradigm shift, moving beyond the fan and into a more efficient, quieter, and sustainable solution. Liquid cooling is emerging as the critical technology, offering vastly superior thermal management capabilities that are essential for unlocking the next wave of computing power. This transition is although keeping servers cool, but also improving energy efficiency in the form of better Power Usage Effectiveness (PUE) metric and ensuring the long-term viability of the digital world.
Industry analysis shows air cooling is nearing its limits under AI/HPC loads major hyperscalers are adopting liquid systems, and liquid cooling delivers up to 1,000× better heat transfer, lower PUE and waste-heat reuse for sustainability.
a) What is PUE?
PUE is a ratio that measures the total energy entering a data center against the energy actually used by the IT equipment.
PUE = [Total Energy fed into data centre facility] / [Total energy used by IT Equipment]
A perfect PUE of 1.0 would mean that all energy is directly utilised for computing, with no loss to cooling, lighting, or other overheads. Traditional air-cooled data centers often struggle to achieve PUEs much lower than 1.5-1.8, which means 50-80% more energy is fed into facility to support the IT load.
b) How Liquid Cooling Transforms PUE (Power Usage Effectiveness):
Elimination/Reduction of CRAC (Computer Room air Conditioners)- Liquid cooling targets heat directly, requiring much less air movement in the facility.
Exploitation of Free Cooling: Liquid coolants can operate efficiently at much higher temperatures (e.g., 40-50°C) than air systems.
Superior Heat Transfer Efficiency: Liquid's thermal conductivity is up to 25 times greater than air.
Drastic Reduction in Fan Power: Immersion cooling eliminates all server and cabinet fans.
Higher Heat Density in Less Space: Liquid cooling allows for higher server density in racks, meaning a smaller overall facility footprint.
These transformations are the main reasons why liquid-cooled data centers can consistently achieve PUEs as low as 1.05–1.2, a massive leap in efficiency compared to the typical PUE 1.5–1.8 of older air-cooled facilities.
c) Environmental Impact:
The direct consequence of a lower PUE is a substantial reduction in energy consumption and, by extension, carbon emissions. A data center operating at a PUE of 1.2 instead of 1.6 can cut its total energy overhead by half! This translates to:
a) Decreased Operating Costs: Lower energy bills for the facility.
b) Reduced Carbon Footprint: Less electricity demand means lower greenhouse gas emissions, especially when drawing from fossil fuel-based grids.
c) Water Conservation (Paradoxically): While liquid cooling uses liquid, its efficiency often allows for less overall water consumption for evaporative cooling towers (if used) compared to air-cooled facilities trying to achieve similar thermal performance in hot climates.
By improving PUE, liquid cooling doesn't just enable higher performance; it fundamentally green lights the growth of hyperscale infrastructure in an environmentally responsible way.
2) Liquid Cooling vs. Air Cooling: Which One Will Dominate Data Centers in 2025-26?
a) The Thermal Imperative: AI Workloads and the End of Air-Cooled Density
The emergence of Artificial Intelligence (AI) and High-Performance Computing (HPC) has fundamentally changed the thermal landscape of modern data centers. The rapid advancement of AI, driven particularly by complex, transformer-based Generative AI architectures, necessitates increasingly powerful computing resources. Consequently, cooling technology exploration has become critical strategic bottleneck for AI deployment.
b) Current Trends: AI as the Catalyst for Cooling Transformation
The demand for high-density computing has resulted in surge in rack densities in data center. Average rack power densities are now projected to reach or exceed 50 kW by 2027. This far surpasses the historical effective limit of traditional air cooling, which typically reaches 10–15 kW per rack. As powerful GPUs and CPUs—the core components for machine learning and AI applications—generate increasingly intense heat, traditional air-cooling methods are simply exhausting their physical and economic limits.
Liquid cooling is transitioning from a niche solution to an essential infrastructure requirement. This adoption is driven by its inherent thermodynamic principle-water can absorb heat 1,000 times more effectively than air. Because of this efficiency gain, liquid cooling is expected to dominate new hyperscale data centres by 2026. While the upfront investment for this specialized infrastructure can be higher, liquid cooling solutions pay for themselves over time through reduced power consumption, lower operational costs, and longer hardware lifespan due to lower thermal stress. Moreover the heat emitted out in liquid cooling is reused like feeding to Swimming pools for heating the water.
Although liquid cooling is paramount for high-density AI/HPC environments, air cooling will not disappear overnight, particularly in legacy facilities like network devices, low compute systems. For organizations in transition, hybrid cooling systems—which strategically blend air and liquid cooling—are emerging as the most practical solution for balancing immediate performance needs, efficiency gains, and cost management.
c) Shifting Bottlenecks and Long-Term Value Creation
Improvements in power distribution combined with the massive power demands of modern GPU architectures mean that a single server, stuffed with high-performance GPUs and two high-end CPUs, can consume most of a 10 kW Power Distribution Unit (PDU). As power capacity increases, the resulting heat generation becomes the new critical constraint—the bottleneck shifts decisively to thermal management.
Liquid cooling is a strong financial transition for maximizing long-term ROI. Although it has higher CAPEX for infrastructure (CDUs, plumbing) than air cooling, this is quickly offset by substantially lower OPEX. Reduced reliance on energy-intensive AC and chillers, combined with enhanced hardware longevity, drives operational savings, often supported by government incentives.
The architectural benefits of liquid cooling stem from its superior heat absorption, allowing for:
i. Elimination of large external cooling units, facility fans, and hot-aisle/cold-aisle separation.
ii. Modular, space-saving designs with a smaller physical footprint.
iii. At-need scalability in deployment.
a) Defining Power Density in Modern AI Racks
Conventional enterprise racks housing standard compute servers typically operate within a manageable power range of 5–10 kW, which is easily accommodated by traditional air-cooling systems. However, AI training or workload driven environments rapidly push these limits. The escalating thermal density of AI hardware makes liquid cooling an absolute necessity:
Current High-Density Racks: Medium-scale AI GPU servers (A100/H100) typically demand 15–30 kW per rack, already straining air cooling capacity.
Extreme-Density AI: Dedicated training configurations reaching 30–42 kW per rack mandates the use of modalities like Direct Liquid Cooling (DLC) or Immersion Cooling for stable operation.
Next-Generation Readiness: Future integrated platforms, exemplified by the NVIDIA GB200 NVL72, are designed to push demands up to 132 kW per rack, relying entirely on high-capacity liquid cooling systems.
The 42kW thermal density of high-performance AI racks routinely exceeds the 10–15kW limit of conventional air cooling. To maximize GPU density, all new AI deployments must budget for liquid cooling infrastructure from the outset.
Table 1: Power Density Requirements for 42U Data Center Racks (AI Focus)
Rack Type | Typical Power Range (kW) | Thermal Strategy Requirement | Supporting AI Workloads |
Standard Compute Servers | 5–10 kW | Air Cooling (Sufficient) | General ML, Data Processing |
High-Density GPU Servers (A100/H100) | 15–30 kW | Hybrid (Air/DLC Recommended) | Medium-Scale AI Training/Inference |
Extreme-Density AI Training Racks | 30–42 kW+ | Liquid Cooling (DLC or Immersion Essential) | Large-Scale Generative AI/HPC |
Future AI Architectures (e.g., GB200) | 100–132 kW+ | Dedicated Immersion or Advanced DLC | Hyperscale AI Factories |
The AI scale-up paradigm treats all rack GPUs as one massive coherent GPU, demanding higher power management and compute density. Achieving the required nanosecond-scale latency and coherence hinges on extreme and uniform thermal stability to eliminate localized hotspots. Because air cooling cannot meet this uniform standard, advanced liquid cooling is essential for supporting true rack-scale coherence.
3) Comparative Analysis: Air Cooling vs. Liquid Cooling Modalities
Data center cooling is broadly categorized into Air Cooling and Liquid Cooling.
Liquid Cooling is further divided into:
a) Direct Liquid Cooling (DLC)
b) Immersion Cooling (Single-Phase and Two-Phase)
The choice depends on three key factors: thermal density requirements, financial constraints, and long-term efficiency goals.
a) Direct Liquid Cooling (DLC): The Phased Approach
Direct Liquid Cooling (DLC), or direct-to-chip cooling, involves circulating a specialized coolant through cold plates directly on the highest-heat components (CPUs/GPUs).
a) Thermal Performance: Offers significantly better performance than air cooling.
b) Density Support: Supports high-density racks with thermal loads typically in the 40–100+ kW range.
c) Integration: Managed by Coolant Distribution Units (CDUs), DLC seamlessly integrates into existing server designs, minimizing operational disruption.
d) Efficiency: Generally, leads to improved energy consumption and lower operating costs compared to air-only cooling.
b) Immersion Cooling: Maximum Density and Efficiency
Immersion Cooling submerges the entire server in a dielectric (non-conductive) fluid for silent and uniform heat dissipation without internal fans.
Cooling Mechanisms
Type | Process | Key Feature |
Single-Phase (1-PIC) | Coolant remains liquid, absorbing heat and transferring it to an external heat exchanger. | Minimal evaporation; uses "open baths." |
Two-Phase (2-PIC) | Leverages a phase change: fluid boils on hot components, rises as vapor, condenses on a cold coil, and rains back down. | Provides superior heat transfer and eliminates temperature stratification. |
Efficiency & Performance
Immersion delivers maximum power capacity for extreme densities and effectively eliminates hot spots.
Energy Savings: Studies show up to 95% reduction in cooling energy use.
PUE: Achieves the lowest Power Usage Effectiveness (PUE) figures, often reaching 1.02–1.03.
4) Measuring True Cooling Efficiency: PUE vs. TUE
Relying solely on PUE (Power Usage Effectiveness) can mask liquid cooling benefits. While a hybrid liquid system may only minimally drop PUE (e.g., 1.38 to 1.34), it drastically reduces IT power consumption by cutting server fan power by up to 80%.
i. The proposed metric, Total Usage Effectiveness (TUE), which improved by 15.5% in the study, is a more comprehensive measure of liquid cooling's true efficiency.
Cooling Modality Comparison (Features & Strategy)
The choice between technologies is a trade-off in density, cost, and complexity:
Feature | Air Cooling | DLC (Direct Liquid) | 2-PIC (Two-Phase Immersion) |
Max Rack Density | 10–15 kW | 40–100+ kW | 80–150+ kW (Highest) |
Typical PUE | 1.30–1.60 | 1.20–1.35 | 1.02–1.05 (Lowest) |
Upfront CAPEX | Low | Moderate | Very High |
Primary Issue | Density limitation, High OPEX | Higher initial investment | High fluid cost, complex maintenance |
Tactical vs. Strategic Choice
i. DLC is the tactical, incremental choice for quick adoption, fitting existing server form factors.
ii. Immersion Cooling (especially 2-PIC) is the strategic architectural commitment for extreme densities (>100 kW).
a. 2-PIC offers superior reliability due to its passive, constant-temperature phase change, eliminating the temperature stratification risks associated with pump-reliant Single-Phase Immersion (1-PIC).
5) Strategic Recommendations for AI Data Center Thermal Strategy
The choice of cooling is dictated by facility type and required density:
🏭 New Hyperscale AI Factories (Density > 30 kW/rack)
Recommendation: Two-Phase Immersion Cooling (2-PIC).
Justification: Provides the highest density (80–150+ kW), superior PUE (1.02), and the best long-term sustainability (using ultra-low GWP fluids). This is the strategy for peak efficiency.
🏢 Brownfield Data Center Retrofits (Medium-to-High Density: 15–30 kW/rack)
Recommendation: Direct Liquid Cooling (DLC).
Justification: Offers immediate efficiency gains for GPU clusters, utilizes existing server form factors, and manages financial risk with lower capital expenditure than a full immersion retrofit.
6) Financial & ESG Advantages of Liquid Cooling
Liquid cooling provides deep financial and environmental benefits beyond basic energy efficiency:
a) Financial Benefits (OPEX & CAPEX)
Energy Savings: Architectures (like NVIDIA GB200 NVL72) demonstrate up to a 25% reduction in annual facility energy consumption.
Space Savings: Systems can achieve up to 75% less rack space required, freeing up expensive data center floor area.
Total Savings: The combined efficiencies can lead to over $4 million in annual operational cost savings for a large hyperscale facility, reframing the initial Capital Expenditure (CAPEX) as a strategic, high-yield investment.
b) Environmental (ESG) Benefits
Waste Heat Reuse: Liquid cooling operates at higher internal temperatures, making the captured waste heat easier to reuse for facility or district heating, which supports ESG objectives and further reduces operational costs.
Standards: Organizations should leverage standards from bodies like ASHRAE TC 9.9, which guides the implementation of resilient cooling systems for extreme AI chip power.
7) Conclusion: Future-Proofing AI Infrastructure
The power demands of modern AI/HPC workloads necessitate a decisive shift from air to liquid cooling.
a) Thermal Imperative & Risk
Air Cooling Limit: Standard air cooling is structurally incapable of managing densities exceeding 15 kW.
Liquid Cooling Capacity: Liquid systems are already managing 42 kW and are projected to surpass 130 kW soon.
Strategic Risk: Delaying liquid cooling is a strategic risk, as it constrains computational density and prevents deployment of the most powerful AI architectures.
b) Strategic Adoption
Tactical Path (DLC): Direct Liquid Cooling (DLC) offers a proven, phased path for incremental high-density adoption.
Strategic Path (2-PIC): Two-Phase Immersion Cooling (2-PIC) offers maximum long-term capacity, efficiency (PUE approx. 1.02), and sustainability (using ultra-low GWP fluids).
8) Future Outlook
Liquid cooling is a fundamental prerequisite for scale-up AI. By 2025-2026, it is expected to dominate new hyperscale data centers, though hybrid approaches (combining liquid and air) will remain practical for balancing cost and performance in retrofitted facilities.
About the Author:
The author is an industry veteran with 30 years of engineering experience spanning IT, Telecom, and Data Centers. They currently serve in a senior role at a major private company—a provider of IT, Telecom, and Cloud Services with national and international importance.
Their expertise is focused on driving business solutions and profitability through the strategic deployment of technology and innovation.
Comments