The rapid advancement of AI infrastructure has made efficient cooling a critical factor for data centers. With NVIDIA's Vera Rubin architecture setting new standards, liquid cooling has become essential rather than optional. This article examines the key differences between two-phase and single-phase liquid cooling for 90kW AI server racks, focusing on performance, costs, and long-term returns.
NVIDIA Rubin Ushers in a New Era of Liquid Cooling
NVIDIA's Rubin platform represents a major shift in data center design. It mandates full liquid cooling, supports warm-water operation up to 45°C, and targets extremely low Power Usage Effectiveness (PUE) levels around 1.06. Key drivers include:
GPU power levels reaching 1.8–2.0 kW each, pushing full racks toward 150–200 kW.
Chip heat flux exceeding 200 W/cm².
The move away from air cooling entirely for next-generation systems.
While 200 kW+ racks are on the horizon, 90 kW configurations remain a practical and widely deployed density level in 2026. The choice between single-phase and two-phase cooling at this level significantly affects both initial investments and ongoing operational expenses.
Technical Comparison: Two-Phase vs. Single-Phase
Two cooling approaches dominate high-density deployments:
Single-phase cold plate cooling relies on sensible heat transfer — the liquid absorbs heat and rises in temperature as it flows.
Two-phase cold plate cooling uses latent heat of vaporization, where the dielectric fluid boils at the heat source, providing far more efficient heat removal at constant temperatures.
Performance MetricsSingle-Phase Cold PlateTwo-Phase Cold PlateAdvantageMax Heat Flux Support≤ 150 W/cm²≥ 200 W/cm²+33% thermal headroomTypical Design PUE1.25 – 1.351.05 – 1.15 (1.08 typical)15–30% energy savingsEstimated Cost per kW~$700 / kW~$900 / kW~30% higher CAPEXHeat Transfer MechanismSensible heat (temp rise)Latent heat (phase change)>60% better efficiency
Two-phase systems excel in handling the extreme heat densities of modern AI chips, making them better prepared for Rubin-class hardware.
5-Year TCO and ROI Analysis
For a 90 kW rack running 8,000 hours annually at $0.11/kWh (typical U.S. industrial rate), the financial picture favors two-phase cooling over time despite higher upfront costs.
5-Year Lifecycle Cost Comparison (USD):
Cost ComponentSingle-Phase (PUE 1.25)Two-Phase (PUE 1.08)Difference (Two-Phase)Initial CAPEX$63,000$81,000+$18,000Annual Power Cost$19,800$17,107-$2,693 savingsAnnual Maintenance$8,500$6,200-$2,300 savings5-Year Total TCO$204,500$197,535-$6,965 net savingsPayback Period—~3.6 yearsROI after ~43 months
Calculation Basis:
Annual Power Cost = 90 kW × 8,000 h × $0.11/kWh × (PUE – 1)
Two-phase systems also benefit from lower mechanical stress on components due to more stable operating temperatures and pressures.
Factors That Influence ROI
Several variables can accelerate or alter the payback timeline:
Higher Electricity Rates (e.g., $0.16/kWh in California or Northeast regions): Annual savings increase to ~$3,917, shortening payback to about 2.8 years.
Further PUE Optimization (down to 1.05): Boosts savings and maintains strong ROI even with modestly higher initial costs.
Higher Rack Densities (150 kW+): Single-phase cooling hits physical limits, making two-phase not just cost-effective but necessary to avoid thermal throttling.
Real-World Deployments
Major operators have already validated two-phase cooling:
Microsoft Azure achieved PUE as low as 1.03 in certain two-phase immersion deployments, significantly reducing chiller dependency and earning carbon credits.
Telecom Retrofit Projects successfully transitioned facilities to two-phase cold plates, lowering cooling PUE from 1.45 to around 1.20 with minimal infrastructure changes.
Maturing supply chains in 2026 have narrowed the cost gap between single- and two-phase components, making adoption more accessible.
Choosing the Right Cooling Strategy
Different deployment scenarios call for tailored approaches:
Brownfield Retrofits (60–90 kW): Pumped two-phase cold plate systems offer fast deployment and strong OPEX reduction.
Greenfield Hyperscale (>100 kW): Two-phase immersion or advanced cold plates for maximum efficiency and density.
Edge Computing (30–60 kW): Single-phase with air backup for simplicity and lower maintenance.
Experimental High-Density (>150 kW): Custom two-phase solutions with direct die attachment.
Quick Tip：
Lian Li is a global leader in data center liquid cooling solutions, driving PUE levels down to below 1.15 and achieving an overall energy savings rate exceeding 30%. We independently research and manufacture our entire product line—ranging from cold plates, liquid-cooled server cabinets, and data center CDUs to containerized liquid-cooled data centers—ensuring compatibility with the full range of NVIDIA and AMD servers, while also offering customized liquid cooling solutions.
Conclusion
As AI servers grow more powerful, two-phase liquid cooling delivers both superior thermal performance and better long-term economics for 90 kW deployments. With a typical payback under 4 years — and faster in high-power-cost regions — it provides a strategic advantage in scalability and efficiency. For operators planning Rubin-era infrastructure, two-phase cooling is increasingly the smart investment for sustainable, high-performance computing.
Lian Li specializes in advanced data center liquid cooling technologies, offering complete solutions from cold plates and server cabinets to full containerized systems, all designed for compatibility with leading AI hardware platforms.
Top comments (0)