Liquid Cooling Servers: Industry Chain Analysis and Selection Guide

1. Why Liquid Cooling Has Shifted from “Optional” to “Mandatory”

As rack power density evolves from the traditional 5–8kW to 20kW, 40kW, and beyond, air cooling is approaching its physical limits. Air has a low specific heat capacity and poor thermal conductivity. As chip power consumption continues to climb, air cooling systems require larger airflow and higher wind speeds, leading to increased fan energy consumption, elevated noise levels, and persistent local hotspots. Crucially, the Power Usage Effectiveness (PUE) of air-cooled data centers typically ranges from 1.3 to 1.5, meaning the cooling process consumes a massive amount of non-IT electrical power.

Liquid cooling uses liquid as a coolant. The specific heat capacity of liquids is thousands of times higher than that of air, resulting in significantly better thermal conductivity and the ability to remove more heat within the same volume. In engineering practices, data centers adopting liquid cooling can reduce PUE to 1.1–1.2 or even lower, while supporting higher rack power densities, lower chip operating temperatures, and longer hardware lifespans. Driven by sustainability goals and surging computing demands, liquid cooling has transitioned from an experimental solution to an engineering necessity for high-density data centers.

2. Two Mainstream Technical Paths: Cold Plate and Immersion

2.1 Cold Plate Liquid Cooling (Indirect Liquid Cooling)

Cold plate liquid cooling works by mounting metal cold plates over high-heat-generating components like CPUs and GPUs. The heat is transferred to the coolant flowing inside the cold plate, and then a CDU (Coolant Distribution Unit) and an external cooling system complete the heat exchange. Its core characteristic is that “the liquid only contacts the heat-generating components, not the entire server.” Consequently, it requires minimal modifications to server motherboards, hard drives, power supplies, and other parts, offering strong compatibility. This makes it the mainstream choice for retrofitting existing data centers.

Key engineering parameters for cold plate solutions include: cold plate flow resistance and flow distribution, coolant inlet/outlet temperatures (the supply water temperature can often be raised to 40–45°C to utilize free cooling), CDU heat exchange capacity and redundancy design, and the reliability and anti-leak design of quick-disconnect fittings. Its advantages lie in a mature industry chain, operational habits similar to traditional air cooling, and controllable retrofit costs. Its limitation is that it still relies on fans for auxiliary cooling of components not covered by cold plates, meaning there is a ceiling for the overall cooling capacity.

2.2 Immersion Liquid Cooling (Direct Liquid Cooling)

Immersion liquid cooling submerges the entire server or rack into a dielectric coolant. The heat is directly absorbed by the liquid and expelled through circulation and heat exchange. Based on whether the coolant undergoes a phase change, it can be further divided into single-phase immersion and two-phase immersion. The advantages of immersion cooling include maximum cooling capacity, extremely low noise, fan-less operation, and support for exceptionally high power densities. The challenges involve coolant costs, material compatibility of server components (such as seals, optical modules, and capacitors’ liquid resistance), restructured operational workflows, and rack-level overall design.

From an engineering selection perspective, cold plate cooling is suitable for “progressive retrofits,” while immersion cooling is ideal for “new high-density clusters.” The two are not mutually exclusive. Some designs adopt a hybrid “liquid + air” cooling strategy: relying primarily on liquid cooling during high-load or hot seasons, and switching to air cooling during low-load or cold seasons to minimize overall energy consumption.

3. Site Selection and Cooling Synergy for Liquid Cooling Data Centers

The energy efficiency of a liquid cooling solution is highly correlated with the data center’s geographic location. In dry regions with large temperature swings, traditional air cooling has a limited window for utilizing free cooling, and dry air introduces electrostatic and dust issues requiring extra treatment. Conversely, liquid cooling systems drastically reduce reliance on the ambient air environment. Paired with dry coolers, they can achieve prolonged free cooling, significantly reducing annual cooling energy consumption. This is a primary technical driver behind the deployment of large-scale computing hubs in regions with favorable, cooler climates.

In system design, “power supply—cooling—rack” must be treated as an integrated whole:

• Rack power density planning: The supply/return liquid temperature, flow rate, and pressure differential of liquid cooling racks must match the IT equipment’s power consumption to avoid over-provisioning or insufficient local flow.
• Power distribution synergy: High-density racks impose higher demands on the capacity and layout of UPS, busways, and PDUs. Liquid cooling piping routes must maintain a safe distance from high-voltage electrical wiring.
• Monitoring and alerting: Real-time monitoring is required for coolant temperature, flow rate, pressure, leak detection, and CDU operational status, all integrated into a unified operations platform.
• Redundancy design: CDUs, pump groups, and heat exchangers typically adopt N+1 or 2N configurations to prevent single points of failure from causing rack-wide overheating.

4. Key Metrics for Rack Power Density and Cold Plate Cooling

When evaluating cold plate liquid cooling solutions, focus on the following engineering metrics (these are common industry reference ranges; exact values depend on vendor specifications and actual testing):

1. Rack power density: Cold plate solutions typically support 20–50kW per rack; immersion solutions can support even higher densities.
2. PUE: The target PUE for liquid-cooled data centers commonly ranges from 1.1 to 1.25, depending on climate conditions and free cooling duration.
3. Coolant supply/return temperature: Raising the supply temperature helps extend free cooling time; a common design range is 35–45°C.
4. Chip junction temperature control: Cold plate solutions must ensure a thermal margin for high-power chips under full load, typically optimized through TIM (Thermal Interface Material) and cold plate flow channel design.
5. Leak risk control: Utilize multi-layer protection such as quick-disconnect fittings, leak detection cables, and pipe pressure monitoring.

A simplified example for estimating cold plate cooling capacity (the code below is illustrative and only represents calculation logic without specific product parameters):

# Illustrative estimation: Estimate removable heat based on coolant flow rate and temperature difference
# Q = m * cp * ΔT
def estimate_cooling_capacity(flow_lpm, delta_t_c, cp_kj_per_kgk=4.0, rho_kg_per_l=1.0):
    """
    flow_lpm: Coolant flow rate (liters/minute)
    delta_t_c: Temperature difference between supply and return (°C)
    Returns: Estimated cooling capacity (kW)
    """
    mass_flow_kg_per_s = flow_lpm * rho_kg_per_l / 60.0
    q_kw = mass_flow_kg_per_s * cp_kj_per_kgk * delta_t_c
    return q_kw
# Example: Flow rate 20 L/min, temperature difference 10°C
print(estimate_cooling_capacity(20, 10))  # Approx. 13.3 kW

This example illustrates that flow rate and temperature difference are the two direct levers for cold plate cooling capacity. In engineering, a balance must be struck among pump power consumption, pipe pressure drop, and thermal margin.

5. Industry Chain Segments and Technical Value Distribution

The liquid cooling server industry chain can be broadly divided into three layers: upstream components, midstream system integration, and downstream data center operations.

• Upstream: Cold plates, CDUs, quick-disconnect fittings, coolants, sealing materials, sensors, pumps, and valves. Among these, CDUs and cold plates represent high technical barriers, involving flow channel design, material corrosion resistance, and sealing reliability.
• Midstream: Liquid-cooled server chassis, liquid cooling racks, and room-level cooling system integration. Server manufacturers must complete compatibility verification between cold plates, motherboards, and CPUs/GPUs, while providing full-rack delivery capabilities.
• Downstream: Data center planning, construction, and operation, focusing on PUE, availability, operational costs, and retrofit timelines.

From a technical value perspective, liquid cooling is not a single piece of hardware but a systematic engineering project comprising “thermal management + fluidics + control + operations.” When evaluating solutions, priority should be given to whether the vendor possesses rack-level delivery and joint commissioning capabilities, provides leak detection and operations toolchains, and has reliable long-term operational data to back their claims.

6. Implementation Path Recommendations

For teams planning to adopt liquid cooling, the following steps are recommended:

1. Requirements Assessment: Map the current rack power density distribution and growth trends to identify high-density hotspot areas.
2. Path Selection: Prioritize cold plate evaluation for existing retrofits; assess immersion or hybrid solutions for new high-density clusters.
3. Pilot Validation: Select a small number of racks for a pilot deployment. Focus on validating thermal margins, leak protection, operational workflows, and monitoring alerts.
4. Standardization: Incorporate cold plate interfaces, CDU specifications, and monitoring protocols into procurement and construction standards to avoid compatibility issues from multi-vendor assemblies.
5. Operations System Development: Establish operational protocols covering coolant testing, pipeline inspections, and emergency leak response drills.

Selecting a liquid cooling server is fundamentally a trade-off among “power density, energy efficiency, reliability, and retrofit cost.” Understanding the engineering constraints behind the technical paths is far more valuable than chasing single conceptual labels.