Data Center Liquid Cooling: Cold Plate vs. Immersion in 2026

Data Center Liquid Cooling: Cold Plate vs. Immersion in 2026

AI infrastructure is running into a thermal engineering problem. Modern accelerators consume so much power that removing their heat has become a fundamental limitation on server density, rack design, and data center capacity.

Traditional air cooling remains effective for conventional server workloads, but high-density AI systems are pushing beyond what conventional heatsinks, fans, and room-level airflow can handle economically.

NVIDIA Blackwell systems provide a clear example. Rack-scale platforms such as GB200 NVL72 and GB300 NVL72 integrate dozens of GPUs into a single computing system, with rack power and thermal loads reaching well over 100 kW. NVIDIA specifies a fully liquid-cooled architecture for GB300 NVL72.

The emerging Rubin generation pushes the same engineering challenge further.

As a result, liquid cooling has moved from specialized HPC environments into mainstream AI infrastructure planning.

However, liquid cooling is not a single technology. Direct-to-chip cold plates, rear-door heat exchangers, single-phase immersion, and two-phase immersion all have different thermal characteristics, operational requirements, and deployment costs.

The right choice depends on accelerator power density, rack configuration, facility infrastructure, serviceability, and long-term operating costs.

💧 Why AI Data Centers Are Moving Toward Liquid Cooling

#

The basic problem is straightforward: nearly all electrical power consumed by computing hardware eventually becomes heat that the facility must remove.

As processor and accelerator power increases, the cooling system must transfer more heat through the same physical footprint.

This becomes particularly difficult in AI racks.

The Rack Power Density Problem

#

Traditional enterprise servers typically operate at relatively moderate rack power densities. General-purpose CPU workloads can often be cooled with conventional air conditioning and carefully managed airflow.

AI servers are different.

A rack containing multiple high-performance GPUs can consume tens or hundreds of kilowatts.

For example, NVIDIA GB200 NVL72 systems combine 72 Blackwell GPUs and 36 Grace CPUs in a rack-scale configuration. HPE’s published GB200 NVL72 specifications identify a nominal rack power of approximately 132 kW, with peak electrical design power reaching approximately 192 kW.

These figures illustrate why conventional rack cooling architectures are becoming difficult to scale.

The thermal load increases faster than the practical capacity of conventional airflow systems.

Several constraints emerge:

• Increasing fan speeds consumes additional electrical power.
• Higher airflow requirements create pressure and noise problems.
• Heatsinks require more physical space.
• Air temperature gradients become harder to control.
• Hotspots develop around high-power processors.
• Increasing rack density makes hot-air recirculation more difficult to manage.

At some point, increasing airflow becomes an expensive way to solve a problem that is fundamentally about heat transfer.

Liquid provides a more effective medium for transporting large amounts of thermal energy.

Why Liquid Transfers Heat More Effectively Than Air

#

The thermal conductivity and volumetric heat capacity of common cooling liquids are significantly higher than those of air.

Water, for example, has a volumetric heat capacity roughly 3,500 times greater than air under typical ambient conditions.

This means a relatively small volume of water can transport substantial thermal energy compared with an equivalent volume of air.

The practical advantages include:

• Higher heat removal capacity per unit of volume.
• Reduced dependence on high-speed server fans.
• Smaller heat exchanger footprints for comparable thermal loads.
• Better control of high-density processor temperatures.
• Support for higher rack power densities.

This does not mean every component should be liquid-cooled.

A practical AI server typically uses a hybrid thermal architecture, with liquid cooling for the highest-power components and air cooling for the remaining hardware.

Liquid Cooling Is Not Required Everywhere

#

One misconception is that all modern data centers need to replace air cooling with liquid.

That is not the case.

The transition is driven primarily by rack-level thermal density.

General-purpose servers, storage systems, and networking equipment can still operate efficiently with conventional air cooling or rear-door heat exchangers.

High-density GPU infrastructure has different requirements.

A more realistic 2026 deployment strategy is to create separate thermal zones:

The objective is to match the cooling architecture to the actual heat load rather than applying the same solution throughout the facility.

❄️ Direct-to-Chip Liquid Cooling: Cold Plate Architecture

#

Direct-to-chip (D2C) liquid cooling, commonly called cold plate cooling, is currently the most practical approach for large-scale AI server deployments.

It combines conventional server architecture with liquid-based heat removal.

Unlike immersion cooling, the coolant does not come into direct contact with the electronic components.

Instead, heat travels through a cold plate attached to the processor package.

How Cold Plate Cooling Works

#

A typical cold plate contains internal microchannels through which liquid coolant circulates.

The cooling assembly is mounted directly above a CPU or GPU package, using a thermal interface material (TIM) to transfer heat from the processor into the cold plate.

The basic thermal path is:

GPU / CPU Silicon
        |
        v
    Chip Package
        |
        v
 Thermal Interface
      Material
        |
        v
     Cold Plate
        |
        v
   Microchannels
        |
        v
   Liquid Coolant
        |
        v
 Cooling Distribution
      Unit (CDU)
        |
        v
 Facility Cooling Loop
        |
        v
   Heat Rejection

The cold plate absorbs heat through conduction.

The circulating coolant then transports that thermal energy away from the processor and toward a heat exchanger.

Most high-performance implementations use copper cold plates because of their favorable thermal conductivity.

Internal microchannel structures increase the effective heat transfer area and help maintain temperature uniformity across the processor surface.

The Role of the Cooling Distribution Unit (CDU)

#

The CDU is one of the most important components in a liquid-cooled data center.

It manages coolant circulation between the IT equipment and the facility cooling infrastructure.

A typical system contains two separate loops.

1. Technology Cooling System (Secondary Loop)

The secondary loop supplies coolant directly to the servers.

Its components typically include:

• Pumps.
• Coolant reservoirs.
• Temperature sensors.
• Pressure sensors.
• Flow meters.
• Rack manifolds.
• Supply and return piping.

The CDU regulates coolant temperature and flow while maintaining the required operating pressure.

2. Facility Water System (Primary Loop)

The primary loop receives heat from the CDU through a heat exchanger.

It transports that heat to the facility’s cooling infrastructure.

Depending on the site, heat rejection may involve:

• Dry coolers.
• Cooling towers.
• Chillers.
• Evaporative cooling systems.
• Hybrid heat rejection equipment.

The two-loop architecture isolates the server coolant from the facility water system.

It also allows server manufacturers to specify coolant chemistry, pressure, and temperature independently of the facility’s primary cooling infrastructure.

Some systems use alternative configurations, but the separated-loop design is common in large AI deployments.

Why Cold Plate Cooling Dominates High-Density AI Deployments

#

Cold plate systems offer several practical advantages.

Compatibility With Existing Server Architecture

#

The servers remain conventional rack-mounted systems.

Components can still be accessed through standard server trays, rails, and maintenance procedures.

This is a major advantage over immersion systems, which require different chassis and servicing workflows.

Hybrid Cooling Support

#

Cold plates can target the hottest components without requiring the entire server to be submerged.

For example:

• GPUs use cold plates.
• CPUs use cold plates.
• Memory modules may use dedicated airflow or liquid cooling, depending on the design.
• Network adapters use heatsinks and airflow.
• SSDs use conventional thermal solutions.
• Power supplies retain their own cooling systems.

This hybrid approach allows manufacturers to focus liquid cooling where it provides the greatest benefit.

It also avoids introducing unnecessary complexity into components that can still be cooled effectively with air.

Retrofit Potential

#

Existing data centers can sometimes adopt direct liquid cooling without rebuilding the entire facility.

A deployment may require:

• Additional CDU capacity.
• Rack manifolds.
• Coolant distribution piping.
• Leak detection.
• Additional facility heat rejection capacity.

However, the electrical distribution system, floor loading, ceiling height, and available mechanical capacity must also be evaluated.

A facility designed around 10–20 kW racks cannot automatically support 100 kW AI racks simply by adding cold plates.

The infrastructure upgrade can be substantial.

Cold Plate Engineering Challenges

#

Cold plate cooling is mature relative to immersion, but it introduces several engineering constraints.

Thermal interface resistance

The TIM between the processor package and cold plate contributes to total thermal resistance.

Poor contact pressure, uneven mounting, or TIM degradation can create localized hotspots.

Coolant compatibility

The coolant must be compatible with the cold plate material, seals, tubing, and connectors.

Mixing copper and aluminum components without proper material and coolant management can introduce corrosion concerns.

Pressure drop

Microchannels improve heat transfer but increase hydraulic resistance.

Higher pressure drops require additional pumping capacity, which increases system power consumption.

Leak management

Even a small coolant leak inside an energized server can create serious operational problems.

Modern systems therefore require reliable quick disconnects, leak sensors, pressure monitoring, and isolation mechanisms.

For another look at cold plate engineering and the thermal path from the processor to the facility, see Frore Systems Shows How Advanced Cooling Can Boost NVIDIA Rubin.

🚪 Rear-Door Heat Exchangers: The Hybrid Alternative

#

Rear-door heat exchangers (RDHx) provide another approach to liquid-assisted data center cooling.

Instead of attaching cold plates directly to CPUs and GPUs, an RDHx system installs a liquid-cooled heat exchanger at the rear of the server rack.

Hot exhaust air passes through the heat exchanger, transferring heat into circulating liquid.

The cooled air then returns to the data center environment.

How Rear-Door Cooling Works

#

A simplified architecture looks like this:

Server Rack
    |
    v
Hot Exhaust Air
    |
    v
Rear-Door Heat Exchanger
    |
    v
Heat Transfer to Liquid
    |
    v
Facility Cooling Loop

There are two common approaches:

• Passive rear-door heat exchangers, which rely on server fans to move air through the exchanger.
• Active rear-door heat exchangers, which use additional fans to control airflow.

RDHx can reduce the amount of heat released into the data hall without changing the internal cooling design of every server.

This makes it attractive for retrofitting existing facilities.

RDHx Limitations

#

Rear-door systems still depend on air moving through the servers.

At extremely high rack power densities, the airflow requirement can become difficult to manage.

They are therefore more appropriate for moderate-density deployments or hybrid architectures than for the highest-power GPU racks.

A rack with 100 kW or more of concentrated accelerator load will generally require a more direct heat removal strategy.

RDHx remains useful as part of a mixed cooling design, especially when different rack types coexist in the same data hall.

🌊 Immersion Cooling: Direct Liquid Contact

#

Immersion cooling takes a different approach.

Instead of transferring heat through a cold plate, it places electronic components or complete server assemblies inside an electrically non-conductive dielectric liquid.

The fluid surrounds the heat-generating components and absorbs thermal energy directly from their surfaces.

This removes the need for conventional heatsinks and large volumes of server airflow.

However, immersion cooling is not simply a more efficient version of cold plate cooling.

It requires different hardware, materials, maintenance procedures, and facility infrastructure.

There are two main approaches:

• Single-phase immersion.
• Two-phase immersion.

Single-Phase Immersion Cooling

#

In single-phase immersion systems, the dielectric coolant remains liquid throughout normal operation.

Heat from the processors and other electronic components transfers into the surrounding fluid.

Pumps circulate the heated fluid through a heat exchanger, where it transfers thermal energy to a secondary cooling loop.

The coolant then returns to the immersion tank.

A simplified architecture looks like this:

Immersed Server Components
          |
          v
   Dielectric Fluid
          |
          v
    Circulation Pump
          |
          v
    Heat Exchanger
          |
          v
  Secondary Cooling Loop
          |
          v
    Heat Rejection
          |
          v
   Cooled Fluid Returns

Advantages

#

Single-phase immersion offers several engineering benefits.

• Large heat transfer surface area.
• Reduced dependence on server fans.
• Relatively simple circulation compared with two-phase systems.
• High rack-level thermal capacity.
• Potentially effective heat reuse through higher-temperature coolant loops.

Because the entire server assembly is surrounded by fluid, cooling is not restricted to CPU and GPU packages.

Memory modules, voltage regulators, and other compatible components also receive direct fluid cooling.

Limitations

#

Immersion requires a different approach to server maintenance.

Technicians must handle fluid-filled equipment, and certain maintenance procedures require lifting hardware out of the tank.

Additional considerations include:

• Dielectric fluid compatibility.
• Material degradation.
• Connector and cable qualification.
• Fluid filtration.
• Tank design.
• Service equipment.
• Fluid storage and replacement.

The server platform must also be designed and qualified for immersion.

Conventional air-cooled servers cannot automatically be submerged in dielectric fluid without checking component compatibility and manufacturer approval.

Two-Phase Immersion Cooling

#

Two-phase immersion uses a dielectric fluid with a relatively low boiling point.

Instead of relying entirely on sensible heat transfer, it uses phase change to transport thermal energy.

The process works as follows:

1. Electronic components heat the surrounding liquid.
2. The fluid reaches its boiling temperature.
3. Part of the liquid vaporizes.
4. Vapor rises toward a condenser located above the immersion bath.
5. The condenser removes heat and converts vapor back into liquid.
6. The condensed fluid returns to the tank.

The underlying principle is latent heat transfer.

A relatively small amount of fluid can carry substantial thermal energy through vaporization and condensation.

Advantages

#

Two-phase systems can provide high heat flux handling and relatively uniform component temperatures.

They reduce dependence on high-flow liquid circulation directly around the electronic components.

The architecture can be attractive for specialized HPC workloads and extremely dense computing environments.

Engineering Challenges

#

The thermal advantages come with additional complexity.

Fluid cost and availability

Specialized dielectric fluids can be expensive, and long-term supply availability must be considered.

Fluid lifecycle

Fluid degradation, contamination, and replacement procedures affect operating costs.

Environmental considerations

Some engineered fluorinated fluids face increasing environmental scrutiny and regulatory uncertainty. Fluid selection therefore needs to account for long-term availability, environmental compliance, and disposal requirements.

Containment

The system must prevent vapor loss and maintain suitable operating conditions.

Serviceability

Removing or replacing server hardware is more complicated than working with conventional rack-mounted systems.

These factors make two-phase immersion a specialized option rather than an automatic choice for every high-density data center.

Immersion Cooling Does Not Guarantee Zero Water Usage

#

A closed dielectric-fluid loop does not necessarily mean a data center consumes no water.

The facility still needs to reject heat into the environment.

If the external heat rejection system uses evaporative cooling towers, it can consume water even though the immersion circuit itself is sealed.

A dry-cooler-based design can reduce operational water consumption, but may require different heat exchanger sizing, larger equipment, or additional mechanical cooling depending on ambient conditions.

The complete facility design determines water usage.

📊 Liquid Cooling Technology Comparison

#

The major cooling architectures have different strengths.

The following table summarizes their practical characteristics.

There is no universal winner.

Cold plate systems currently provide the most practical combination of thermal performance, maintainability, and compatibility with mainstream AI server architectures.

Immersion becomes more attractive when the full system can be designed around it and its thermal advantages justify the additional operational complexity.

Market Share Requires Careful Interpretation

#

Market statistics for liquid cooling are not always directly comparable.

Some reports measure the percentage of data center operators using a technology. Others measure facilities, racks, installed IT capacity, or revenue.

These denominators produce different results.

For example, Uptime Institute’s 2024 cooling survey found that 22% of surveyed operators reported using direct liquid cooling. Among respondents already using direct liquid cooling, water-cooled cold plates represented the most commonly reported approach.

That figure should not be interpreted as a 2026 global market share.

The market is changing quickly, and different cooling technologies can coexist within the same facility.

For engineering decisions, rack density, cooling capacity, serviceability, and total operating cost are more useful metrics than an isolated market-share percentage.

⚡ The Blackwell and Rubin Impact

#

NVIDIA’s Blackwell generation has made liquid cooling a central consideration in high-density AI system design.

The transition is not driven by GPU power consumption alone.

It comes from the combined thermal load of the accelerators, CPUs, memory, networking equipment, and power conversion infrastructure.

NVIDIA GB200 and GB300: Rack-Scale Liquid Cooling

#

The GB200 NVL72 architecture integrates:

• 72 Blackwell GPUs.
• 36 Grace CPUs.
• NVLink switching infrastructure.
• High-speed networking.
• Rack-level power distribution.
• Liquid cooling infrastructure.

The system is designed as a complete rack-scale computing platform rather than a collection of independent GPU servers.

This requires coordinated thermal engineering across compute trays, switch trays, manifolds, and cooling distribution equipment.

NVIDIA’s GB300 NVL72 extends this architecture with 72 Blackwell Ultra GPUs and 36 Grace CPUs.

NVIDIA explicitly describes GB300 NVL72 as a fully liquid-cooled rack-scale system.

See the official NVIDIA GB300 NVL72 specifications for the platform configuration.

The engineering implications are significant.

Cooling is no longer something that server manufacturers can treat as an optional accessory after the compute architecture has been finalized.

The cold plates, coolant distribution, rack manifolds, power system, and facility interface must be designed together.

Rubin Pushes Thermal Engineering Further

#

NVIDIA’s Vera Rubin platform represents the next step in rack-scale AI infrastructure.

Its architecture integrates Rubin GPUs, Vera CPUs, networking, and rack-level management into larger computing systems.

The move toward higher-density AI platforms increases pressure on:

• Cold plate heat flux.
• Thermal interface materials.
• Coolant flow rates.
• Manifold design.
• CDU capacity.
• Rack-level heat rejection.
• Facility cooling infrastructure.

The thermal path must be optimized from the processor die all the way to the external heat rejection system.

This is increasingly becoming a system-engineering problem rather than an isolated GPU cooling problem.

For a related look at next-generation cold plate designs and the processor-to-chiller thermal path, see Frore Systems Shows How Advanced Cooling Can Boost NVIDIA Rubin.

Beyond NVIDIA: Custom AI Accelerators

#

The shift toward liquid cooling is not limited to NVIDIA GPUs.

Custom AI accelerators from companies such as Google, Amazon, and other hyperscalers also face increasing thermal requirements.

These platforms may use different package layouts, memory configurations, and interconnect architectures, but the underlying engineering challenge is similar.

Higher compute density produces more heat per unit of physical space.

As a result, liquid cooling is becoming relevant to:

• Google TPU systems.
• AWS Trainium infrastructure.
• AMD Instinct accelerator racks.
• Custom hyperscaler ASICs.
• High-performance networking platforms.

The exact implementation depends on the accelerator package and server design.

Some use conventional cold plates, while others require custom thermal interfaces and specialized rack-level cooling infrastructure.

🌡️ PUE, WUE, and Data Center Cooling Efficiency

#

Liquid cooling is frequently associated with better Power Usage Effectiveness (PUE).

However, PUE is a facility-level metric, not a direct measurement of how efficiently an individual cold plate removes heat.

The distinction is important.

Understanding PUE

#

Power Usage Effectiveness measures total facility energy consumption relative to IT equipment energy consumption.

\[ PUE = \frac{\text{Total Facility Energy}}{\text{IT Equipment Energy}} \]

A PUE of 1.0 represents the theoretical minimum, where all facility energy is consumed by IT equipment.

A lower PUE generally indicates less overhead from cooling, electrical distribution, and other facility systems.

Liquid cooling can improve PUE by reducing the energy required for server fans and room-level air conditioning.

However, the result depends on:

• Coolant supply temperature.
• Pump efficiency.
• Heat exchanger performance.
• Outdoor climate.
• Chiller requirements.
• Cooling tower configuration.
• Server fan power.
• Facility utilization.

A liquid-cooled facility does not automatically achieve a particular PUE.

A poorly designed liquid cooling system can consume substantial pumping and mechanical cooling power.

Understanding WUE

#

Water Usage Effectiveness measures data center water consumption relative to IT energy consumption.

\[ WUE = \frac{\text{Annual Water Usage}}{\text{Annual IT Energy}} \]

Liquid cooling does not inherently guarantee low WUE.

A cold plate system connected to an evaporative cooling tower may consume substantial water.

An immersion system using dry coolers may achieve much lower direct water consumption, but could require more electrical energy under certain ambient conditions.

Operators need to evaluate both metrics.

Heat Reuse

#

Liquid cooling also creates opportunities for heat recovery.

Compared with conventional low-temperature air exhaust, liquid cooling loops can provide relatively concentrated heat at temperatures suitable for certain reuse applications.

Potential applications include:

• Building heating.
• District heating networks.
• Industrial process heating.
• Domestic hot water preheating.
• Agricultural facilities.

The economic feasibility depends on the availability of a nearby heat consumer and the temperature requirements of that application.

Heat recovery should be designed into the facility rather than treated as an automatic benefit of installing liquid cooling.

🔧 Operational Challenges: What Engineers Need to Plan For

#

A production liquid cooling deployment requires much more than selecting cold plates or immersion tanks.

The facility must maintain reliable coolant circulation while supporting maintenance, fault isolation, and continuous operation.

Leak Detection and Isolation

#

Liquid systems introduce failure modes that conventional air-cooled servers do not have.

Potential problems include:

• Damaged tubing.
• Loose connectors.
• Seal degradation.
• Pump failures.
• Corrosion.
• Coolant contamination.
• Pressure loss.

A production system should include leak detection at suitable points, isolation valves, pressure monitoring, and redundant pumping where required.

Quick-disconnect fittings also need to be qualified for the expected service life and maintenance frequency.

Coolant Quality Management

#

Coolant chemistry affects the reliability of the entire system.

Important parameters include:

• Conductivity.
• pH.
• Corrosion inhibitor concentration.
• Particulate contamination.
• Microbial growth.
• Material compatibility.

The appropriate limits depend on the coolant chemistry and server manufacturer’s requirements.

Using ordinary facility water directly inside a server loop without proper treatment and compatibility validation is not an acceptable substitute for a qualified technology cooling system.

Redundancy and Failure Handling

#

AI training systems may operate continuously for extended periods.

A cooling failure can trigger thermal throttling, accelerator shutdown, or complete rack unavailability.

Cooling architecture should therefore consider:

• Redundant pumps.
• Redundant power supplies for CDUs.
• Backup coolant distribution paths.
• Temperature and flow telemetry.
• Automatic fault isolation.
• Controlled thermal shutdown.
• Spare cooling capacity.

Redundancy must cover the complete thermal path.

A server may have redundant pumps, but that does not help if the facility-side heat exchanger or common coolant header becomes unavailable.

Negative-Pressure Liquid Cooling

#

Another approach to leak mitigation is negative-pressure cooling.

Instead of maintaining positive pressure throughout the liquid loop, the system operates below atmospheric pressure in selected sections.

If a small leak develops, the pressure differential can draw air into the system rather than forcing coolant outward.

This can reduce the risk of coolant spraying onto electronic components.

It does not eliminate every failure mode. Air ingress, pump behavior, coolant flow, and pressure stability still require careful monitoring.

For a more detailed discussion of this architecture, see Negative Pressure Liquid Cooling for Data Centers.

🏗️ Designing AI Data Centers Around the Thermal System

#

One of the biggest changes in AI infrastructure design is the shift from server-centric planning to rack- and facility-level thermal planning.

Traditional data centers were often designed around relatively standardized server power requirements.

Modern AI installations require a much tighter relationship between compute architecture and cooling infrastructure.

Cooling Distribution Must Match Rack Power

#

CDU capacity should be sized according to the actual thermal load and redundancy requirements.

A rack consuming 100 kW of IT power requires the cooling system to transport approximately 100 kW of heat under sustained full-load conditions, subject to the distribution of power and heat sources.

This means facility engineers must account for:

• Rack heat load.
• Coolant flow rate.
• Supply and return temperatures.
• Heat exchanger approach temperature.
• Pressure drop.
• Pump efficiency.
• N+1 or 2N redundancy requirements.

The coolant flow rate is determined by the heat load and the permitted temperature rise.

A simplified relationship is:

\[ Q = \dot{m} \times c_p \times \Delta T \]

Where:

• \(Q\) is the heat transferred.
• \(\dot{m}\) is the coolant mass flow rate.
• \(c_p\) is the coolant specific heat capacity.
• \(\Delta T\) is the temperature increase across the cooling loop.

For water-based coolant, increasing the permitted temperature difference can reduce the required mass flow rate for a given thermal load.

However, supply temperature must remain within the operating limits of the processors and cold plates.

Warm-Water Cooling and Free Cooling

#

Many direct-to-chip systems can operate with relatively warm coolant.

This creates an opportunity to reduce or eliminate mechanical chiller operation during favorable ambient conditions.

Depending on climate and facility design, operators can use:

• Dry coolers.
• Waterside economizers.
• Cooling towers.
• Adiabatic heat rejection.
• Hybrid chiller and free-cooling systems.

Higher coolant temperatures can improve the efficiency of external heat rejection.

They can also make heat recovery easier.

The trade-off is thermal headroom: warmer coolant leaves less temperature difference available for removing heat from the hottest components.

This is why the cold plate, processor package, coolant loop, and facility heat exchanger must be engineered as one thermal system.

Rack-Level Cooling Is Becoming Part of Compute Architecture

#

Modern AI racks should be evaluated as integrated systems containing:

• CPUs and GPUs.
• Memory and networking.
• Power shelves.
• Busbars.
• Liquid manifolds.
• Cold plates.
• CDUs.
• Monitoring and control systems.

The cooling architecture influences how much compute can fit inside a rack, how reliably it can operate, and how efficiently the facility can reject heat.

The relationship works in both directions.

More efficient cooling allows greater compute density within a fixed thermal envelope, while higher compute density places additional demands on the cooling system.

For AI infrastructure, thermal design is no longer a secondary facility concern. It is one of the constraints that determines the architecture of the computing system itself.