Planning Guide — High-Density AI Liquid Cooling

Liquid Cooling Strategies
for High-Density AI Compute
Planning for 100kW+ Racks

How to plan liquid cooling for very high rack densities — including direct-to-chip architectures, CDUs, secondary loops, heat rejection, redundancy, retrofit constraints, and the broader thermal infrastructure decisions that determine whether a 100kW+ deployment performs as planned.

Why 100kW+ Rack Planning Changes the Cooling Conversation

The best high-density AI cooling projects do not begin with a rack. They begin with the full thermal chain — from chip to facility to heat rejection.

As AI infrastructure scales, rack density is changing the way data centers must think about cooling. At lower densities, operators may still have multiple architecture options that rely heavily on air. At very high rack densities, liquid cooling becomes less of an optimization and more of an infrastructure planning requirement.

For many operators, the real challenge is not just cooling the rack. It is building the surrounding thermal system that allows the rack to operate reliably, efficiently, and at scale — through deployment phases, changing loads, and the operational realities of a live facility.

MRX approaches liquid cooling for 100kW+ racks as an infrastructure architecture problem built around the heat load, the rack strategy, the heat-rejection path, the controls environment, the maintainability model, and the way the thermal system must fit into the broader facility.

Key Takeaway

Liquid cooling at 100kW+ should be planned as infrastructure — not as equipment. The rack-level decision is one part of a thermal chain that includes CDU strategy, secondary-loop architecture, heat rejection, controls, redundancy, and the operational model the site can actually support.

Infrastructure Architecture

What Liquid Cooling Strategy Really Means

A liquid cooling strategy is more than choosing direct-to-chip or immersion. The right strategy is one that aligns all of these elements into a coherent thermal system — from the chip to the facility boundary.

Rack-level cooling approach
CDU strategy
Secondary-loop architecture
Fluid distribution
Heat exchanger selection
Pumping & pressure management
Controls & monitoring
Leak management strategy

See how MRX engineers custom heat exchangers and thermal management systems and pump skids and fluid-circulation systems for high-density cooling infrastructure.

Direct-to-Chip Cooling

Where Direct-to-Chip Often Fits into 100kW+ Planning

For many high-density AI deployments, direct-to-chip cooling is one of the most practical paths to supporting rack densities beyond what air cooling can reliably handle. It does not make every deployment simpler — CDU and secondary-loop planning remain critical — but it can align well with how many real-world high-density projects are structured.

See the MRX Direct-to-Chip vs. Immersion Cooling Guide for a deeper comparison of these architectures.

Often attractive when the project wants:

Chip-level liquid heat removal without full immersion
A rack architecture close to conventional server deployments
CDU-supported secondary loops with facility water integration
A hybrid air/liquid transition path rather than a full architecture change
Scalability without redesigning the server service model
Strong integration with conventional facility water infrastructure

Immersion Cooling

Where Immersion May Enter the Planning Conversation

At very high thermal densities, immersion may also enter the evaluation set. It is not the right answer for every site — particularly for retrofits or operators who require conventional rack-based service models — but when the facility architecture and operational model genuinely support it, it can be a practical option for reaching very high heat-capture percentages.

Often considered when:

The deployment is planned around immersion — not retrofitted reluctantly into it
The operator is comfortable with tank-based service architecture and workflow
A very high portion of heat load is intended to be captured directly by liquid
The facility is being designed from the ground up to support immersion
The operational team is equipped for immersion-specific maintenance workflows

For a detailed comparison of architectures, see the DTC vs. Immersion Cooling Guide.

Thermal Infrastructure

Why the CDU and Secondary Loop Matter So Much

In most direct-to-chip deployments, the rack-level cooling architecture is only part of the answer. The CDU and secondary-loop design are often central to whether the system performs well — in normal operation, under thermal events, and as the deployment scales.

When these areas are underplanned, the problem usually appears later: in deployment, in operations, or when the first scaling phase begins.

Liquid-to-Liquid Heat Exchange

The CDU is the thermal interface between the rack-level loop and the facility. Sizing, approach temperatures, and control logic all determine whether the rack operates within thermal limits under all load conditions.

Fluid Distribution to the Rack

Flow balance, manifold design, and pressure stability at each rack connection determine whether liquid cooling actually performs as designed — or creates inconsistent results across the deployment.

Pressure & Flow Stability

Pressure and flow management across the secondary loop must account for startup transients, partial-load conditions, and future rack additions — not just the nominal steady-state operating point.

Controls & Monitoring

CDU controls strategy determines how the system responds to thermal events, load changes, and fault conditions. Monitoring coverage determines how quickly the operations team knows when something needs attention.

Maintenance Access

CDU placement and service-access design determine whether maintenance can be performed without disrupting rack operations — and how long that maintenance takes in practice.

Redundancy Planning

Redundancy at the CDU level — including pump redundancy, controls failover, and fluid inventory management — is often the most important determinant of uptime in a liquid-cooled deployment.

The Deciding Factor

The Heat-Rejection Path Still Controls the Outcome

Liquid cooling captures heat efficiently at the rack, but it does not eliminate heat — it moves it. Every kilowatt captured by the liquid loop still has to be rejected somewhere: to facility chilled water, to a cooling tower or dry cooler, to ambient air, or to some combination.

For 100kW+ deployments, the heat-rejection path is often the most consequential infrastructure planning decision — and the one most often underplanned in early project phases.

100kW+ planning often depends on:

Facility chilled water conditions — supply temperature, flow capacity, chemistry
Outdoor heat-rejection strategy — cooling tower, dry cooler, or combined
Heat exchanger selection and approach temperatures
Pumping architecture and system pressure management
Utility capacity and headroom for planned future phases
Redundancy planning at the heat-rejection level
Controls integration and alarm logic for rejection equipment
How future capacity increases will affect the full thermal chain

Operational Reality

What Affects Real-World Performance at 100kW+

Performance at 100kW+ is not determined by rack density alone. Two projects targeting similar rack densities can require very different thermal system designs if the facility constraints, operational model, and scaling plan are materially different.

Liquid Heat-Capture %

Percentage of heat actually captured by liquid under all load conditions — including partial loads and transients.

CDU Strategy

CDU sizing, control logic, and redundancy determine whether the rack operates within thermal limits under all operating conditions.

Loop Stability

Pressure and flow stability across the secondary loop, including startup, load changes, and future rack additions.

Heat-Rejection Capacity

Available rejection capacity and margin under peak loads and worst-case ambient conditions — not just design-day conditions.

Controls Quality

Monitoring coverage, alarm logic, and control responsiveness determine how quickly thermal events are detected and managed.

Deployment Phasing

How the thermal system scales as rack density increases — and whether the infrastructure investment today supports future phases without major rework.

Before You Request a Quote

What Buyers Should Define Before Requesting a Quote

A better RFQ leads to a better thermal system. The more clearly the rack-density, facility, and reliability assumptions are defined up front, the more likely the final architecture will fit the project without unnecessary redesign or scaling problems.

Rack & Thermal Duty

  • Target rack density — near-term and future range
  • Total heat load and liquid heat-capture percentage
  • Cooling architecture under consideration
  • Allowable thermal margin and operating temperature range
  • Expected operating schedule and load profile
  • Deployment phasing assumptions

Facility & Heat Rejection

  • Facility water availability and supply conditions
  • Heat-rejection path and outdoor-rejection strategy
  • Utility constraints — electrical, chilled water, structural
  • Floor layout and space available for CDU and infrastructure
  • Retrofit versus new-build context
  • Future scaling capacity requirements

Controls, Reliability & Operations

  • Controls and monitoring requirements
  • Leak-detection or fluid-management strategy
  • Redundancy expectations — CDU, pumps, controls
  • Alarm and interlock logic
  • Unattended-operation assumptions
  • Maintenance workflow and service-response model
  • Operator visibility requirements

Ready to define your liquid-cooling architecture?

MRX engineers help define the right thermal system before fabrication begins.

Planning Reference

What a Better 100kW+ Cooling Specification Looks Like

A strong high-density liquid-cooling specification defines more than rack power. Each of the following parameters should be resolved before architecture decisions are finalized — use this as a practical reference for evaluating specification completeness or preparing an RFQ.

100kW+ Liquid Cooling — Planning Parameters
01
Target rack densityNear-term target and future density range — define both clearly to allow for scaling margin in the architecture
02
Liquid heat-capture %Percentage of total heat load expected to be captured by liquid — affects CDU sizing and air-side requirements
03
Total heat loadPer-rack and aggregate — including future-phase additions at planned build-out
04
Cooling architectureDirect-to-chip, immersion, hybrid — and stage-gate assumptions if transitioning over time
05
CDU strategyRack-mounted, row-level, or room-level CDU; redundancy requirements; control strategy
06
Secondary-loop designFluid type, operating pressure range, flow rates, manifold approach, balancing strategy
07
Heat-rejection pathFacility chilled water, cooling tower, dry cooler, or combined; supply temperature and flow capacity
08
Thermal marginRequired thermal headroom above nominal load — affects CDU sizing and heat-rejection capacity planning
09
Controls & monitoringSensor coverage, alarm and interlock logic, BMS integration, remote monitoring requirements
10
Redundancy requirementsN+1, 2N, or other — for pumps, CDU units, controls, and heat-rejection equipment
11
Leak managementLeak-detection strategy, secondary containment, fluid inventory management approach
12
Facility constraintsFloor loading, height, footprint, utility tie-in locations, access corridors
13
Deployment phasingBuild-out stages, when liquid cooling is required at each phase, infrastructure pre-investment decisions
14
Service modelMaintenance workflow, service-access requirements, operator staffing and capability assumptions
Avoid These

Common Planning and Selection Mistakes

Many high-density liquid-cooling problems begin in planning, not commissioning. The more AI-dense the deployment, the more expensive these planning mistakes become.

01

Treating 100kW+ cooling as a rack-only problem

The rack cooling decision is one part of a larger thermal system. CDU strategy, secondary-loop design, heat-rejection path, controls, and redundancy are all part of what determines whether the system works in practice — not just whether the rack stays cool.

02

Underdefining the heat-rejection path

Efficient heat capture at the rack is only half the problem. The other half is rejecting that heat somewhere. Facility water conditions, cooling tower or dry-cooler capacity, and utility constraints must be defined before the system can be properly sized.

03

Waiting too long on CDU architecture

CDU placement, sizing, and control strategy affect the floor plan, utility requirements, service access, and system integration in ways that are very difficult to change after the infrastructure layout is fixed.

04

Assuming direct-to-chip and immersion are interchangeable

They are fundamentally different system architectures — with different service models, facility requirements, scaling logic, and operational workflows. Treating them as equivalent hardware options that can be swapped late in planning creates serious project risk.

05

Ignoring retrofit and white-space realities

A cooling architecture designed for a new build may not be practical in a retrofit context. Floor loading, structural constraints, utility tie-ins, and existing infrastructure must all be part of the planning conversation before the architecture is fixed.

06

Underestimating service and maintenance workflow

A liquid-cooled deployment that cannot be maintained efficiently carries its operational burden forward for the life of the system. Service access, leak-response workflow, and maintenance disruption to production must be planned before installation — not discovered after.

07

Failing to define redundancy and controls expectations early

Redundancy at the CDU level and controls coverage across the thermal chain are infrastructure decisions — not late-stage additions. Both affect system sizing, layout, and cost in ways that cannot be retrofitted economically.

08

Treating this as a hardware purchase instead of an infrastructure decision

High-density liquid cooling is an infrastructure architecture choice. The hardware is part of it — but the architecture, integration, and operational model are what determine whether the deployment performs as planned over time.

Why Custom Engineering Matters

Custom Engineering for High-Density AI Cooling

Many 100kW+ rack deployments do not fit a standard infrastructure package well. Facility constraints, retrofit conditions, operational preferences, scaling goals, redundancy requirements, and heat-rejection realities often require a more tailored approach.

Custom engineering means designing the thermal system around the real deployment — the facility it has to fit, the operational model it has to support, and the scaling path it has to enable.

For very high-density AI infrastructure, that alignment often makes the difference between a system that technically cools a rack and one that truly supports performance, uptime, scale, and long-term operability.

Thermal architecture designed around the real facility — not a catalog configuration
CDU and secondary-loop engineering integrated from the start
Heat-rejection sizing matched to actual operating conditions and future phases
Redundancy and controls strategy appropriate for the uptime requirement
Retrofit constraints addressed at the engineering level — not worked around in the field
Scaling path planned into the architecture from the beginning
Service access and maintenance workflow designed for real operational teams
Integration with facility BMS and controls architecture

How MRX Approaches High-Density Liquid Cooling

MRX approaches liquid cooling for high-density AI compute as an infrastructure engineering problem built around the deployment. We support projects where thermal architecture, CDU strategy, heat rejection, pumping, controls, maintainability, and broader system integration all matter.

Our role is not just to supply thermal hardware. It is to help define, engineer, and build the right thermal infrastructure for the way the facility actually needs to operate — in-house in Canby, Oregon, from engineering through fabrication and delivery.

Frequently Asked Questions

Liquid Cooling Strategy FAQs — 100kW+ Racks

Oregon Engineered — Canby, OR

Need Help Planning Liquid Cooling
for 100kW+ Racks?

Talk with MRX about your rack density targets, thermal duty, heat-rejection path, redundancy expectations, facility constraints, and deployment goals. We can help define the right liquid-cooling strategy from engineering through fabrication and delivery.