Case Study
AI Compute Infrastructure

Case Study:
High-Density AI Compute
Liquid Cooling Infrastructure

How MRX engineered and delivered a complete direct-to-chip cooling system for a 200-rack AI training facility — reducing PUE from 1.8 to 1.06 through integrated thermal architecture, not just hardware selection.

200

Racks

AI Training Facility

1.8→1.06

PUE

Achieved in Deployment

40%

Reduction

Cooling Energy vs. Baseline

D2C

Turnkey

Direct-to-Chip System

The Challenge

When Rack Density Outgrows What Air Can Handle

As AI rack density rises, data-center projects face a different class of infrastructure challenge. The problem is no longer only how to deploy more compute. It is how to support that compute thermally — reliably, maintainably, and at a scale that can grow.

For high-density environments, the real engineering challenge often sits around the rack rather than inside it. You can select the best cold plates and CDUs available, but if the thermal infrastructure surrounding them is underspecified or poorly integrated, the whole system underperforms.

Liquid cooling architecture

How heat moves from the rack to the facility — CDU, secondary loops, and heat rejection.

Fluid distribution

Flow, pressure, and manifold strategy across a 200-rack deployment.

Controls and monitoring

Visibility into thermal performance before problems become failures.

Serviceability

Designed to be maintained without full-system downtime.

Heat rejection

Capturing heat at the rack is only half the problem — rejecting it efficiently is the other half.

Facility integration

Aligning the cooling system with what actually exists at the site.

Key Takeaway

High-density AI cooling is an infrastructure problem, not a rack-hardware problem. The systems that sit around and between the racks — CDUs, secondary loops, heat exchangers, pump circuits, controls — must be engineered as a coherent system, not assembled from independent selections.

Project Context

A 200-Rack AI Training Facility With a System-Level Cooling Requirement

The customer required a liquid-cooling infrastructure approach for a high-density AI compute environment where thermal performance, deployment practicality, and infrastructure integration all mattered.

The facility was designed around AI training workloads — continuous, high-utilization GPU compute that generates sustained heat loads that air cooling cannot manage at this density. A 200-rack deployment at this density requires a system-level thermal solution, not a rack-by-rack hardware decision.

The need was not limited to cooling hardware alone. The project required a broader thermal-system mindset that addressed heat capture, heat rejection, fluid distribution, controls, packaging, and integration with the facility — from a single engineering and fabrication source.

Deployment Profile

Facility typeAI training cluster
Rack count200 racks
Cooling strategyDirect-to-chip liquid cooling
System approachTurnkey, design-build
Engagement typeConfidential customer — anonymized

The Engineering Problem

Defining the Right Thermal Architecture Was the Core Challenge

The core challenge was not simply selecting equipment. It was defining the right thermal architecture — and then building it as a coherent, manufacturable system.

This is where many high-density AI projects become infrastructure problems rather than single-product problems. Each decision ripples through the system: CDU sizing, secondary-loop temperature targets, pump curves, heat exchanger capacity, fluid distribution headers, and controls logic must all align around the same thermal duty.

At this density level, the project needed to account for:

Rack-level cooling strategy — direct-to-chip cold plate architecture and flow requirements
How heat would be transferred out of the compute environment through the CDU
Secondary-loop thermal management and heat exchanger strategy
How liquid would be distributed and controlled across 200 racks
How the infrastructure would be packaged for the deployment environment
How the cooling strategy would align with the broader facility infrastructure
Maintainability — designing for service without full-system shutdown

See the MRX guide on liquid cooling strategies for 100kW+ racks and the technical comparison of direct-to-chip vs. immersion cooling.

MRX Approach

Design-Build Thermal Infrastructure — Not Equipment Supply

MRX approached the project as a design-build thermal infrastructure problem. The focus was on building a practical, manufacturable solution around the deployment requirements — not forcing the deployment into a generic package.

That distinction matters. When engineering and fabrication are managed together, the decisions that shape system performance — thermal architecture, packaging, integration, serviceability — can be made in coordination rather than in sequence. That coordination is often where project risk is either created or controlled.

Key priorities in MRX's approach:

Align the liquid-cooling architecture with the actual thermal duty — density, heat load, and ambient conditions
Engineer the CDU and secondary-loop strategy around the facility infrastructure that exists
Design fluid distribution for the 200-rack deployment with appropriate flow and pressure
Integrate heat-exchange and heat-rejection components as part of the system — not as afterthoughts
Design for maintainability from the start — service access, isolation provisions, monitoring
Preserve visibility between engineering, fabrication, and deployment planning
Support a complete thermal chain — from cold plate to heat rejection — rather than only a rack-level answer

What Was Delivered

A Complete Turnkey Direct-to-Chip Cooling System

MRX engineered and delivered a complete direct-to-chip cooling system for the 200-rack AI training facility. The scope was designed as an integrated turnkey thermal infrastructure package — not a collection of independently sourced components.

Direct-to-Chip Cold Plate Architecture

Rack-level liquid cooling capturing heat directly at the processor across the 200-rack deployment.

CDU and Secondary-Loop Infrastructure

Coolant distribution units and secondary fluid loops engineered around the facility interface.

Heat Exchange and Heat Rejection

Thermal support equipment integrated into the system to manage heat from CDU to final rejection point.

Pump and Fluid Distribution

Pumping systems and distribution manifolds sized and designed for the deployment scale.

Controls and Monitoring

System controls providing visibility into thermal performance and supporting reliable continuous operation.

Turnkey Packaged System

Engineering, fabrication, and integration managed in-house in Canby, Oregon — single-source accountability.

Why the Solution Worked

Engineering Logic, Not Just Hardware

The value of the solution was not just in the hardware itself. It was in the alignment between the thermal challenge and the system architecture — and in the coordination between engineering and fabrication that made that alignment practical to build and deploy.

Projects that treat high-density cooling as a component-selection problem often find the gaps during commissioning: the CDU is undersized for the actual heat load, the secondary loop cannot reach the temperature target, or the distribution manifold creates flow imbalances across the rack array. Those problems are expensive to fix after the system is deployed.

Stronger match between thermal duty and infrastructure design

The system was sized and designed around the actual heat load, not around a generic envelope. That alignment is the foundation of the PUE outcome achieved.

Better coordination between engineering and fabrication

With both functions managed in-house, engineering decisions translated directly into fabrication without the integration gaps that appear when systems are sourced from multiple vendors.

Practical packaging around deployment realities

The system was packaged around how it would actually be installed and operated — not around what was easiest to fabricate in isolation.

System-level approach to heat movement and rejection

Cold plate, CDU, secondary loop, and heat rejection were designed as a coherent chain — not as independent selections optimized separately.

Design choices that supported maintainability

Service access, isolation provisions, and monitoring were built into the design from the start — not added as modifications after the architecture was set.

Eliminating the root cause of PUE inefficiency

Air cooling inefficiency is structural — fans, CRAC units, hot-aisle pressure losses. Direct-to-chip liquid cooling removes those losses from the architecture entirely.

Project Results

PUE 1.8 to 1.06 — Reduced by 40%

The delivered system achieved a power usage effectiveness of 1.06 for the 200-rack AI training facility — a 40% reduction from the pre-deployment baseline of 1.8. These are approved outcomes from MRX's delivered project record.

Baseline PUE

1.80

Air-cooled at this rack density

↓

Achieved PUE

1.06

40% reduction in cooling overhead

Additional Outcomes

Improved thermal readiness across the 200-rack AI training deployment
Stronger infrastructure alignment with rack-density and duty requirements
Practical path to liquid-cooling deployment via turnkey design-build
Improved maintainability through built-in service access and monitoring
Single-source accountability from engineering through fabrication and delivery
Thermal chain continuity from cold plate to heat rejection without integration gaps
“
Their liquid cooling system dropped our PUE from 1.8 to 1.06. The engineering team understood our rack density challenges from day one — no hand-holding required.

Confidential Customer — AI Training Cluster Deployment

Why This Case Study Matters

A Broader Reality in AI Infrastructure

This case study is useful because it reflects a challenge facing many operators building high-density AI environments today. The question is not simply whether liquid cooling is needed — at 100kW+ rack densities, it is. The question is how to build the surrounding thermal infrastructure in a way that is manufacturable, maintainable, and aligned with real deployment conditions.

That is the level at which MRX works. The 40% PUE reduction achieved in this deployment was not a product of hardware specification alone — it was the result of engineering the full thermal chain as a coherent system, coordinating that engineering with fabrication, and packaging the result around what the deployment actually required.

What makes this outcome replicable for similar projects:

Defining the thermal architecture before selecting hardware
Engineering CDU, secondary loop, and heat rejection as a system
Coordinating engineering and fabrication under single-source accountability
Designing maintainability in from the start — not as a retrofit
Aligning the deployment package with real facility constraints

For Similar Projects

Liquid Cooling Should Be Evaluated as Infrastructure, Not Equipment

For other high-density AI projects, the lesson is clear: liquid cooling should be evaluated as infrastructure, not just as equipment. The more density-driven the deployment, the more important it is to think early about the systems that sit around and between the racks — not only the hardware inside them.

Early decisions that shape project outcomes:

Heat-rejection path

Defining this early determines whether the cooling system can achieve its PUE target under real ambient conditions.

CDU and secondary-loop architecture

Undersized or mismatched loops create constraints that cannot be resolved by changing rack-level hardware.

Fluid distribution strategy

Flow balance across a large rack array requires deliberate manifold and pump design — not field improvisation.

Controls and monitoring

Thermal visibility built in from the start — not added after deployment reveals performance gaps.

Maintainability planning

Service access, isolation valves, and drain provisions must be designed in before the architecture is set.

Facility integration

The cooling system must align with what the facility actually provides — chilled water, ambient conditions, space.

MRX Approach to High-Density AI Cooling

Infrastructure Engineering Built Around the Deployment

MRX approaches high-density AI liquid cooling as an infrastructure engineering problem built around the deployment — not as a hardware catalog. We support projects where thermal architecture, heat rejection, fluid distribution, packaging, controls, and broader system integration all matter.

Our role is not just to provide hardware. It is to help define and build the real thermal infrastructure around the compute environment — from engineering through fabrication and delivery, in-house in Canby, Oregon.

Define the Thermal Architecture

Before hardware, we define the system: heat load, cooling strategy, CDU sizing, loop temperatures, and heat-rejection path.

Engineer and Fabricate In-House

Engineering and fabrication under one roof in Canby, Oregon — so decisions translate directly into what gets built.

Deliver a Deployable System

Packaged, tested, and documented for deployment — not a collection of components requiring field integration.

Frequently Asked Questions

AI Liquid Cooling Infrastructure — FAQs

Oregon Engineered — Canby, OR

Need Help Solving the Infrastructure
Side of AI Liquid Cooling?

Talk with MRX about your rack density targets, thermal architecture, heat-rejection path, fluid-distribution needs, and deployment goals. We can help define the right liquid-cooling infrastructure strategy from engineering through fabrication and delivery.