A practical framework for selecting the right cooling strategy. Understand how density, scale, and facility design shape your liquid cooling strategy. Selecting the right CDU architecture is one of the most important decisions in designing liquid-cooled data center infrastructure. It impacts performance, efficiency, scalability, and long-term cost.
Why Architecture Matters
CDU architecture is not a standalone decision – it must align with your broader data center systems, including facility water, controls, and compute infrastructure.
Choosing a CDU is not just about capacity – it’s about how cooling is delivered across your infrastructure.
The architecture you choose will determine:
- how efficiently heat is removed
- how easily your system can scale
- how complex deployment and operations will be
Making the right decision early can reduce risk, improve performance, and avoid costly redesigns later.
The 4 Key Decisions
1. Density (Thermal Load)
As rack densities increase, cooling requirements change dramatically.
- <100 kW per rack
Early-stage or hybrid environments - 100–300 kW per rack
Typical AI and HPC deployments - 300 kW+ per rack
Ultra-high-density workloads requiring advanced cooling strategies
Higher densities require:
- increased flow rates
- tighter thermal control
- optimized system design
2. Scale (Deployment Size)
Your architecture should align with how large your deployment is today and where it’s going.
- Single rack / pilot
Fast deployment, minimal infrastructure - Row-level deployments
Shared cooling across multiple racks - Facility-scale
Centralized systems supporting large AI clusters
As scale increases, centralized approaches often become more efficient and easier to manage.
3. Facility Constraints
Your existing or planned facility plays a major role in determining the right architecture.
Key considerations include:
- Availability of water
- Space and floor loading constraints
- Integration with existing systems (BMS/DCIM)
- Retrofit vs new build
For example:
- Retrofit environments often favor flexible, perimeter systems
- New builds allow for optimized, centralized designs
4. Growth Strategy
AI infrastructure rarely stays static.
Your cooling architecture should support:
- phased expansion
- modular growth
- future increases in rack density
Choosing a scalable architecture early can prevent overprovisioning and reduce long-term cost.
CDU Architecture Types
In-Row CDU
Installed close to IT equipment within the rack row.
Best for:
- high-density deployments
- precision cooling
Trade-offs:
- consumes white space
- less flexible once deployed
Perimeter CDU
Installed outside the rack row with flexible placement.
Best for:
- retrofit environments
- mixed cooling deployments
Trade-offs:
- longer piping runs
- slightly reduced efficiency
Centralized CDU
Deployed at facility level to support multiple rows or zones.
Best for:
- hyperscale and AI factory deployments
- large-scale infrastructure
Trade-offs:
- higher design complexity
- requires upfront planning
Trade-Offs to Consider
Every architecture involves trade-offs.
| Priority | Best Fit |
|---|---|
| Speed | CDU-1MW |
| Flexibility | GigaModular |
| Efficiency at scale | GigaModular |
| Scalability | GigaModular |
Understanding these trade-offs helps ensure your design aligns with your operational goals.
How This Maps to LiquidStack
LiquidStack solutions are designed to support different architectural needs.
CDU-1MW
- Ideal for fast deployment
- Supports in-row and perimeter configurations
- Best for retrofit and expansion projects
GigaModular CDU
- Centralized control with flexible placement — in-row, perimeter, or skidded mechanical corridor
- Enables modular, large-scale growth
- Ideal for hyperscale and AI factory deployments
Next Step
Choosing the right architecture is just the first step.
Use our guided selector to identify the best solution for your environment.
