How to Plan AI-Ready Cooling for High-Density Racks: A Step-by-Step Guide for Data Center Engineers and Procurement Managers
August 10, 2026
Planning cooling for AI-ready, high-density rack deployments starts earlier than most teams expect. The decisions made in the planning phase determine whether your infrastructure holds up under AI workloads or becomes a bottleneck as density climbs: heat load per rack, cooling architecture, rack platform, power delivery, and monitoring.
Why AI Infrastructure Changes the Cooling Conversation
Traditional data center cooling was sized around average rack densities of 5 to 10 kW. Most facilities planned for that range, and room-level air cooling handled it well enough. AI infrastructure operates in a different category entirely. A single rack of GPU servers running AI training or inference workloads can generate 30 kW, 50 kW, 100 kW, or more, and that heat concentrates in a far smaller footprint than general-purpose compute ever required.
The infrastructure decisions that worked for traditional server deployments don’t carry over automatically. Rack platforms, power delivery, cooling architecture, and monitoring all need to be evaluated against a different set of requirements. To get this right, your organization needs to start planning long before a purchase order is placed.
This guide walks through that planning process step by step, from heat load assessment through cooling selection, rack infrastructure, and monitoring. Products are introduced where they’re relevant to a decision you’re already working through.
Step 1: Start With Heat Load
The most common planning mistake in high-density deployments is treating cooling as something to sort out after the servers are specified. By the time the hardware arrives, the options narrow considerably. Heat load is where the planning process should begin.
Per-rack heat load matters more than total facility load. A data center with 100 racks averaging 5 kW each and a data center with 10 racks at 50 kW each have the same total load on paper, but the cooling requirements are completely different. AI deployments concentrate heat in a small number of racks, which means the cooling solution must address localized thermal density per rack.
GPU servers and AI accelerator platforms also behave differently from general-purpose compute under production conditions. General-purpose servers cycle between idle and peak throughout the day. AI training workloads are among the most power-intensive in production, often pushing hardware to or near nameplate draw. Inference workloads are less intensive but still run at sustained power levels well above what general-purpose compute typically requires. An infrastructure plan sized for average utilization ends up undersized for either.
Future density growth belongs in the calculation, too. GPU generations turn over faster than data center infrastructure does, and the hardware deployed today will likely be replaced with higher-density successors within a few years. A plan with no headroom for that growth will need expensive retrofits sooner than most timelines account for.
Rittal’s RiTherm thermal calculation tool can help validate heat load assumptions at the individual cabinet level. Engineers can input specific hardware configurations and ambient conditions to determine the thermal output of a given enclosure before committing to a cooling approach. RiTherm is calibrated for enclosure-level analysis, so it’s most useful early in the process as a check on per-rack heat load assumptions.
For facility-level planning, Rittal Solution Architects provide engineer-to-engineer support from the earliest stage of a project. With hands-on expertise across data center infrastructure, their services are available at no charge for organizations that are seriously evaluating a Rittal deployment. Getting a Solution Architect involved before your hardware is specified often produces better outcomes than bringing one in after the key decisions have already been made.
Step 2: Match the Cooling Architecture to the Density
Heat load per rack determines which cooling architecture fits the application. Air cooling, rear-door heat exchange, and direct liquid cooling each serve a different density range, and selecting the wrong one for the application creates problems that compound as workloads grow.
Air Cooling
Room-level air cooling works well for rack densities up to roughly 10 to 15 kW. Above that, the volume of airflow needed to remove the heat becomes difficult to manage. Distribution becomes uneven, hot spots develop, and energy consumption climbs as more cooling infrastructure is added to compensate. For AI deployments running at 30 kW per rack or above, air cooling can’t carry the load as a primary strategy. It may still serve as supplemental cooling in a hybrid architecture, but it can’t stand alone.
Rear-Door Heat Exchanger
A rear-door heat exchanger (RDHEx) mounts to the back of the rack and uses a liquid circuit to capture heat as it exits the server exhaust before it reaches the room. It works with existing rack infrastructure and doesn’t require changes to the server hardware, which makes it a good option for facilities that need to increase cooling capacity without redesigning everything around it.
The ceiling is approximately 70 kW per unit, which covers moderate-density deployments but falls short of what dense GPU racks require. Rittal’s RDHEx hits that mark with N+1 redundant fans and a hot-swappable control module. Like air cooling, RDHEx can also provide supplemental cooling within a full DLC design.
Direct Liquid Cooling
When rack densities push above 100 kW, which is where GPU-dense AI racks routinely operate, direct liquid cooling (DLC) is the architecture that can keep up. DLC brings coolant directly to the heat source at the chip or server level. The thermal efficiency is higher than any air-based approach, and it’s why DLC has become the standard for the highest-density AI deployments.
Rittal’s DLC platform is a modular, OCP-compatible family of climate control options built around a common architecture. Some products in the platform install without tools, running hot-swappable components while carrying redundant pumps and fans, leak detection, and flexible manifolds throughout. The platform scales from rack-level to multi-megawatt deployments, and N+1 redundancy is built in throughout, not added after the fact.
A DLC deployment involves a few decisions that need to be settled early.
CDU placement is the first. A coolant distribution unit (CDU) manages the interface between the facility’s chilled water loop and the liquid cooling circuits at the rack level. In-rack CDUs fit within the rack footprint in a compact 4OU form and provide up to 150 kW of localized cooling per unit, suited for installations where space is constrained and rack-level control is a priority. In-row CDUs sit between rack rows and scale to much higher capacities: Rittal’s In-Row CDU grows beyond 1 MW via 250 kW incremental N+1 additions, which means cooling capacity can follow compute capacity without a full re-build each time density increases. For large AI clusters where the workload is expected to grow, the in-row configuration is where your planning should start.
Manifold design is the next consideration. A DLC deployment requires a manifold system to distribute coolant to individual servers inside the rack. Manifold design affects how easily the system can be reconfigured when hardware changes and how leak risk is managed. Rittal’s DLC platform uses leak-aware manifolds with tri-clamps and dripless connectors, which address the leak risk concern that comes up most often with liquid cooling.
Redundancy and serviceability round out the picture. AI workloads are unforgiving of unplanned downtime, and a cooling architecture that requires taking systems offline for maintenance is a liability in production. Rittal’s DLC platform is built around redundant pump units, fail-safe modes for pumps and fans, hot-swappable sensors and controllers, and component-level leakage monitoring.
Step 3: Select the Rack Platform Before Specifying Hardware
Cooling architecture and rack platform are interdependent decisions, but they’re often treated as sequential ones. Specifying hardware first and then finding a rack to fit it creates compatibility issues that show up at installation.
AI infrastructure increasingly uses hardware built around the Open Compute Project (OCP) standard, which uses a 21-inch form factor instead of the traditional 19-inch. The wider format enables more efficient component arrangement and higher component density, both of which matter for AI applications where packing more compute per rack is a core objective.
Rittal’s ORV3 rack is built to the OCP standard and available off-the-shelf in 44OU, which means fast, repeatable installation without long lead times. A few specifics come into play when planning your rack infrastructure:
The ORV3 integrates a 48V DC busbar that delivers power directly to components through the rack’s power rail system. The rack ships ready to deploy, with castors, power rails, and front-side cable routing already integrated. Most of the assembly is tool-free, using click-in mounting throughout. If your organization is standing up multiple rows of AI infrastructure on a compressed timeline, that installation efficiency compounds.
The ORV3 also accommodates 19-inch components through an optional adapter, which matters for facilities running a mix of OCP-compatible AI hardware and conventional server equipment in the same infrastructure. The rack is CE and UL certified.
If your current hardware isn’t OCP-compatible yet, Rittal’s broader IT enclosure portfolio provides options that work with the DLC cooling systems described above. A Rittal Solution Architect can help map the right rack platform to the specific hardware and cooling architecture being considered.
Step 4: Plan Power Delivery Alongside Cooling
Power delivery and cooling are connected in ways that aren’t always obvious until something goes wrong. The thermal load a rack generates is tied to the power it consumes, and the facility infrastructure required to support high-density AI racks (switchgear, PDUs, cabling, UPS) needs to be sized for the actual load before deployment, not once it’s running.
A row of ten racks at 100 kW each draws 1 MW from your facility. That power infrastructure should be planned at the same time as the cooling architecture.
The ORV3’s integrated 48V DC busbar is an important piece of this. AI server platforms that support 48V power input benefit from fewer voltage conversion steps between the utility supply and the hardware, which reduces conversion losses and the heat those losses generate. Less waste heat means less cooling load for the same compute output.
PDU selection also affects monitoring capability, and intelligent monitored PDUs provide per-outlet power measurement and environmental sensor support that feeds into the monitoring architecture. In a high-density AI deployment, per-rack power visibility is how operators correlate workload changes with thermal output and catch anomalies before they become incidents.
Step 5: Build Monitoring in From the Beginning
Monitoring is the part of the planning process that most commonly gets deferred, and in a high-density AI deployment, that tends to surface as an incident. Temperature excursions, coolant leaks, and airflow disruptions in a 100 kW rack happen faster and with more consequence than in a 5 kW rack. The monitoring architecture needs to be in place before the workloads go live.
The monitoring scope in an AI-ready deployment covers temperature at the rack inlet, inside the cabinet, and at the server exhaust; coolant flow and leak detection at the component level; airflow and pressure; power consumption per rack; and physical security including vibration and door access. In a DLC deployment, leak detection at the component level is the one that can’t be skimped on. A coolant leak in a high-density rack is a serious incident, and facility-level detection is too slow to prevent hardware damage.
Rittal’s CMC III monitoring suite is a rack-ready platform that centralizes this data and feeds it into existing DCIM or network management systems. The CMC III Processing Unit ships off-the-shelf and supports up to 32 direct sensor connections, covering temperature, humidity, airflow and pressure, leakage, smoke and fire, vibration, and door access. CAN-bus connectivity links multiple units into a centralized view, and DCIM compatibility means the data integrates with existing management infrastructure. Hot-pluggable sensors keep maintenance manageable in a production environment where downtime for sensor replacement isn’t acceptable.
Step 6: Build Scalability into the Design Before It’s Finalized
An AI-ready data center built around today’s workloads needs a path to accommodate the next hardware generation, and that path is easier to build in at the design stage than to add later. A few important things belong in your design conversation before anything is finalized.
- Cooling Scalability: Rittal’s In-Row CDU adds capacity in 250 kW N+1 increments, which means cooling can follow compute without a full infrastructure re-deployment each time density increases. That scalability is built into the architecture from the start.
- Rack Flexibility: When the next generation of GPU servers arrives with different power requirements or a different physical form factor, a standardized modular rack platform is easier to reconfigure than a purpose-built alternative. The ORV3’s accessory system and 19-inch adapter support that flexibility.
- Monitoring Headroom: The CMC III’s CAN-bus architecture networks multiple processing units together, so monitoring coverage can expand alongside the physical deployment without replacing the base platform.
Where Rittal Solution Architects Fit
Every step in this planning process involves decisions where the outcome depends on your specific workloads, facility constraints, hardware mix, and growth timeline. A framework like this one establishes the sequence. Getting the details right for your specific situation requires someone who has worked through these decisions across many projects.
Rittal Solution Architects are engineers. They understand data center infrastructure from the facility level down to the rack, and they work with organizations from the earliest stage of a project through deployment. Their support covers heat load assessment, cooling architecture selection, rack platform specification, power delivery planning, and monitoring design: the full scope of what this guide addresses.
Solution Architect services are available at no charge for organizations that are seriously evaluating a Rittal infrastructure deployment. There’s no requirement to have the design figured out before the first conversation—by relying on a Rittal Solution Architect early in the design process, you receive more support for the entire process.
The Planning Sequence in Summary
Heat load comes first. Per-rack thermal output that is sized for production conditions drives every decision that follows. Cooling architecture follows from that number: air cooling for lower densities, rear-door heat exchange for moderate densities, and direct liquid cooling for the high-density AI rack environments where air-based approaches can’t keep up.
The rack platform and cooling architecture need to be selected in parallel. Power delivery planning runs alongside both. Monitoring gets designed into the architecture from the start. Add to that room for design to grow, because the hardware that arrives in two years will be denser than today’s deployment.
