Liquid Cooling for GPU Data Centers: What Operators Need to Know
GB200 and GB300 NVL72 require liquid cooling. Learn CDU installation, manifold routing, pressure testing, and thermal commissioning for GPU deployments.
The latest generation of NVIDIA GPU platforms—GB200 NVL72 and GB300 NVL72—represent a fundamental shift in data center infrastructure requirements. These systems deliver unprecedented compute density, but that performance comes with thermal loads that make liquid cooling not just preferable, but mandatory. For data center operators planning GPU deployments, understanding what liquid cooling adds to your project scope is critical to avoiding costly delays and facility surprises.
Why GB200 and GB300 Require Liquid Cooling
The physics are straightforward: GB200 and GB300 NVL72 racks generate over 120kW of heat per rack. At this density, air cooling simply doesn't work. Traditional air-cooled infrastructure tops out around 20-30kW per rack before you hit fundamental limitations in airflow velocity, temperature differentials, and acoustic constraints.
To remove 120kW+ of heat with air, you would need airflow rates and temperature rises that are physically impractical in a data center environment. The volume of air required would create noise levels incompatible with human occupancy, and the temperature differentials would exceed what silicon can tolerate. Even if you could move enough air, the energy cost of the fans alone would be prohibitive.
Liquid cooling solves this by using water's superior thermal properties. Water has roughly 3,500 times the heat capacity of air by volume, meaning you can remove far more heat with much smaller flow rates and temperature differentials. This is why NVIDIA designed GB200 and GB300 platforms with direct liquid cooling from the start—it's the only viable path to the compute densities these systems deliver.
What Liquid Cooling Adds to Your Deployment Scope
Deploying liquid-cooled GPU infrastructure introduces five major work categories that don't exist in air-cooled deployments. Each requires specialized expertise, specific equipment, and careful coordination with facility systems. Understanding these categories upfront is essential for accurate project planning and budgeting.
CDU Installation and Facility Connection
The Coolant Distribution Unit (CDU) is the heart of your liquid cooling system. Each CDU contains a heat exchanger, pumps, filters, control systems, and monitoring equipment. It acts as the interface between your facility's chilled water loop and the secondary coolant loop that serves the GPU racks.
Typical deployments require one CDU per one to two racks, depending on the specific platform configuration and CDU capacity. Each CDU must be physically installed, connected to facility chilled water supply and return lines, integrated with facility power, and wired into monitoring systems. This isn't plug-and-play equipment—proper installation requires understanding of both the CDU's requirements and your facility's capabilities.
The facility connection is where many projects encounter their first surprise. Your chilled water system must have adequate capacity, deliver water at the correct temperature (typically 10-20°C), and provide sufficient flow rate and pressure. If your facility was designed for traditional air-cooled loads, you may need chilled water system upgrades before you can support liquid-cooled GPU racks.
Rack-Level Manifold Routing
Once coolant reaches the rack from the CDU, it must be distributed to individual server trays and network switches. This requires manifold systems with quick-disconnect fittings that allow components to be serviced without draining the entire cooling loop.
The routing is platform-specific and differs significantly between GB200 and GB300 configurations. Each platform has its own manifold design, fitting types, and routing paths. The manifolds must be installed with precise attention to fitting torque specifications, routing clearances, and service access requirements. Improper installation can lead to leaks, restricted flow, or inability to service components.
This work requires technicians who understand both the mechanical aspects of fluid systems and the specific requirements of the GPU platform. Generic plumbing experience isn't sufficient—the tolerances, materials, and procedures are specific to data center liquid cooling systems.
Leak Detection Systems
Liquid cooling introduces a risk that doesn't exist with air cooling: leaks. While modern liquid cooling systems are highly reliable, leak detection is a mandatory safety layer. Sensors must be installed at all critical points—under manifolds, at fitting clusters, beneath CDUs, and anywhere coolant lines run above or near equipment.
These sensors must be wired into your facility's alarm system with appropriate escalation procedures. A leak detection event should trigger immediate alerts to operations staff and, depending on severity, automatic shutdown procedures. The detection system isn't just about protecting equipment—it's about giving your team time to respond before a small leak becomes a major incident.
Installation of leak detection requires coordination between the liquid cooling team, facility electrical contractors, and your building management system integrator. All parties must understand the sensor specifications, wiring requirements, and alarm logic before installation begins.
Pressure Testing
Before any coolant enters the system, every circuit must be pressure tested. This involves pressurizing the cooling loop with nitrogen to a specified pressure (typically 1.5x operating pressure) and monitoring for pressure decay over a defined period, usually 24 hours.
Pressure testing validates that the entire cooling circuit is leak-free before you introduce coolant. Using nitrogen instead of coolant means that if there is a leak, you're dealing with an inert gas rather than liquid near expensive equipment. Any pressure drop during the test period indicates a leak that must be found and corrected before proceeding.
This step is non-negotiable. Skipping pressure testing to save time is one of the most common—and costly—mistakes in liquid cooling deployments. Finding a leak after coolant fill means draining the system, locating the leak, making repairs, and starting over. Finding it during pressure testing means fixing it once, correctly, before any coolant is involved.
Thermal Commissioning
Once the cooling system is filled and operational, thermal commissioning validates that everything works correctly under actual load. This means running the GPUs at full power while monitoring coolant flow rates, inlet and outlet temperatures, system pressures, and GPU junction temperatures.
Every parameter must fall within specification. If GPU temperatures are higher than expected, it could indicate insufficient coolant flow, incorrect coolant temperature, air in the cooling loop, or improper cold plate contact. If pressures are out of range, it might point to pump issues, flow restrictions, or incorrect manifold configuration.
Thermal commissioning isn't a quick check—it requires sustained full-load operation while systematically verifying every measurement point. This is where you prove that the entire cooling system, from facility chilled water through CDUs and manifolds to GPU cold plates, is functioning as designed. Cutting corners here means you won't discover problems until production workloads are running, which is exactly when you don't want surprises.
Facility Requirements for Liquid Cooling
Your facility must meet specific requirements before liquid-cooled GPU infrastructure can be deployed. These aren't optional upgrades—they're prerequisites for system operation.
Chilled Water Supply
Your facility's chilled water system must have adequate capacity to handle the heat load from your GPU deployment. For a 10-rack GB300 deployment generating 1.2MW of heat, your chilled water system must be able to remove that heat continuously. If your facility was designed for 20kW/rack air-cooled loads, you may need significant chilled water plant upgrades.
The chilled water must be delivered at the correct temperature, typically 10-20°C depending on your CDU specifications. It must also provide sufficient flow rate and pressure to meet CDU requirements. These parameters aren't negotiable—if your facility can't deliver them, the cooling system won't function correctly.
Chilled Water Piping
Chilled water supply and return lines must be routed to each CDU location. This often requires new piping installation, especially if your facility wasn't originally designed for liquid-cooled loads. The piping must be sized correctly for the flow rates required, properly insulated to prevent condensation, and installed with appropriate isolation valves for maintenance.
Piping routes must be planned to avoid conflicts with other infrastructure, provide adequate support, and allow for thermal expansion. This work typically requires coordination with facility engineering and may need to happen during scheduled maintenance windows if it affects existing systems.
Floor Space for CDUs
CDUs are substantial pieces of equipment, typically occupying several square feet of floor space each. They need to be positioned near the racks they serve, with adequate clearance for service access, and located where chilled water connections are practical. In dense deployments, CDU placement becomes a significant space planning consideration.
You also need to account for the secondary coolant lines running from CDUs to racks. These lines must be routed safely, with appropriate protection and support, while maintaining service access to both the CDUs and the racks.
Building Management System Integration
CDUs and leak detection systems must integrate with your facility's building management system (BMS). This provides centralized monitoring of coolant temperatures, flow rates, pressures, pump status, and alarm conditions. Integration requires coordination between the liquid cooling vendor, your BMS provider, and facility operations staff to ensure all monitoring points are correctly configured and alarm thresholds are appropriately set.
Drainage Plan
You need a plan for handling coolant in both routine and emergency situations. This includes drainage points for system maintenance, containment for potential leaks, and procedures for coolant disposal or recycling. Floor drains near CDUs and under manifold clusters provide a safety margin if leaks occur, but they must be positioned and sized appropriately.
The Deployment Partner's Role
A qualified deployment partner should handle everything from the CDU output to the rack and back. This includes manifold installation, all rack-level coolant connections, pressure testing, leak detection installation, and thermal commissioning. The facility is responsible for chilled water infrastructure up to the CDU input; the deployment partner owns everything downstream.
This division of responsibility is important because it creates a clear boundary between facility infrastructure and GPU deployment scope. Your deployment partner should have the expertise, tools, and experience to execute all liquid cooling work without requiring facility staff to become liquid cooling experts. They should also coordinate with facility teams on chilled water connections, BMS integration, and any facility modifications required.
Common Mistakes to Avoid
Several mistakes appear repeatedly in liquid-cooled GPU deployments. Avoiding these can save significant time and cost.
Ordering Hardware Before Confirming Facility Capacity
The most expensive mistake is ordering liquid-cooled GPU hardware before confirming your facility's chilled water system can support it. If you discover after hardware arrival that you need chilled water plant upgrades, you're looking at months of delay while expensive equipment sits idle. Always validate facility capacity before committing to hardware orders.
Skipping Pressure Testing
Pressure testing feels like it slows down the project, but skipping it is false economy. Finding leaks after coolant fill means draining the system, making repairs, and starting over—a process that takes far longer than proper pressure testing would have. Always pressure test with nitrogen before introducing coolant.
Treating Liquid Cooling as an Afterthought
Liquid cooling requirements must be integrated into project planning from the start. Treating it as a detail to be figured out later leads to schedule delays, budget overruns, and facility surprises. Engage with liquid cooling expertise during the planning phase, not after hardware is ordered.
No Leak Detection Validation
Installing leak detection sensors isn't enough—you must validate that they work and that alarms reach the right people. Test each sensor by simulating a leak condition and verifying that the alarm triggers correctly and reaches operations staff. An untested leak detection system provides false confidence.
Planning Your Liquid-Cooled GPU Deployment
Successful liquid-cooled GPU deployments require early planning, facility validation, and experienced execution. Start by assessing your facility's chilled water capacity and identifying any upgrades needed. Engage with deployment partners who have proven liquid cooling experience—this isn't the time for on-the-job training.
Build realistic schedules that account for all five work categories: CDU installation, manifold routing, leak detection, pressure testing, and thermal commissioning. Each requires time and expertise. Budget for facility modifications if needed—chilled water piping, BMS integration, and floor space for CDUs all have real costs.
Most importantly, understand that liquid cooling is not optional for GB200 and GB300 platforms. The physics of 120kW+ per rack make it mandatory. The question isn't whether to implement liquid cooling, but how to implement it correctly, on schedule, and without facility surprises.
Leviathan Systems handles liquid cooling integration in-house for GB200 and GB300 deployments—not subcontracted to third parties. Our engineering team manages CDU installation, manifold routing, pressure testing, leak detection, and thermal commissioning as part of our turnkey deployment service. We're currently executing active GB300 deployments with full liquid cooling integration. If you're planning a liquid-cooled GPU deployment and want a partner who owns the entire cooling stack, our team can help you navigate facility requirements, avoid common mistakes, and execute on schedule.