How to Plan Liquid Cooling for AI Infrastructure

Liquid cooling planning begins before a server order. High-density AI systems combine compute, networking, power delivery, and thermal management so tightly that a facility can have enough floor space yet still be unable to operate the equipment. The useful planning question is therefore not whether liquid transfers heat better than air. It is whether the entire path from silicon to outdoor heat rejection can carry the intended load safely, continuously, and serviceably.

A credible plan connects five decisions: the rack heat load, the cooling architecture, the building's mechanical and electrical limits, the reliability model, and the evidence required before deployment expands. Treating any one of these as a separate procurement exercise creates integration risk.

Translate the compute plan into a heat budget

Nearly all electrical power consumed by a rack ultimately becomes heat. Start with the expected rack configuration and power envelope, including accelerators, CPUs, switches, optical modules, storage, and power-conversion losses. Do not plan around a chip's headline rating alone. Leading accelerators can exceed one kilowatt, while complete rack-scale systems can reach hundreds of kilowatts. Density turns component heat into a facility problem.

Model at least three conditions: normal sustained operation, the highest credible workload, and a degraded cooling state. Include growth if later equipment may occupy the same row or loop. The objective is not merely to remove average heat but to keep temperatures, flow, and pressure within approved ranges when workloads vary or one part of the system is unavailable.

Also separate the heat captured by liquid from residual heat released to air. Direct-to-chip designs may cool processors and selected networking components while memory, storage, power hardware, or auxiliaries still warm the room. Some newer rack designs extend liquid connections across compute and switching trays, but that does not make the data hall's air-management needs disappear automatically. The design documents for each actual system should determine the split.

This heat budget becomes the common input for rack layout, pipe sizing, CDU selection, power distribution, controls, and outdoor equipment. If those teams use different assumptions, the site can pass individual design reviews and still fail as an integrated system.

Choose an architecture by boundary, not label

“Liquid cooling” describes several architectures, not one interchangeable product. Direct-to-chip cooling places cold plates against heat-producing devices. A server-side loop carries heat through hoses and manifolds to a coolant distribution unit, or CDU. The CDU manages flow, pressure, temperature, and separation between the equipment circuit and the facility circuit. The building loop then carries the heat to a dry cooler, cooling tower, chiller, or another rejection system.

Immersion cooling instead places equipment in a dielectric fluid. Its fluid, service procedures, component compatibility, and environmental considerations differ from those of a water-based cold-plate loop. A market forecast or supplier claim about “liquid cooling” is incomplete unless it identifies the architecture.

For each option, draw the full thermal chain and mark ownership at every boundary. Identify who specifies the coolant, who maintains its chemistry, who supplies the CDU controls, who can isolate a rack, and who is responsible for a leak or contamination event. Ambiguous ownership is especially dangerous at the interface between IT equipment and facility systems.

Published platforms can provide useful design references without becoming universal templates. NVIDIA's GB200 NVL72 system connects 36 Grace CPUs and 72 Blackwell GPUs in a liquid-cooled rack-scale design. The company has also shared rack mechanics and cooling specifications through an open hardware contribution. These examples can reduce interface uncertainty, but every selected server, facility, and operating environment still needs validation.

Test the building before selecting components

A CDU is a bridge, not a substitute for facility capacity. The site must accept the incoming heat and reject it outdoors. Survey available pipe routes, floor loading, service clearances, electrical feeds, backup power, controls integration, and the capacity of existing mechanical equipment. Retrofitting an air-cooled hall may require changes far beyond adding pipes beside a rack.

Supply temperature affects the facility design. A vendor design that accepts coolant entering at up to 45 degrees Celsius can create opportunities for chiller-free heat rejection with dry coolers. NVIDIA describes this warm-water approach, but it is a capability, not a result that applies everywhere. Climate, seasonal weather, approach temperatures, workload, and existing equipment determine whether chillers or cooling towers remain necessary.

Evaluate water and energy claims at the whole-site boundary. A closed loop may reduce direct water use, particularly when paired with dry coolers, yet the outcome varies by location and design. Account for pumps, fans, chillers, treatment, and residual room cooling. Comparing isolated component efficiency can hide energy that moved elsewhere in the system.

Electrical and thermal capacity must be commissioned together. Dense racks may need larger feeds, busbars, distribution gear, and backup systems. A cooling design that is ready months before the electrical system does not create usable compute capacity, and the reverse is equally true.

Engineer failure containment and maintenance

Liquid introduces pumps, seals, connectors, valves, sensors, and fluid quality into the availability model. The goal is not to claim that leaks never happen; it is to prevent a single fault from spreading and to detect abnormal conditions before equipment is damaged.

Define redundancy for CDUs and pumps based on the failure domain. Determine whether one CDU serves a rack, a row, or a larger group, then model what happens during maintenance and failure. Include isolation valves, bypass paths, spare capacity, and the control behavior after loss of power or communications. A nominal N+1 label is not enough if several components share the same controller, electrical feed, or facility loop.

Ultra-quick disconnects should limit fluid release during installation and service, but their evidence must cover repeated mating, temperature changes, vibration, pressure variation, and long operating periods. Maintenance procedures should prevent particles and trapped air from entering small passages. Technicians need defined steps for draining, reconnecting, purging, verifying flow, and returning equipment to service.

Monitoring should include supply and return temperature, flow, pressure, pump condition, and moisture or leak signals where appropriate. Set alert thresholds and automated responses before production. A completed maintenance action should be verified through both cooling telemetry and restored system health, not merely a closed work order.

Treat materials as a system-level compatibility decision

Cold plates, connectors, hoses, manifolds, seals, and coolant must remain compatible across the expected temperature, pressure, and service life. Cold-plate performance depends on channel geometry, bonding quality, pressure loss, surface contact, and manufacturing consistency. Microchannel designs can improve heat transfer near the device, but narrow passages increase sensitivity to particles, corrosion, bonding defects, and uneven flow.

Ask suppliers for the qualified material set rather than approving each component independently. The fluid specification should address cleanliness, corrosion control, thermal stability, handling, and replacement intervals. Changes to a coolant or wetted material should trigger compatibility review because a locally acceptable substitution can affect seals, metals, or deposits elsewhere in the loop.

Keep direct-to-chip and immersion-fluid risks separate. Water-based coolants are common in cold-plate systems. Some immersion systems have used fluorinated dielectric fluids. 3M completed its PFAS manufacturing exit at the end of 2025, which is relevant to affected fluid supply chains but does not imply that every liquid-cooled rack depends on those fluids. Procurement should document fluid composition, availability, environmental handling, and end-of-life plans for the chosen architecture.

Demand deployment evidence, not just capacity claims

High order volumes show interest, not operating reliability. Component readiness progresses through design, samples, performance tests, production testing, customer qualification, repeat manufacturing, and field operation. These stages should not be collapsed into a single claim of availability. Precision parts can take much longer to qualify than assembly capacity takes to expand.

Request evidence with denominators and operating context. Useful measures include production yield, leak-test results, pump durability, connector mating cycles, pressure and temperature ranges, coolant maintenance, field incidents, repair time, and performance across seasons. Distinguish small-batch shipments from qualified volume and recognized revenue.

Forecasts need similar discipline. TrendForce projected that liquid cooling penetration among AI chips would rise from about 33% in 2025 to 53% in 2026 and approach 60% in 2027 in its market outlook. Those figures describe a forecast for AI chips, not observed adoption across all servers or data centers. Use projections for scenario planning, then size commitments around actual platform schedules, construction readiness, and qualification results.

Pilot at the smallest meaningful failure domain. Run representative workloads, maintenance procedures, failover events, and abnormal-condition tests. A successful thermal demonstration proves heat transfer under its test conditions; it does not establish multi-season uptime, serviceability, or supply continuity.

Implementation checklist

Before approving a design or expanding a pilot, confirm the following:

• Heat load: Rack power, peak conditions, residual air load, growth, and degraded-state assumptions share one documented model.
• Architecture: The server loop, CDU, facility loop, and heat-rejection path are drawn end to end, with fluids and operating ranges identified.
• Facility fit: Pipe routes, floor loading, electrical distribution, backup power, controls, clearances, climate, water, and outdoor equipment have been surveyed.
• Failure domains: CDU, pump, controller, power, and facility-loop dependencies are mapped; isolation and redundancy work during planned maintenance.
• Materials: Coolant, metals, seals, hoses, connectors, and cold plates are qualified as a set, with contamination and corrosion controls.
• Service: Staff can connect, isolate, drain, purge, inspect, and restore equipment using documented procedures and available spares.
• Monitoring: Temperature, flow, pressure, pump, and leak signals have thresholds, owners, escalation paths, and tested responses.
• Evidence: Supplier claims identify test conditions, sample size, failures, production stage, qualification status, and field duration.
• Commissioning: Thermal load, failover, loss of power, alarms, maintenance, and recovery are tested with compute and facility teams together.
• Expansion rule: Scaling depends on uptime, incident, maintenance, energy, water, and capacity evidence rather than rack delivery alone.

A sound liquid-cooling program treats the rack and building as one operating system. Density establishes the heat problem; architecture defines the boundaries; facility engineering carries the load; materials protect the loop; and measured deployment results determine when to scale. That sequence turns cooling from a late mechanical add-on into a controlled infrastructure decision.