Liquid cooling for AI data centers is the use of a liquid coolant — rather than air — to remove heat directly from the high-density processors that AI workloads demand, and it has shifted from a niche technique to a near-necessity as rack densities climb past the limits of air. The reason is physical: liquid carries far more heat per unit volume than air, so when an AI rack draws 80, 100, or over 130 kW, only a liquid can remove that heat from such a concentrated source. Two approaches dominate. Direct-to-chip (DLC) cooling circulates coolant through cold plates mounted on the processors, capturing the bulk of the heat at its source while air handles the rest of the rack; it integrates with conventional facilities relatively readily and is the most common path for AI deployments. Immersion cooling submerges entire servers in a dielectric fluid, capturing essentially all the heat and enabling the highest densities, but requiring a fundamentally different facility and operating model. Either way, liquid cooling changes the facility: it introduces coolant distribution units (CDUs), fluid loops, leak detection, and new mechanical and structural requirements, and it changes how heat is ultimately rejected to the environment. Designing it well is now central to whether a data center can host the most valuable AI compute at all.
Axiom Utility Solutions engineers liquid cooling and the supporting mechanical and electrical systems with the constructable focus dense facilities require. This guide explains the liquid cooling approaches and their facility impact.
What Are the Main Liquid Cooling Approaches?
Two approaches account for most AI liquid cooling, with a hybrid middle ground. Direct-to-chip (DLC) cooling brings coolant to cold plates mounted directly on the CPUs and GPUs, where most of the heat is generated. The coolant absorbs that heat and carries it to a coolant distribution unit, which transfers it to the facility’s chilled-water or condenser-water loop. DLC captures roughly 70–80% of the rack’s heat in liquid, with air cooling handling the remainder, and it can be retrofitted into many existing facilities more readily than immersion, which is why it is the dominant approach for current AI deployments.
Immersion cooling submerges whole servers in a tank of dielectric (non-conductive) fluid, either single-phase (the fluid stays liquid) or two-phase (the fluid boils and condenses). Immersion captures essentially all the heat in liquid, enables very high densities, and can be highly efficient, but it requires purpose-built tanks, different server hardware handling, and a distinct operating model, so it is less common but growing. The choice between DLC and immersion depends on density, the facility’s existing infrastructure, and the operator’s appetite for a new operating model — a decision that belongs in the mission-critical design conversation early, because it shapes the whole mechanical system.
How Does Liquid Cooling Change the Facility Design?
Adopting liquid cooling reshapes the mechanical and structural design of a data center. The coolant distribution unit (CDU) becomes a key piece of equipment, transferring heat from the technology coolant loop to the facility loop and isolating the two; CDUs must be sized, located, and made redundant to match the load and reliability tier. The facility needs fluid distribution piping to and from the racks, which introduces routing, manifold, and connection requirements that air systems do not have. Leak detection and containment become essential, since liquid near energized electronics demands careful engineering to manage the (small but real) risk. Structural loading rises, particularly for immersion tanks filled with fluid, requiring floor capacity that air-cooled designs do not.
Liquid cooling also changes heat rejection. Because liquid captures heat at higher temperatures than air systems typically use, liquid-cooled facilities can often reject heat more efficiently and make greater use of free cooling, improving PUE. But the design must integrate the liquid loops, CDUs, and heat rejection coherently with the electrical system and the mission-critical facility as a whole. This is why liquid cooling is not a bolt-on but a design decision that ripples through the mechanical, structural, and electrical engineering, and why it must be engineered by a team that sees the whole system.
How Is Liquid Cooling Engineered, Step by Step?
Engineering liquid cooling for an AI facility follows a defined sequence integrated with the broader design.
1. Establish density and heat load. Determine the rack densities and total heat the facility must reject, which drives the cooling approach.
2. Select the approach. Choose direct-to-chip, immersion, or a hybrid based on density, existing infrastructure, and operating model.
3. Design the coolant loops. Engineer the technology coolant loop, the facility loop, and the CDUs that couple them, with the required redundancy.
4. Engineer distribution and connections. Design the piping, manifolds, and rack connections, including the routing that liquid demands.
5. Design leak detection and containment. Engineer detection, isolation, and containment to manage the risk of liquid near energized equipment.
6. Address structural needs. Account for the added floor loading, especially for immersion tanks.
7. Integrate heat rejection and efficiency. Couple the liquid system to efficient heat rejection and free cooling, optimizing PUE without compromising reliability.
8. Coordinate and commission. Integrate with the electrical design and verify performance, including failure scenarios, through commissioning.
The throughline is integration: liquid cooling touches mechanical, structural, and electrical systems, so it must be engineered as part of the whole facility rather than in isolation.
What Should You Look For in a Liquid Cooling Design Partner?
Because liquid cooling is both essential for AI density and consequential for the whole facility, the design partner should bring integrated, current expertise. Look for a team experienced with both direct-to-chip and immersion, since the right choice is project-specific and a partner biased to one may not give objective counsel. Look for the ability to integrate liquid cooling with the electrical and structural design, since it touches all of them. Look for rigor on leak detection and containment, the area where liquid cooling’s risk concentrates. Look for fluency in heat rejection and efficiency, since liquid cooling’s PUE benefit is a major part of its value. And look for engineers grounded in constructability and operation, not just the cooling concept.
Axiom Utility Solutions engineers liquid cooling as an integrated part of mission-critical design, with the practical focus the brand is built on. The value is a facility that can host the densest AI compute reliably and efficiently, with the cooling engineered to work where the work gets done.