Before generative AI reshaped compute demand, data center cooling didn't require much strategic thought. Facilities sized air handlers to keep server inlets within a safe temperature range, added hot aisle and cold aisle containment, and moved on to other priorities. That approach worked because IT equipment, until recently, didn't generate enough heat to challenge it.
That's no longer the case. AI training, high-performance computing, and dense inference deployments are concentrating far more power into each rack than many existing facilities were designed to support. The result is a greater volume of heat concentrated in less space.
Conventional enterprise racks have commonly operated in the single-digit to low-teens kilowatt range, while modern AI clusters can drive rack densities into the tens of kilowatts and, in some cases, beyond 100 kW.
Cooling is not a minor line item, either. Depending on the facility’s design and operating conditions, cooling systems can account for a substantial share of total data center electricity use, with some estimates placing that share near 40 percent. The cooling approach a facility chooses can therefore affect both operating costs and sustainability goals.
The question for operators is no longer simply whether liquid cooling is coming. It is whether their facility can support the mix of power distribution, heat rejection, piping, controls, monitoring, maintenance practices, and residual air cooling that high-density IT will require. In many cases, the answer will not be an all-air or all-liquid architecture, but a carefully planned hybrid environment.
Why Liquid Cooling Is Gaining Ground
Liquid cooling encompasses several approaches. Rear-door heat exchangers can capture heat from a rack’s exhaust air and may be used in retrofit projects. Direct-to-chip systems circulate coolant through cold plates attached to processors and other high-heat components. Immersion cooling submerges IT equipment in a dielectric fluid and is generally reserved for specialized, very-high-density applications. Each approach presents different considerations for retrofit complexity, facility-water infrastructure, maintenance, and serviceability.
Adoption is accelerating as AI systems concentrate more computing power in each rack. Goldman Sachs has forecast that liquid-cooled AI servers will rise from roughly 15% of deployments in 2024 to 54% in 2025 and 76% in 2026, driven in part by next-generation, full-rack AI systems designed for liquid cooling.
The shift does not mean every facility must replace air cooling everywhere. Instead, many operators are adopting hybrid designs that use liquid cooling for the highest-density AI hardware while retaining air cooling for residual rack heat and conventional IT equipment. That approach can help facilities add high-density capacity without treating every workload as though it has the same thermal requirements.
The Water Question
Water use has become a prominent part of the public conversation around data centers, and it is worth addressing directly. The issue is often less about whether heat is collected by air or liquid at the rack and more about how that heat is ultimately rejected from the facility. Data centers that rely on evaporative cooling towers can consume significant amounts of water because a large share of the water used for heat rejection is lost through evaporation.
Liquid cooling can reduce a facility’s onsite water use when it is paired with a closed-loop design and water-saving heat-rejection equipment, such as dry coolers or air-cooled chillers. Direct-to-chip and immersion systems move heat efficiently from high-density IT, allowing facilities to operate at higher coolant temperatures and, in appropriate climates and designs, reduce dependence on evaporative cooling. Schneider Electric notes that closed-loop, air-cooled chillers can save millions of gallons of water annually for every megawatt of cooling capacity
Schneider Electric’s modeled data-center scenarios in Dallas, Texas, and Paris, France illustrate the potential impact of those design decisions. In the Dallas model, projected annual water consumption fell from 382,000 cubic meters in a traditional air-cooled scenario to 197,000 cubic meters in a closed-loop liquid-cooled scenario, a 48% reduction. In Paris, the projection fell from 108,000 to 51,000 cubic meters, a 53% reduction. Those are modeled, location-specific outcomes, not universal liquid-cooling results, but they show why cooling architecture matters.
Liquid cooling is not automatically water-neutral. A liquid-cooled facility may still use a water-cooled chiller plant, cooling towers, or another water-dependent heat-rejection system. The deciding factor is the complete cooling and heat-rejection architecture, not simply whether a facility uses air or liquid at the rack. For operators in water-stressed regions or areas with tighter water-use requirements, that distinction is increasingly a design requirement.
What Facility Readiness Actually Requires
None of this means air cooling is obsolete. Many enterprise workloads and lower-density colocation deployments remain well served by air cooling, and for many operators, the practical path forward is a hybrid environment rather than a wholesale replacement.
Reaching that point is rarely just a mechanical decision. It touches nearly every layer of facility infrastructure:
- Power density planning: Facilities need a clear view of which racks will support AI or other high-density workloads over the next three to five years, since retrofitting power and cooling infrastructure after deployment is typically more disruptive than planning for it up front.
- Liquid-distribution and facility-water infrastructure: Direct-to-chip, rear-door and immersion systems often require CDUs, piping, manifolds, pumps and leak-detection systems that many legacy sites were never built to accommodate.
- Residual air cooling and heat rejection: Liquid cooling rarely captures every source of rack heat, so facilities still need to account for air-cooled components, room-level airflow and how collected heat is ultimately rejected outside the data hall.
- Structured cabling and pathway design: CDUs, piping, manifolds and denser compute racks can reduce routing space and complicate rack access, so cabling pathways, fiber management and capacity for future connections should be planned alongside the cooling system, not added after installation.
- Serviceability and staff training: Liquid cooling introduces maintenance tasks and incident-response procedures, such as reading coolant flow and pressure data or responding to leak alarms, that teams may not encounter in air-cooled environments.
Facilities that treat cooling strategy and physical infrastructure planning as separate conversations tend to pay for it later, in rework, operational risk, unplanned downtime or capacity constraints they didn't anticipate.
Plan the Facility Around the Workload
For most data center operators, the decision is not whether to choose air cooling or liquid cooling forever. It is whether current and projected rack densities justify the additional infrastructure liquid cooling requires, and whether the rest of the facility, from power distribution and heat rejection to piping, pathways and connectivity, is prepared to support that transition.
Facilities that plan for high-density workloads before thermal limits become a problem will be better positioned to deploy AI infrastructure without costly rework, avoidable downtime or stranded capacity. In many cases, the most practical approach will be a hybrid environment that applies liquid cooling where density demands it while preserving air cooling where it remains effective.
Planning a high-density data center build or upgrade? Hexatronic Data Center supports projects from the design phase through the operational facility, with passive fiber connectivity solutions and services that help teams build scalable, maintainable data center infrastructure.
Talk with Hexatronic today about designing the fiber infrastructure and pathways your evolving data center will need.
