Air cooling has not become obsolete because AI racks got hotter. Liquid cooling has not become the obvious answer for every new data center either. In 2026, both statements are easy to find online, and both are too simple.
The real divide is density. For a conventional enterprise rack, air is still familiar, serviceable and economically hard to beat. As rack power climbs, however, the fans, airflow, containment and room-level cooling needed to keep air viable become progressively more demanding. Eventually the economics stop favoring “more air” and start favoring getting liquid much closer to the heat source.
Uptime Institute puts the industry's rough economic crossover in the 20–30 kW per rack range, while stressing that it is not a universal engineering limit. Its July 2026 analysis says typical rack densities are only now shifting toward about 10 kW, even as AI hardware pushes a much smaller part of the market toward dramatically higher levels. That distinction is worth keeping in view: the high-density frontier is moving very quickly, but most installed racks are not living at the frontier.
That middle region is important because there is no magic rack density at which air stops working on Monday and liquid becomes mandatory on Tuesday. Hardware design, server form factor, allowable inlet temperatures, redundancy, climate and the amount of floor space available all shift the crossover.
Air cooling wins when the workload does not need to be dense
Air cooling has several advantages that are easy to understate when the conversation is dominated by AI. Operators know how to design it, maintain it and recover from failures. A data hall itself provides a large thermal buffer, and redundant room-level air systems can often share that common environment without being tightly coupled to each individual server.
Uptime's April 2026 assessment is unusually direct on this point: direct liquid cooling remains concentrated in applications where air cooling is no longer practical, and Uptime does not expect liquid cooling to become truly mainstream across mission-critical enterprise IT during the next three years. Its reasoning is not that liquid cooling performs poorly. It is that many business applications have no compelling reason to accept the additional operational and resiliency complexity that comes with it.
There is also a less fashionable way to extend air cooling: do not pack the servers so tightly.
Uptime notes that larger server chassis can use larger, slower fans and more generous heat sinks. That can materially reduce fan power compared with highly compact 1U systems and can keep air cooling viable at higher processor power levels. In other words, some of the pressure to adopt liquid cooling is created by the decision to maximize compute density in the first place.
That trade-off is economic as much as technical. More floor space costs money. So does a liquid-cooling system. The cheapest design depends on which constraint is actually scarce.
At around 20–25 kW per rack, densification starts giving less back
Packing more IT into each rack produces obvious early savings. The same compute capacity can use fewer cabinets, a smaller data hall, shorter cable runs and fewer distribution components. But the savings do not continue at the same rate forever.
Uptime's 2026 cost analysis says the capital benefit of densification begins to thin out at around 20–25 kW per rack. Beyond that region, the savings from squeezing the IT footprint become smaller while high-density power distribution and thermal-management equipment start carrying their own premium.
That observation helps explain why the “densest possible rack” is not automatically the lowest-cost architecture. If floor space is cheap and available, spreading the same compute across more racks can sometimes preserve simpler cooling and reduce the amount of specialized infrastructure around each cabinet.
The useful question is not “How dense can this rack be?” It is “What density minimizes the cost and operational complexity of delivering the required compute?” Those are not always the same number.
Liquid cooling costs more to introduce today
For comparable U.S. projects, the public benchmark data is fairly consistent. Turner & Townsend estimates that high-density liquid-cooled data centers cost about 7%–10% more to construct than air-cooled facilities of similar IT capacity. Its current index also shows why: mechanical systems account for roughly 22% of the indexed cost mix in its air-cooled model and 33% in the liquid-cooled model.
Uptime arrives at a similar order of magnitude using a different methodology. It estimates that supporting direct liquid cooling in a new build can add roughly 5%–10% to capital expenditure, depending on the design and assumptions. Existing facilities are likely to pay more because the new system has to be fitted around an infrastructure that was designed for air.
The extra cost is not just a set of cold plates. Coolant distribution units, piping, manifolds, controls, commissioning and the supporting heat-rejection system all become part of the facility. Redundancy can add another layer, especially when the workload expects high availability.
Then there is the awkward middle period: air and liquid at the same time
Few operators can replace an entire installed IT estate in one move. A campus may add liquid-cooled AI halls while conventional servers remain on air for years. That creates a period in which both systems have to exist at meaningful scale.
Uptime specifically flags this as a near-term cost problem. Facilities are likely to oversize total air-plus-liquid cooling capacity while workloads migrate, rather than sizing one system down at exactly the same rate the other grows.
Existing enterprise and conventional cloud loads dominate.
Air remains substantial while liquid capacity grows around new dense clusters.
Liquid can dominate high-density zones without eliminating air from the campus.
That coexistence is one reason the early economics of liquid cooling can look worse than the long-term theoretical economics. A future facility designed around standardized liquid-cooled hardware may be able to simplify more of the air-side infrastructure. A 2026 facility usually cannot assume that future has already arrived.
Liquid cooling's strongest argument is capacity, not fashion
Where liquid cooling becomes genuinely compelling is when the workload cannot be delivered economically with air at the required density.
Current AI roadmaps are pushing that boundary quickly. Uptime says major AI hardware platforms are heading toward rack power levels above 200 kW. At that point, the question is no longer whether a few extra air handlers can solve the problem. Heat has to be captured much closer to the components generating it. Uptime's 2026 analysis describes this shift as a change from asking whether liquid cooling is needed to asking how much liquid-cooled capacity should be planned.
This does not mean 200 kW racks are about to become normal across enterprise data centers. They are an edge case with outsized influence because AI training campuses are being built at enormous scale. The average facility still looks much more conventional.
Higher density can offset some of the construction premium
Turner & Townsend makes an important qualification to its 7%–10% premium: liquid-cooled AI facilities can sometimes reduce the building footprint needed to deliver the same IT load. Very large AI campuses can also benefit from economies of scale.
Suppose one design spreads 20 MW of IT across 2,000 racks averaging 10 kW, while another can support the same 20 MW across 400 racks averaging 50 kW. That is not a realistic complete design comparison — redundancy, usable density and support equipment would all need to be modeled — but it shows why rack count and white-space requirement can move dramatically.
A higher mechanical budget can therefore coexist with a smaller building. Whether the total project becomes cheaper or more expensive depends on how much shell, land, distribution and support infrastructure can actually be removed.
Energy efficiency is more nuanced than “liquid uses less power”
Direct liquid cooling can eliminate or reduce server fan power and move heat using fluids that are far more effective than air. That creates real efficiency potential. But the economic result depends on the complete cooling chain.
Pumps, CDUs and heat-rejection equipment still consume power. A liquid-cooled system that requires low coolant temperatures may still rely heavily on chillers. A design operating at warmer fluid temperatures may open more opportunities for efficient dry cooling or economization.
The most reliable way to compare operating cost is therefore to model the resulting whole-facility PUE and electricity price rather than assigning a generic percentage saving to “liquid cooling.”
There is another subtlety: some fan power is counted inside the IT denominator of PUE. Removing server fans can reduce total electricity use while changing the ratio in ways that are less intuitive than a simple cooling-energy comparison suggests. PUE remains useful, but total kWh is ultimately what appears on the energy bill.
Resilience is easier to overlook until the cooling stops
One of the practical advantages of air is thermal inertia. A large room full of air does not instantly become unusable when one cooling component fails. At conventional densities, operators can have minutes of thermal ride-through while backup systems start or control systems recover.
Direct liquid cooling couples the cooling infrastructure much more closely to the IT hardware. Uptime notes that cold-plate systems at high density may tolerate loss of coolant circulation for only seconds, which can create requirements for UPS-backed pumps, thermal storage or other resilience measures.
This does not make liquid cooling unreliable. It means equivalent resilience can require a different design — and sometimes more cost — than teams accustomed to room-level air cooling expect.
Retrofit is where the economic comparison becomes least tidy
A greenfield project can place CDUs, piping, heat exchangers and plant capacity where they make sense from day one. An existing data center inherits whatever space, floor loading, pipe routes and heat-rejection equipment already exist.
That can turn an apparently straightforward cooling upgrade into a building project carried out around live IT. The retrofit may need temporary capacity, phased shutdowns, new structural support or a separate liquid loop that was never anticipated in the original design.
Uptime's estimate that existing facilities are likely to face a higher liquid-cooling premium than new builds is therefore unsurprising. The biggest cost is often not the cooling technology itself but the constraints around introducing it safely into a working facility.
Air and liquid cooling are also starting to split by workload
The most plausible near-term picture is not one technology replacing the other. It is different workloads pulling the cooling architecture in different directions.
AI training and HPC have a strong incentive to densify because expensive accelerators benefit from tight, high-bandwidth clusters and the workload can justify specialized infrastructure. Traditional databases, enterprise applications and many cloud services often care more about resilience, flexibility and operational familiarity than extreme density.
Uptime's 2026 view is that this split will persist. Direct liquid cooling will expand substantially because the AI/HPC niche itself is becoming enormous, while air cooling continues to support a large conventional market that has no urgent reason to change.
Three projects can reach three different answers
Consider a regional enterprise data center expecting mostly 5–12 kW racks. There is little reason to accept a liquid-cooling premium simply because AI exists elsewhere in the industry. A well-designed air-cooled facility can remain the lower-risk and lower-cost option.
Now consider a new colocation hall expected to host a mixed customer base, including some 30–50 kW racks. A hybrid design becomes much more interesting. The operator may preserve air cooling for conventional customers while creating liquid-ready zones or reserving plant capacity for future deployment.
Finally, consider an AI training campus designed around racks well above 50 kW and a roadmap toward much higher density. There, designing around air first and planning to retrofit liquid later can be the expensive choice. The workload itself has already decided much of the cooling architecture.
The decision should start with the workload, not the cooling technology
A sensible comparison asks what the facility must support over its life: current rack density, likely future density, uptime requirements, building constraints, electricity and water conditions, and how quickly the IT estate will turn over.
Only then does the air-versus-liquid question become useful. Air cooling tends to win where density remains moderate and operational familiarity has high value. Liquid cooling tends to win where density makes air costly, inefficient or physically impractical. Hybrid infrastructure makes sense when the workload mix genuinely requires both — not because “liquid-ready” sounds modern in a project brief.
The 2026 market is not moving from air to liquid in one straight line. It is splitting. Conventional compute still gives air cooling a large and durable role, while AI is creating a high-density market where liquid cooling is becoming basic infrastructure. The cost question is therefore less about choosing a winner and more about avoiding the wrong cooling system for the workload you actually have.