As AI rack densities blow past what air cooling can handle, operators are stuck between two hard truths: greenfield liquid-cooled capacity takes too long to build, and communities are pushing back on data center water and energy use. Retrofitting existing facilities for closed-loop liquid cooling is emerging as the pragmatic middle path — cutting water consumption by roughly half while unlocking AI-ready capacity in months, not years. But a retrofit is not a simple swap; it's a coordinated overhaul of power, cooling, and live-site operations. This post breaks down when retrofitting makes strategic sense, what it costs, and how to execute one without disrupting the facility you're trying to upgrade.
Every operator racing to deploy AI capacity is caught in the same bind. Rack densities are climbing past the point where air cooling can keep up, but building new liquid-cooled capacity from the ground up can take years—time most operators don't have in today's speed-to-power arms race. At the same time, data centers are under real public scrutiny for how much water and energy they consume, with community pushback and regulatory attention becoming a normal part of siting conversations. Retrofitting existing facilities for closed-loop liquid cooling is emerging as the answer to both problems at once. Here's what the evidence says about whether—and how—to do it.
Retrofitting is urgent because AI rack densities have outrun what air cooling can physically remove, and greenfield construction can't come online fast enough to meet demand. Above roughly 40kW per rack, air-cooling systems hit a "threshold effect" where they can no longer safely remove heat, and interviewees in one industry report cited GPU performance losses of up to 40% from thermal throttling on air-cooled racks. With major cloud providers moving to phase out new air-cooled AI racks, the pressure to act is not theoretical.
The strategic case is strongest where power is hardest to get. In power-constrained markets like Northern Virginia, utilities have already had to pause new connection requests, pushing operators toward upgrading facilities where power is already secured rather than waiting years for new grid capacity.¹ Retrofitting existing, powered-shell facilities lets operators repurpose power and building envelopes they already have — a meaningfully faster route to AI-ready capacity than permitting and building new.
A facility assessment is the right starting point for any operator weighing this decision — before committing capital, you need a clear, data-backed picture of where your facility actually stands against retrofit readiness criteria.
Yes — closed-loop liquid cooling can cut a data center's annual water consumption by roughly half compared to traditional air cooling, because the cooling fluid stays sealed in the system instead of evaporating. Traditional evaporative cooling towers lose water continuously to the atmosphere to remove heat. Closed-loop systems circulate the same fluid indefinitely, rejecting heat to outside air or a radiator instead of evaporating water into it.
Schneider Electric's Tuan Hoang made this point directly in a recent DCD Talks conversation on closed-loop cooling, arguing that water consumption in AI data centers is fundamentally a design choice, not an inevitable cost of doing business. In modeled comparisons of a Dallas, Texas facility and a Paris, France facility, transitioning from air to liquid cooling cut projected annual water use by 48% and 53% respectively.² As Hoang put it, "It's a choice to how you reject the heat"² — not a fixed trade-off between AI performance and community water resources.
This matters because water usage has become one of the most visible flashpoints in public pushback against new data center development. A retrofit that meaningfully cuts water draw isn't just an operational upgrade — it changes the conversation with the communities and regulators a facility depends on.
Closed-loop and direct-to-chip liquid cooling typically reduces cooling-related energy consumption by 30% to 60% compared to air cooling, because moving heat through liquid is far more efficient than moving it through air. Cooling already accounts for roughly 30% to 40% of a typical data center's total energy consumption, so improvements here have an outsized effect on overall facility efficiency.³
The efficiency gain compounds with density. As GPU thermal design power climbs, air cooling requires increasingly aggressive—and energy-intensive—fan and chiller operation just to keep pace, while liquid cooling removes heat directly at the source with far less parasitic energy loss. For facilities carrying legacy air-cooled infrastructure, this is often the single biggest lever available for reducing both operating costs and carbon footprint without waiting for a full facility rebuild.
Retrofitting existing capacity for liquid cooling costs roughly $2 million per MW, compared to $11 million or more per MW for greenfield liquid-cooled construction — making retrofit the more capital-efficient path when the underlying facility and power are already sound. This is according to a 2026 industry report from STL Partners, prepared on commission from Airedale, based on interviews with data center operators, developers, and cooling specialists across the sector.⁴
That $2 million figure is the hard cost only — equipment, materials, installation, and associated electrical work. It doesn't capture the operational risk of upgrading a live facility: lost tenancy revenue during the upgrade, customer workload migration, and the disruption of converting revenue-generating air-cooled space into liquid-cooled capacity. Those hidden costs are real, which is why the same report frames greenfield as the strategic ideal when time and power allow—retrofit is the pragmatic choice when they don't.
Retrofit projects also vary widely in scope. Light retrofits (rack-level power and floor changes) can run three to four months with minimal disruption; a full liquid cooling conversion is a heavy retrofit that can take eight months or more in a live facility, depending on scale and sequencing.⁴
No—retrofit viability depends on two factors: how much pressure the market is putting on you (through AI demand and power constraints), and your facility's operational archetype. A single-tenant hyperscale facility with strong internal demand visibility is a far stronger retrofit candidate than a multi-tenant retail colocation facility, where tenant coordination and lease timing make upgrades much harder to execute.
According to the STL Partners framework, operators generally fall into one of three postures:⁴
Facility age and design also matter. Data centers built from roughly 2015 onward, with modular "removable panel" construction, are typically far easier to retrofit than earlier facilities built with rigid concrete structures that weren't designed with future infrastructure changes in mind.⁴ An honest assessment of your facility against these criteria — power availability, structural flexibility, and customer demand quality — should come before any retrofit commitment, not after.
A liquid cooling retrofit is a coordinated overhaul across power distribution, cooling strategy, structural capacity, and live-site operations — not a simple equipment swap. Operators need to work through a clear technical checklist before and during physical works:⁴
Assessment phase:
Execution phase:
Floor loading is frequently the detail that catches operators off guard — AI racks can weigh up to five times more than standard server racks, and many older facilities simply weren't built to carry that load without reinforcement.⁴
Retrofits succeed or fail based on execution discipline — phased sequencing, tight vendor coordination, and clear risk protocols for the parts of the facility that stay live during the upgrade. Most retrofits happen in facilities that remain at least partially operational, which means the retrofit plan has to account for what stays running just as much as what's being upgraded.
This is where structured project management makes the difference between a retrofit that protects uptime and one that puts customer workloads at risk. It requires the same discipline Steadfast Operations brings to every mission-critical upgrade: risk-based sequencing, real-time tracking, and contingency planning built in from day one — not bolted on after something goes wrong. Across the facilities we've supported, this kind of structured incident and risk management has driven a 58% reduction in operational incidents, a track record documented in our case studies.
The retrofit itself is also only the beginning. Ongoing fluid management, monitoring, and operational oversight are what determine whether a facility actually delivers reliably at higher densities over time — which is why retrofit should be treated as an operational program, not a one-time project.
The speed-to-power arms race isn't slowing down, and neither is public attention on data center water and energy use. Retrofitting for closed-loop liquid cooling gives operators a way to answer both pressures at once — but only when it's grounded in an honest read of the facility, the market, and the execution risk involved.
With 20+ combined years of experience in data center operations, Steadfast Operations has helped operators move through exactly this kind of decision — from the initial assessment that determines retrofit readiness, through project management that protects uptime during execution. We've prevented 170+ issues before they became incidents across the facilities we support, and we bring that same proactive discipline to every retrofit engagement. Explore our full solutions or review our case studies to see how it plays out in practice.
.png)
.png)