Summary
- STL Partners research reported by Data Centre Review puts liquid cooling retrofit cost at around $2m per MW.
- New greenfield liquid-cooled capacity can cost upwards of $11m per MW, according to the same research.
- Retrofit economics depend on power availability, downtime tolerance, live-site constraints, and the condition of existing electrical and mechanical plant.
STL Partners research has put a sharp number on the retrofit argument for AI data centres, suggesting that liquid cooling upgrades to existing facilities can cost around $2m per MW, compared with more than $11m per MW for new greenfield liquid-cooled capacity.
The figures support a growing view that AI capacity will not be delivered only through new campuses. Existing facilities with secured power, suitable space, and adaptable mechanical and electrical systems may offer a faster and cheaper route to high-density deployments.
The headline saving is large, although it depends heavily on the building. Liquid cooling retrofits involve more than installing new rack-level hardware. They can require changes to chilled-water systems, CDUs, pipework, containment, leak detection, controls, power distribution, floor loading, maintenance access, and operating procedures.
Facilities with available power allocations have a clear advantage, particularly in European markets where grid queues and planning processes are delaying new capacity. A site that already has energised infrastructure, resilient utility feeds, and operating permissions can be valuable even if its original design assumed far lower rack densities.
Power remains the hard limit
The economics of liquid cooling retrofit are strongest where power already exists. Without additional electrical capacity, a mechanical upgrade cannot create AI-ready megawatts. Operators may be able to concentrate load into fewer halls, free capacity by consolidating legacy IT, or support denser deployments within an existing electrical envelope. Where the electrical infrastructure is already fully committed, cooling alone cannot solve the constraint.
Downtime is the second major variable. Retrofitting a live data centre is not the same as fitting out an empty shell. Work has to be phased around customer environments, change-control windows, redundancy requirements, and operational risk. Pipework routes, hot works, water systems, and commissioning are all more difficult when production workloads are running nearby.
Commercial demand also has to be clear. Colocation providers need to know whether enough customers will pay for high-density capacity in a retrofitted facility, and whether contracts justify disruption and capital spend. Some enterprise workloads will continue to sit comfortably in lower-density air-cooled environments. AI-focused clusters require a different plant and service model.
Across Europe, the retrofit argument is strengthened by the scarcity of new power and land. Upgrading existing assets can reduce embodied carbon compared with new build, shorten delivery timelines, and make better use of scarce grid allocations. It can also become an expensive engineering exercise where older buildings have low floor-to-ceiling heights, limited plant space, inadequate risers, constrained floor loading, or legacy electrical layouts.
Existing powered assets are becoming more valuable as AI demand rises faster than grid reinforcement and planning systems can respond. The best candidates will be facilities with known plant condition, credible customer demand, and enough operational control to upgrade without turning live-site risk into the hidden cost of cheaper capacity.
