Liquid cooling: Fab or fad?

For most organisations, the default design choice for new builds and upgrades remains air-cooled hot/cold aisle containment. It’s familiar, cost-effective, and—done well—highly efficient. But with rack densities rising and AI/HPC moving from the lab into mainstream production, many operators are asking whether liquid cooling has finally moved from hype to practical reality.

From our project teams’ experience across enterprise, public sector and edge deployments, the answer is: liquid cooling is now a mature option with clear, real-world benefits when applied in the right places—and it can be integrated pragmatically alongside existing air systems. The decision is less “either/or” and more “where/when.”

Why interest is surging

Three forces are converging. First, density. The compute per rack you can pack into a cabinet today is extraordinary, driven by GPUs and high-core CPUs for AI and analytics. Traditional air systems can struggle to keep junction temperatures within spec as heat flux concentrates, particularly in retrofits where space and airflow are constrained.

Second, energy and climate pressure. Operators have been navigating volatile energy prices and record summer temperatures—the UK’s all-time high of 40.3°C was set in 2022—which stress cooling plants and expose inefficiencies in control logic and maintenance regimes. And this isn’t a blip: provisional Met Office statistics indicate summer 2025 will almost certainly be the UK’s warmest on record, surpassing 2018, a sign that sustained warmth—not just short spikes—will push cooling systems harder for longer. 

Third, the refrigerant transition. The phasedown of high-GWP HFCs under F-Gas regulation has increased costs and complexity around legacy gases (such as R410A) and replacements, prompting some operators to rethink strategy rather than chase like-for-like swaps. This is an opportunity to plan a phased shift that improves density and future-proofs cooling for the next decade.

What “liquid cooling” really means

Liquid cooling is not monolithic; two mainstream approaches are now proven in live environments:

With Direct-to-chip (D2C) cooling, cold plates draw heat away from CPUs/GPUs (and sometimes memory), carrying it via a coolant loop to a cooling distribution unit (CDU) and heat-rejection plant. Because heat is removed at source, racks can be far more densely populated without dumping heat into the aisle.

Alternatively, with immersion liquid cooling servers (appropriately prepared) are submerged in a dielectric fluid. This enables very high density and quiet operation, though maintenance methods and hardware choices differ from standard rack practice.

The physics is on your side: water has orders of magnitude more volumetric heat capacity than air, so moving heat with liquid is inherently more efficient than pushing large air volumes through restrictive paths.

Where liquid cooling shines

From a design perspective, there are clear scenarios where liquid cooling often wins:

• AI/HPC pods inside general-purpose rooms. A handful of 20–80 kW racks can be handled elegantly as a “pod-within-a-room” using D2C, while the rest of the hall remains air-cooled. This avoids re-engineering the entire plant to cope with short-term hotspots.

• Space- and noise-constrained locations. Micro data centres at the edge—in plant rooms, cupboards, or on manufacturing floors—benefit from compact, high-density cooling that doesn’t require large air paths or create disruptive noise. Cabinet-integrated liquid options reduce footprint and keep sound levels acceptable for adjacent staff.

• Life-extension upgrades. If you’re already touching cooling due to refrigerant changes or end-of-life plant, introducing a liquid loop for your hottest racks can relieve stress on legacy CRACs/CRAHs, deferring major capex while accommodating new workloads.

• Energy-efficiency programmes. By removing fan energy and reducing recirculation, liquid systems can improve PUE and stabilise operating envelopes—particularly when paired with smarter EMS controls, dashboards and, increasingly, AI-assisted optimisation.

Integration without disruption

A common misconception is that adopting liquid cooling means ripping and replacing the entire cooling plant. In practice, it’s often additive and phased.

Start by ring-fencing target racks (for example, an AI training cluster) and design a closed liquid loop serving those enclosures via CDUs. Heat can be rejected to facility water, dry coolers or existing chiller circuits. The wider room remains on air, with setpoints and airflow tuned so the two systems coexist without cross-contamination of heat or humidity targets. This hybrid approach spreads capex and limits risk while you build internal capability and confidence.

Design fundamentals still apply: route power and network to the chosen location, validate floor loading, and address security and service clearances—especially for micro/edge deployments placed outside traditional DC rooms. Sound-proofing, access control, and CCTV are part of a professional deployment, even for a single cabinet on a shop floor.

On the operational side, refresh your EMS rules and dashboards so liquid and air data are visualised together. Most sites already collect rich telemetry (intelligent PDUs, environmental sensors), but the value comes from using it to spot drift, confirm savings, and adjust controls. If you’ve not reviewed your rulesets in years, do it before and after introducing liquid so you can lock in gains and avoid control conflicts.

Maintenance and skills

Liquid introduces different maintenance rhythms. D2C requires attention to quick-disconnect integrity, filtration, coolant quality, and leak detection. Immersion shifts the service model (lift-and-drain vs. swap-and-slide), so spares planning and technician training need revisiting. None of this is insurmountable; it’s about updating SOPs and ensuring the right contractor competencies—just as you would when adopting any new plant technology.

This is also where AI-assisted operations are beginning to help. Models trained on multi-site datasets can flag anomalies in pump curves, predict valve or seal issues, and optimise setpoints across mixed cooling estates. Crucially, you can deploy AI incrementally—recommendation mode first, then controlled automation within guardrails—to earn trust while capturing efficiency gains.

Costs, gases and the long view

The capital case for liquid cooling should be built on total lifecycle economics, not just the hardware bill. Consider avoided spend on oversized air paths, deferred room-level upgrades, lower fan energy, potential heat-re-use, and the refrigerant transition you’ll otherwise face. Between tightening F-Gas quotas and rising energy prices, availability and pricing pressures will only intensify. Planning a measured shift now can reduce risk, rather than forcing a hurried replacement later.

Equally, liquid cooling is not a silver bullet. Many workloads remain perfectly suited to well-engineered air containment, especially where density is modest and room geometry is favourable. The winning strategy is a portfolio mindset: match cooling modalities to workload profiles and site constraints, and design for tomorrow’s heat load—not just today’s nameplate.

A practical adoption roadmap

1. Identify the heat. Map current and forecast hotspots (AI, analytics, VDI bursts) and select initial candidate racks. Validate with measured data, not assumptions.

2. Pilot a pod. Deploy D2C for a small cluster with CDUs and integrated leak detection. Keep the rest of the room on air; tune setpoints and EMS rules.

3. Prove the economics. Track energy, availability and service effort for six to twelve months to build the business case for scale-out.

4. Scale with governance. Extend liquid to additional racks or edge sites, fold maintenance into BAU, and consider AI-assisted optimisation once telemetry is stable.

The bottom line is that liquid cooling is neither fad nor a replacement religion; it’s a powerful tool whose time has come—especially for densifying parts of your estate without upheaval. Used judiciously, it will help you deliver more compute in the same footprint, stabilise operations in hotter summers, and navigate the refrigerant transition on your terms, rather than the market’s.