Liquid cooling: when air runs out
AI accelerators put more heat into a rack than moving air can pull out of it. So the industry is bringing water — and other fluids — closer and closer to the silicon. This guide explains the liquid-cooling methods, from rear-door coils to full immersion, the facility loop that feeds them, the efficiency and heat-reuse upside, and what it takes to bring liquid into a hall. With an interactive method explorer.
Why Air Runs Out
Air is a poor heat-transfer medium. Cooling a rack with it works fine to around 20–30kW, but an AI rack can draw four to five times that — and you simply can’t blow enough cold air through a cabinet to keep the chips in spec. Water carries roughly 3,500 times the heat of the same volume of air, so the answer is to bring liquid to where the heat is. The only real question is how close to the chip you take it.
The physics
Heat removal scales with a fluid’s density and heat capacity. Liquid beats air by orders of magnitude, so a thin loop of water removes what a gale of air can’t.
The driver: AI density
8-GPU servers at ~10kW and GB200-class racks past 130kW put heat fluxes on the chip that air physically cannot keep up with. Liquid is now a requirement, not an upgrade.
The bonus: efficiency
Liquid lets you run warm — cooling with water at 30–45°C — which slashes or removes mechanical chilling and pushes PUE toward ~1.1.
The trade: complexity
Water (or dielectric fluid) in the white space means loops, CDUs, leak management and a new operational discipline. The density and efficiency are worth it — if it’s designed properly.
How Close to the Chip?
Liquid cooling is a spectrum — from a coil on the back of an air-cooled rack, to cold plates on the chips, to dunking the whole server in fluid. Each step takes liquid closer to the heat, handling more density at better efficiency, for more change. Explore them:
Pick a method to see how it works, what density and efficiency it reaches, and what to watch for.
The pattern is consistent: the closer the liquid gets to the silicon, the higher the density and the better the efficiency — but the bigger the change to how you build and operate. Most AI halls today land on direct-to-chip, often hybrid with air for the residual load; immersion is chosen where density and efficiency justify a wholesale change in operations.
The Facility Loop Behind the Rack
Whatever the method, the heat has to get from the chip to the outside world. It does that through (usually) two loops, kept separate by a CDU — a coolant distribution unit — so the clean, controlled fluid near the servers never mixes with the facility’s water.
The CDU
The heart of the system: a heat exchanger plus pumps, filtration and controls that isolates the clean technology loop from facility water, regulates flow and temperature, and is sized N+1 for resilience.
Two loops, separated
Facility water (which can be hard, variable, shared) never touches the servers. The secondary loop is filtered, treated and tightly controlled — protecting cold plates and quick-disconnects from fouling and corrosion.
Warm water by design
Liquid cooling runs warm (often 30–45°C supply). Warm water rejects heat to ambient most of the year — little or no mechanical chilling — which is where the efficiency comes from.
Quick-disconnects & manifolds
Dripless quick-disconnects and in-rack manifolds let servers be added, removed and serviced without draining the loop — the plumbing equivalent of hot-swap.
Efficiency & Heat Reuse
Density is why you need liquid; efficiency is why you might want it even before you’re forced. Because liquid removes heat at high temperature, it changes the energy — and water — equation of the whole hall, and turns waste heat into something useful.
Lower PUE
Warm-water cooling cuts or removes compressor-based chilling and the fans that air cooling lives on, pushing PUE from ~1.5 toward ~1.1 — a direct cut in energy overhead.
Free cooling, more of the year
The warmer the supply water you can use, the more hours a dry or adiabatic cooler handles the load with no chiller running at all — in many climates, nearly year-round.
Water use (WUE)
Liquid cooling can reduce evaporative water use versus some air-cooled chiller designs — but it depends on the heat-rejection choice; WUE has to be designed for, not assumed.
Heat reuse
High-grade captured heat can warm offices, district heating networks or industrial processes — turning a cost into a contribution, and a strong sustainability story.
The sustainability angle is real and increasingly demanded: AI’s energy appetite is under scrutiny, and a hall that runs at low PUE on warm water — ideally reusing its heat — is far easier to justify than one chilling air at 1.5+.
Retrofitting & Designing for Liquid
Few operators flip a whole hall to liquid overnight. The usual path is hybrid — liquid for the dense AI rows, air for the rest — and increasingly, designing every new hall liquid-ready even if it starts air-cooled. Either way, these are the things that decide whether it goes smoothly.
Leak management sits at the very top — water near live electronics is the fear that stops projects, and the answer is engineering: dripless disconnects, leak detection, drip trays, negative-pressure loops and tested procedures. Done properly, liquid cooling is no riskier than the power chain it sits beside. The other half is designing it with the white space and the loop together, not bolting it on — which is exactly where a fit-out partner earns its keep.
How Optronix Helps
Liquid cooling lives where we work — the white space, the containment, the pathways and the loop. We design and deliver liquid-ready and liquid-cooled halls, integrate CDUs and pipework with the rack layout and cabling, and retrofit density into existing air-cooled estates — with our own engineers, across the UK, EMEA and worldwide.
Designed with the white space, not bolted on
We model the loop, the CDUs, the pipework and the rack layout together — so floor loading, pathways, leak management and serviceability are designed in from the start, not retrofitted into a fight.
Air, hybrid or full liquid
Rear-door retrofits, direct-to-chip rows alongside air, or liquid-ready new build — matched to your density today and your roadmap, with the residual air load planned in.
Assess & design
Density targets, method selection, loop and CDU sizing, floor loading and pathways — modelled and costed alongside the white space.
Install & integrate
CDUs, manifolds, pipework and containment installed and integrated with power and the cabling plant — cleanly, by our own teams.
Commission & prove
Leak-tested, flow- and temperature-verified, and commissioned through to integrated systems testing before any live load.
Operate & support
Ongoing operational support, smart hands and maintenance for the loop and the hall around it.
Going dense, or going liquid?
Tell us your target densities and whether it’s a retrofit or new build, and we’ll come back with a costed liquid-cooling design — method, loop, layout and the lot — built to perform and to scale.
Sources & notes
- ASHRAE TC 9.9 — liquid-cooling guidelines and facility-water temperature classes (W17–W45)
- Accelerator / AI-system vendor specifications for rack densities and direct-to-chip heat capture
- Common industry practice for CDUs, primary/secondary loops, immersion (single- and two-phase) and heat reuse; figures are indicative and vary by deployment
General guidance, not a design specification. Density ceilings, PUE, heat-capture and temperature figures are indicative and depend on the specific kit, coolant, temperatures and site; two-phase technologies and dielectric fluids are evolving and face changing environmental/regulatory scrutiny. Confirm against vendor data and an engineered design.