Liquid Cooling and What It Means for Your Money

Why AI densities force liquid cooling, where the extra capital goes, and the two investment consequences that follow — higher build cost per megawatt and sharper obsolescence risk.

Liquid cooling is the engineering change that defines the AI data center, and it has two consequences an investor should care about: it costs more to build, and it makes a building more specialised. Specialisation is excellent while your tenant wants exactly what you have, and expensive when you need a different one.

You do not need to understand cold plate design. You need to know where the money goes and what it does to risk.

Why air stops working

A conventional data center rack — web servers, databases, corporate systems — draws roughly 5 to 10 kilowatts, and fans can move enough air across it to carry the heat away. A rack of AI accelerators built for training draws several times that.

Somewhere above roughly 20 to 30 kW per rack, air cooling runs out. The air cannot carry the heat fast enough at any practical flow rate, and the building has to deliver liquid to the rack instead. There are two main approaches:

Direct-to-chip. Cold plates sit on the processors, fed by a liquid loop through manifolds and a coolant distribution unit. The rest of the room may stay air-cooled. This is the mainstream retrofit and new-build approach.

Immersion. The hardware sits in a dielectric fluid. Thermally very effective, operationally more unusual, and a larger departure from how data centers are serviced.

Either way, a building type designed around fans acquires a plumbing system: pumps, manifolds, heat exchangers, filtration, leak detection, and the procedures to operate them.

Where the capital goes

Relative to an air-cooled facility of the same capacity, the extra cost sits in:

• The liquid loop itself — CDUs, manifolds, piping, pumps, valves, filtration.
• Heat rejection at higher density — the plant that gets the heat outside.
• Structural and floor work for heavier racks and the pipework beneath.
• Leak detection and containment, and the commissioning to prove it.
• Electrical distribution sized for far more power in far less floor area.

Note what gets cheaper: floor area. If one rack draws 100 kW, a 10 MW facility is only about a hundred racks. The building need not be large; the electrical service must be. This is why the unit of value in this industry is the megawatt and why an offering document that leads with acreage is describing the wrong asset. See what is an AI data center.

Consequence one: cost per megawatt, and what to check

Because the capital is concentrated in electrical and mechanical plant rather than shell, cost per megawatt is the comparison that matters — total budget divided by capacity.

When you compare that figure across deals, a number well below the others is not automatically good news. Ask what it excludes. The usual omissions are the utility network upgrade, bridge power, and any capital replacement reserve. A budget that leaves out the substation is not cheaper; it is incomplete.

Consequence two: obsolescence risk, honestly

This is the part that rarely appears in marketing.

A data center's shell and electrical service are long-lived — decades. The mechanical fit-out is not: chillers, pumps, CDUs, UPS and switchgear have finite lives and will need replacing inside the horizon of a fifteen-year projection. A model that forecasts fifteen years of income with no capital replacement reserve is understating cost, and you should ask for the reserve assumption explicitly.

Beyond wear, there is functional obsolescence. A facility is engineered for a load profile and a cooling approach. If densities keep climbing, a building plumbed for today's hardware may need re-engineering for the next generation; if the industry converges on an approach different from the one installed, the retrofit cost falls on the owner. Nobody knows which way this goes, and a sponsor who claims to is overselling.

Set against that: the power service is the durable asset. Tens of megawatts delivered to a specific site is the scarce thing, and it does not become obsolete. Cooling technology can be replaced inside a building whose interconnection remains valuable. That is the real reason a powered site holds value, and it is a better argument for the asset class than any claim about cooling specifications.

What it means for the two routes

Listed REITs absorb this at portfolio scale. Retrofit and replacement are ordinary capital expenditure across hundreds of facilities, funded from a balance sheet and disclosed quarterly. You carry the risk diluted and professionally managed.

A single private project carries it undiluted. One building, one cooling approach, one tenant whose hardware it suits. If that tenant signs a long lease, the specialisation is exactly what you wanted. If the building needs a different tenant with different hardware, the retrofit is yours. This is the concentration point made on REITs vs private deals, in its physical form.

Three questions to ask

1. What is the cost per megawatt, and what does the budget exclude — in particular the utility network upgrade?
2. What capital replacement reserve is in the model, and over what life?
3. If this tenant leaves, who is the next one, and what would the building need for them?

For the engineering itself — cooling topologies, CDU selection, commissioning — The Definitive Guide to AI Data Centers is a free 166-chapter reference and far deeper than we are. Our job stops at what the engineering does to the money.