Immersion Cooling: European Datacenters' Answer to AI GPU Density

Air cooling is hitting its physical limits: beyond 30 to 40 kW per rack, ventilation becomes a losing battle. With AI training GPU racks now exceeding 100 kW, liquid is no longer optional — and immersion, long marginal, is becoming a serious design topic for European datacenters.

A technical reminder. Air comfortably removes 10 to 20 kW per rack. Direct-to-chip liquid goes beyond 100 kW. Recent accelerator racks, from NVIDIA and its competitors alike, are designed for liquid from the outset. Immersion goes further: the entire server is submerged in a dielectric fluid, single-phase or two-phase.

The technology is not exotic: HPC has used it for years, and European suppliers — Submer in Spain, Iceotope in the UK, Asperitas in the Netherlands — have turned it into industrial products. The Open Compute Project publishes open specifications. What has changed is demand: GPU density makes the debate concrete.

What Immersion Changes in a White Room

No more server fans, almost no room air conditioning: according to operating feedback, the partial PUE of immersed installations can approach 1.1, versus 1.3 to 1.5 for conventional air cooling. Heat is recovered at high temperature, and therefore reusable — a direct advantage under the EED directive.

The constraints are real: high floor loading, fluid logistics, operator training, OEM warranties to be validated server by server. Immersion is not an off-the-shelf product; it is an operating choice that commits a site for twenty years.

What This Changes for Investors

The economic trade-off is simple: immersion costs more in capex, less in energy opex. It pays off at high density — beyond 50 to 80 kW per rack — and loses its meaning on a conventional 10 kW colocation site. The question is therefore not 'for or against immersion', but 'what density does this site target, for which customers?'.

In practice, few existing sites will be retrofitted to full immersion: the structure, the slab and the teams were not designed for it. The dominant strategy remains direct-to-chip liquid, with an 'immersion-ready' design that preserves the option for future phases.

How We Apply This at FINXIA

In the TITAN DC AI strategy, we assess each asset on its cooling trajectory: current density, target density of customers, direct liquid compatibility, structural feasibility of an evolution toward immersion. A site frozen on air closes itself to AI customers; a liquid-designed site captures them.

On brown transformations, the structural question is decisive: slab load capacity, ceiling height, hydraulic network to be built. This is a technical due diligence point as much as a valuation argument — an asset's thermal flexibility is worth value.

Key Takeaways

GPU density mandates liquid; immersion is the most accomplished variant for very high densities, with a PUE and heat recovery that air cannot match. But it is a long-term operating choice, not an interchangeable piece of equipment.

Worth watching: the generalization of OEM warranties on immersed servers, and the effect of EED reporting — which publicly rewards the most efficient sites.