Above roughly 30 kW per rack, air cooling stops working. Above 50 kW, it stops being defensible even with hot-aisle containment, raised pressure, and aggressive CRAH design. AI training racks are now routinely above 60 kW; new accelerator generations push toward 120 kW+. Liquid cooling has therefore become a procurement question, not an R&D one.
Two technologies are production-ready in Europe today. The right choice for a given deployment depends on three things: density, retrofit constraints, and what the facility will commit to in writing.
Direct-to-chip (D2C) cooling
A cold plate sits on each high-power component — typically GPUs and CPUs — and a coolant loop carries heat to a coolant distribution unit (CDU) at the rack or row level. The CDU exchanges heat with the facility’s chilled water loop. Servers stay air-cooled at the chassis level for the components that aren’t on cold plates (memory, NVMe, NICs).
Where D2C wins:
- Retrofit-friendly. Most modern Tier III facilities can accept D2C with manageable changes — primarily the addition of a chilled-water loop to the rack, and rack-level CDU placement. We see this happening in selected NL, DE, and BE facilities now.
- Density ceiling. Production deployments today land between 60 and 100 kW per rack with rear-door heat exchangers added for residual air load. Higher possible with redesign.
- Vendor support. All major server vendors ship D2C-ready SKUs and the integration patterns are well-documented.
Where D2C is harder:
- Plumbing in the rack. Quick-disconnect couplings and proper service procedures are required. Operations teams need training. Leak detection and containment are real considerations, not theoretical.
- Mixed cooling complexity. You’re operating two cooling loops (air for chassis, liquid for chips). PUE math is more complicated than air-only or immersion.
- Component coverage. D2C cools the chips. Memory, NICs, and NVMe still need air. At very high density, the residual air load is non-trivial.
Immersion cooling
An entire server (or rack) is submerged in a non-conductive dielectric fluid. Single-phase systems keep the fluid liquid throughout and circulate it to a heat exchanger. Two-phase systems let the fluid evaporate at the chip surface and condense back at the lid; the phase change carries large amounts of heat per kilogram of fluid.
Where immersion wins:
- Density ceiling. Comfortably above 100 kW per rack. Two-phase systems can handle the densest accelerator configurations on the market without architectural redesign.
- Whole-system cooling. Every component is cooled equally, including NICs and memory. This becomes important above ~80 kW per rack where D2C’s residual air load becomes hard to manage.
- Mechanical simplicity. No fans on the servers, no chassis-level air handling, no hot-aisle containment. The rack is the cooling unit.
- Hardware lifespan. Lower thermal cycling, no dust ingress, generally extends component life.
Where immersion is harder:
- Greenfield-shaped. Retrofitting an existing data hall to take immersion tanks is non-trivial. Floor loading, fluid handling, fire suppression interactions, and serviceability all need rethinking. Most production immersion in Europe is in purpose-built halls or carefully selected zones of larger facilities.
- Servicing and warranty. Pulling a server out of fluid for service is a real procedure. Some vendors have warranty terms that change for immersed equipment.
- Fluid management. Fluid lifecycle, top-up, filtration, and disposal are part of the operational cost model. Two-phase fluids have additional environmental considerations under recent EU regulation.
How we filter on cooling in the shortlist
When the buyer’s brief flags 30+ kW per rack, cooling becomes a hard filter:
- We ask the facility for the maximum contractually committed power density per rack, not the maximum achievable in optimal conditions.
- We ask which cooling technologies the facility supports today, in production, with named existing customers (NDA-bound is fine — we just want the operator to confirm production deployments exist).
- We ask for the operational PUE under those high-density configurations specifically. Hyperscale-adjacent facilities sometimes have a great average PUE and a much worse PUE in the dense zones; we want the right number.
- We ask about leak detection, containment, fire suppression interactions, and operational procedures. The technology is mature; the operations sometimes are not.
Facilities that cannot answer these in writing are flagged on the shortlist. They are not automatically excluded — for the right buyer profile and price, the trade-off can be acceptable — but the buyer sees the gap explicitly.
The technology choice is downstream of the facility commitment. Pick the facility that can deliver the density profile your workload needs, and let that constrain the cooling technology rather than the other way around. We have seen too many AI projects pick immersion on principle and then spend nine months looking for a hall that will host it.