AI data centre cooling
The reference for practitioners on AI-density cooling: direct-to-chip liquid cooling (DLC), single-phase and two-phase immersion, cooling distribution unit (CDU) architecture, fluid supply chain, vendor landscape, and hyperscaler procurement patterns.
01Direct answer
02Key concepts
03Structural decomposition
The AI cooling architecture stack
- Facility plant. Chillers, dry coolers, cooling towers; central plant sized to AI density
- Facility loop. Distribution piping (chilled water or elevated hot water) to racks
- CDU. Cooling distribution unit; heat exchange + secondary loop pumping
- Rack loop. Manifolds + hoses + connectors distributing to compute blades
- Cold plates or immersion tank. Point-of-thermal-load: direct contact with GPU/CPU packages
- Return + heat rejection. Return path to facility loop and eventual ambient rejection
04Deep-dive research
- The Thermal StackHow heat became the binding constraint on AI compute.
- The Thermal Stack technical companionFluid chemistry, thermal transport, and leak sensitivity.
- CoolIT direct-to-chipDLC vendor dynamics and the CoolIT (Ecolab) integration.
- Submer immersion coolingImmersion cooling vendor landscape and fluid supply.
- Cooling + 800V DC co-emergenceHow liquid cooling and 800V DC power architectures reinforce each other.
05Standards + regulation
- ASHRAE TC 9.9 Mission Critical FacilitiesData centre thermal operating envelopes; W3-W5 water temperature classes for liquid cooling.
- OCP Advanced Cooling SolutionsWorking group on rack-scale liquid + immersion cooling standards.
- NFPA 70EElectrical safety in workplace; applicable when liquid cooling intersects live electrical.
06Commercial + investment implications
Commercial + investment angles
Cooling is a high-M&A-activity segment. CoolIT (Ecolab) set the vertical-integration precedent; expect more consolidation. Fluid supply concentration (two-phase immersion) is a specific supply-chain thesis. See Case Study 04: Eaton/Boyd one year on for a recent strategic acquirer entry into cooling.
07Frequently asked
When does an AI facility need liquid cooling?
Above roughly 30-40 kW/rack sustained. Air cooling economics break down as density rises, and the incremental facility infrastructure to support high-density air cooling (larger CRACs, cold aisle containment, more airflow) eventually costs more than the incremental liquid-cooling infrastructure. Hyperscaler AI-density (100+ kW/rack) is unambiguously liquid.
DLC vs immersion: which wins?
Both persist. DLC is easier to retrofit into existing rack forms and is dominant for GPU-centric AI training clusters. Immersion offers higher density per square metre and simpler thermal management for edge cases (extreme density, high-ambient siting). Neither is universally superior.
Why is two-phase fluorinated fluid supply a concern?
Manufacturing capacity for the specific fluorinated fluids used in two-phase immersion (3M Novec, Solvay Galden, etc.) is concentrated among a small number of chemical suppliers, some of whom face regulatory pressure on PFAS/fluorochemistry. Growth in immersion demand could outpace fluid supply expansion.
How do cooling and 800V DC power architectures interact?
They co-emerge because both are driven by AI-density economics. High-density racks need both. Sidecar power racks (Mt. Diablo) can share thermal envelopes with liquid-cooled compute racks. See DC-DC VIII: Cooling co-emergence.
What is the biggest vendor concentration risk in cooling?
CDU + high-current disconnect connectors, specifically for hyperscaler-scale deployments. CoolIT (Ecolab) + Motivair + Boyd + Asetek dominate DLC; concentration in specific hyperscaler design wins matters more than aggregate share.
08Further reading
| Rack power | Dominant cooling architecture | Vendor concentration | Typical deployment |
|---|---|---|---|
| Under 30 kW/rack | Air cooling with hot/cold aisle containment | Diverse (Vertiv, Schneider, Stulz, Rittal) | Traditional enterprise + legacy colocation |
| 30-60 kW/rack | Rear-door heat exchanger (liquid-assisted air) | Motivair (Schneider), CoolIT (Ecolab), Vertiv | Colocation + inference workloads |
| 60-150 kW/rack | Direct-to-chip liquid cooling (single-phase) | CoolIT, Motivair, Boyd (Eaton), Vertiv | H100/H200 hyperscale + training clusters |
| 150-500 kW/rack | DLC + facility loop chilled water | Same, with CDU manufacturers layered | GB200 NVL72 + Rubin generation training |
| 500+ kW/rack (emerging) | Two-phase immersion or two-phase cold plates | LiquidStack (Trane, closed 3 March 2026), ZutaCore, Iceotope | First-of-a-kind Rubin Ultra + custom silicon |
Data Centre Cooling · a double-click on Direct-to-Chip
The nine-layer DTC stack, end-to-end: physical schematic from GPU die and TIM through cold plate, QDs, manifold and CDU out to the facility loop. CDU as testable control-point proposition. Single-phase vs two-phase reframed as an economic threshold, not a physics deadline. Chilldyne's negative-pressure architecture as the third distinct route. Value migration map, capex by rack density, consolidation race, geopolitical supply-chain exposure and investor decision map by profile. Anchor deep-dive for this hub.
Data Centre Leak Detection · the six-stage qualified path
Sub-system companion to the DTC anchor. The sensing element (rope, spot, dielectric, optical) is being standardised at the OCP layer while the qualified control path (CLASSIFY → AUTHORISE → ISOLATE) stays proprietary. Reference topology on a 100 MW / 85 MW IT campus · ~$14.7/kW IT · ~$1.25M campus total · 24 named vendors from Chemelex TraceTek and TTK through PermAlert, Vutlan, Tatsuta and CoolIT-embedded telemetry. Companion interactive calculator at /leak-detection-calculator/.
DTC cooling · capex, opex + TCO calculator
Author-built scenario model that turns the DTC essay's Fig C15 anchor stack ($14K / $44K / $87K per rack at 40 / 100 / 250 kW) into a live workbench. Six architectural inputs drive CDU sizing, per-rack + campus capex, year-1 opex and 7-year TCO. Includes a capex waterfall, TCO-over-time chart, payback engine on the 45 °C free-cooling and two-phase switches, a 3×3 sensitivity matrix on electricity price and utilisation, plus URL-hash scenario sharing and one-click Markdown export.