Almost all the electricity consumed by computing equipment ultimately becomes heat. Pack more accelerators into a rack and that heat must be removed from a smaller space without interrupting the workload. Cooling is therefore a limit on what can be deployed, how hard it can run and how reliably it operates. A thermal system that fails to keep up can turn valuable computing capacity into throttled performance or an outage.
The investment question
The cooling investment question is whether higher AI density creates a durable expansion in customer spending and which suppliers capture it. Cold plates, pumps, coolant distribution units, heat exchangers, chillers, controls and services perform different roles. A growing liquid-cooling market does not imply that every component enjoys equal pricing power.
Our view is that integration and operating reliability are central to value. Customers need a complete heat-removal path that works across equipment generations and local conditions. Suppliers can differentiate through qualified designs, commissioning and maintenance, while commoditisation remains a risk for components that become standard and widely available.
How heat leaves the data centre
Air cooling transfers heat from components into air and then through the facility’s cooling system. Direct-to-chip liquid cooling uses cold plates to carry heat away from selected components. Immersion cooling places suitable equipment in a dielectric fluid. Rear-door heat exchangers capture heat from server exhaust. These approaches can coexist within one facility.
A coolant distribution unit, or CDU, helps manage the liquid circuit and heat exchange between equipment and the surrounding cooling system. Vertiv’s CDU explanation describes closed-loop circulation and different heat-exchange arrangements. The CDU is one part of the system: heat still needs to be rejected to the outside environment or used elsewhere.
That distinction prevents a common misconception. Liquid cooling does not necessarily mean high water consumption, and a closed loop at the rack does not automatically mean zero water use across the facility. The external heat-rejection method, climate and operating strategy determine much of the outcome. Electricity and water use should be assessed over clearly defined boundaries.
Market structure and competitive advantage
| Layer | Function | Competitive test |
|---|---|---|
| Cold plates and interfaces | Capture heat at components | Thermal performance and mechanical compatibility |
| Fluid distribution | Move coolant safely and consistently | Reliability, cleanliness and serviceability |
| Facility heat rejection | Transfer heat outside the computing environment | Energy use, climate fit and capacity |
| Controls and services | Keep the system operating correctly | Monitoring, commissioning and response capability |
Vertiv and Schneider Electric provide broad data-centre infrastructure capabilities, while other thermal-management and industrial suppliers compete in specific layers. Server manufacturers and hyperscalers also influence the design and qualification of cold plates, manifolds and associated components. The supplier relationship can therefore be with the server platform, facility operator or both.
Qualification and operating experience create barriers. A cooling component must meet thermal needs without introducing leaks, contamination, corrosion or maintenance problems. ASHRAE’s cold-plate discussion explains the physical heat-transfer role. Commercial differentiation depends on delivering that performance consistently within the customer’s complete operating environment.
Economics: heat removal and productive capacity
The customer’s cooling cost includes initial equipment, installation, pumps, fans, refrigeration where required, water where consumed, maintenance and the space occupied by infrastructure. It also includes the cost of downtime or reduced processor performance. A system with lower purchase cost can be economically inferior if it limits density or requires more energy and service over its life.
PUE can help assess facility overhead, but it is not a complete cooling investment metric. A more efficient cooling design may enable a denser, higher-power computing fleet, increasing total electricity use while improving useful output. Compare energy per relevant workload and total lifetime cost alongside facility-efficiency measures.
Illustratively, reducing facility overhead from 30% to 20% of a constant 10 MW IT load lowers total demand from 13 MW to 12 MW. At continuous operation, the saving is 8,760 MWh a year. This is arithmetic under fixed assumptions, not a forecast for a cooling product. Actual savings depend on climate, utilisation, equipment and the baseline being replaced.
Supplier economics depend on product mix, project execution and service attachment. New technologies can support higher content per rack, while competition and standardisation pressure component pricing. Large projects may also require working capital ahead of delivery. Track cash conversion and margins as closely as orders, especially during rapid capacity expansion.
AI and hyperscalers: density drives co-design
AI changes the thermal problem through higher heat density and concentrated loads. Direct liquid cooling can remove heat close to processors, potentially reducing the burden on air movement. Other server components may still require air cooling, so many deployments remain hybrid. The right architecture follows the full server and facility design rather than a universal density threshold.
Hyperscalers can coordinate chip, server and facility decisions, influencing coolant temperatures, interfaces and maintenance practices. They may prefer repeatable designs that can be deployed across several locations. That creates large supplier opportunities but also gives customers leverage to standardise components and qualify alternatives.
Microsoft’s June 2026 discussion of its cooling designs describes closed-loop, direct-to-chip approaches without evaporative cooling in the relevant designs. This is evidence of an operator’s design strategy, not a statement that every Microsoft facility or every liquid-cooled data centre has the same water profile.
Current market debates — September 2026
The near-term debate is how quickly liquid cooling moves from selected high-density deployments into a larger share of new capacity and retrofits. Vertiv’s second-quarter 2026 results show strong business performance and raised guidance. The company’s broader infrastructure portfolio means its total growth should not be treated as a pure measure of liquid-cooling adoption.
The constructive case is that denser computing makes additional thermal investment unavoidable and creates service opportunities. The countercase is that large buyers standardise designs, component supply expands and competitive pricing captures part of the benefit for customers. Evidence of repeat qualified programmes and margin retention is stronger than the number of newly announced cooling products.
Commissioning is an immediate operational issue. ASHRAE’s performance-validation guidance highlights challenges associated with rapidly deployed liquid-cooled systems, including cleanliness and preparation. A delivered cooling unit is not the same as a tested system operating reliably under full load. Installation and service capability can determine whether demand becomes profitable revenue.
Structural debates: retrofit, standards and water
Retrofitting existing sites may require changes to pipes, electrical systems, floor layouts and maintenance procedures. A solution that works in a purpose-built facility may be uneconomic in an older one. Hybrid approaches can extend existing assets, but the operator must compare retrofit cost with the value of the additional computing capacity enabled.
Standardisation could expand adoption by reducing integration friction. It can also make components easier to substitute. Suppliers are most defensible where they contribute system knowledge, validated performance or ongoing service rather than relying solely on a proprietary fitting or early market availability.
Water and heat reuse create another set of trade-offs. Dry heat rejection can reduce direct water consumption but may change electricity requirements under certain conditions. Recovering heat can be useful where nearby demand, temperature requirements and infrastructure align. Neither approach should be credited with universal savings without a site-specific assessment of the complete system.
What to watch
Track qualified liquid-cooling deployments, content per rack, commissioning performance, service revenue and warranty experience. Compare order growth with manufacturing and installation capacity. Evaluate water and energy claims against the stated design boundary and local operating conditions.
The strongest cooling franchise enables customers to run valuable computing equipment safely, densely and economically over time. Its advantage should persist after the first wave of urgent installations gives way to more standard purchasing decisions.
Explore this sector
AI infrastructure & data centres — sector overview
Related sectors: Semiconductors & chipmaking · Data platforms & analytics
