In data centers, artificial intelligence is not just changing servers: it is overhauling the entire plumbing system. To train and run models, operators are packing increasingly powerful accelerators into racks that air struggles to cool. Liquid then becomes an infrastructure choice, complete with pipes, risks and promises. Looking ahead to September 2026, the strategic question is no longer just how to deliver more computing power, but where to send the heat. This outlook builds on industrial changes already underway; it does not imply they will be universally adopted by that date.
AI upends established cooling practices
A conventional server can be cooled by fans circulating air between its components. At building scale, hot and cold aisles, sometimes enclosed, organize this airflow. But accelerator clusters change the equation: they bring together power-hungry chips, fast memory and dense interconnections in a confined space. Virtually all the electricity consumed ultimately becomes heat that must be removed.
The problem is as much about concentration as total volume. A server room may have enough electrical power available without being able to extract the heat from a few particularly densely loaded racks. Speeding up fans does not solve everything: it consumes more energy and runs into physical limits. The Nvidia Blackwell architectures unveiled in 2024, particularly the liquid-cooled GB200 NVL72 systems, brought this shift in scale into sharp focus. Cooling is becoming a defining feature of computing system design.
Circulating liquid in the right place
In contact with chips, without immersing them
The approach often favored for accelerators involves attaching cold plates to the hottest components. A fluid circulates through them, carrying energy to a heat exchanger. This is known as direct-to-chip cooling. A distribution unit, generally called a CDU, handles tasks including heat exchange between circuits and control of fluid circulation conditions. The building’s liquid therefore does not necessarily flow through the servers.
This approach does not always eliminate air: power supplies, storage and other components may still depend on it. Another option is a rear door fitted with a heat exchanger that captures heat as it leaves the rack. This can make some retrofits easier, but it does not address the same requirements as a plate placed directly on an accelerator. The starting point must be the hardware’s requirements, not an abstract preference for a particular technology.
Immersion: a different setup
Immersion cooling submerges equipment in a nonconductive dielectric fluid. It can provide substantial heat removal capacity, but fundamentally changes maintenance procedures, component choices and logistics. Material compatibility, manufacturer warranties, aging and the environmental status of fluids become decisive factors. There is therefore no single, uniform transition to “liquid,” but several paths, with different supply chains and constraints.
The work extends well beyond the server room
In a new building, incorporating piping, heat exchangers and heat rejection equipment from the design stage is a relatively straightforward proposition. At an existing site, the exercise is more like open-heart surgery. Operators must check floor load capacity, pipe routing, electrical supply, outdoor equipment and the scope for work without interrupting services. A compatible rack does not automatically make the entire room compatible.
Operations change too. Fluid quality, corrosion, fittings, leak detection and flow control become daily concerns. Redundancy must cover pumps and heat exchangers, not just electrical circuits. High-performance cooling that is difficult to maintain can become an availability risk. Operators therefore have good reason to test procedures with their teams and suppliers before scaling up deployments.
Less electricity does not necessarily mean less water
Liquid transports heat more efficiently than air. By reducing some ventilation needs and, depending on the design, the need for mechanical cooling, it can cut the energy devoted to cooling. But the overall outcome depends on operating temperature, climate, computing load and auxiliary equipment. Pumps and outdoor fans also consume energy. Above all, greater efficiency does not prevent a site’s total consumption from rising if computing capacity surges.
The most common confusion concerns water. A closed liquid loop near servers does not imply substantial ongoing water consumption at that location. The crucial question often lies further downstream: how is heat discharged into the environment? An evaporative cooling tower consumes water. A dry cooler can sharply limit this on-site use, but entails other trade-offs, particularly during hot weather. Hybrid systems adapt their operation to conditions.
It is also necessary to look beyond the building: electricity generation can require water, depending on the technologies and regions involved. Decisions should therefore draw on several indicators:
- the site’s energy efficiency, particularly its PUE, without confusing it with computing efficiency;
- water consumption and pressure on local water resources;
- emissions associated with electricity, equipment and construction work.
A strong annual metric can conceal a summer problem. In a water-stressed basin, a few weeks of high demand matter as much as the average reported in an environmental report.
Recovered heat needs a customer
Liquid cooling offers another benefit: capturing heat in a stream that is easier to put to use. When equipment supports higher circulating fluid temperatures, the opportunities for heat reuse improve. District heating networks, neighboring buildings and certain industrial applications can benefit. Nordic projects, such as the one announced in 2022 by Microsoft and Fortum in the Helsinki region, illustrate this approach to bringing computing and district heating together.
But available heat is not automatically useful. It requires a nearby customer, sufficiently steady demand, pipelines and sometimes a heat pump to raise the temperature. The data center produces heat year-round, while heating demand peaks in winter. Who pays for the connection? Who guarantees supply if the computing load varies? Heat recovery is primarily a local infrastructure and contractual undertaking, not a simple hookup.
A strategic advantage, with conditions
For operators, planning ahead for liquid cooling can make it easier to accommodate future generations of accelerators and avoid repeated retrofits. For customers, it is becoming a site selection criterion alongside available power and connectivity. But investing too early in a rigid architecture also creates exposure to future incompatibilities. Modularity, documented interfaces and a clear division of responsibilities between hosting provider, manufacturer and tenant are acquiring very tangible commercial value.
What next? Looking ahead to September 2026, the most plausible scenario is lasting coexistence between air and liquid, rather than an overnight replacement. Sites capable of balancing chip requirements, power availability, water resources and outlets for recovered heat could gain an advantage. The right question will no longer be “do you have liquid cooling?” but “what complete system have you built around it, and what results can you demonstrate?”
