The expansion of artificial intelligence and cloud computing has transformed data centres into unprecedented powerhouses, and with power inevitably comes heat. Today, managing that heat is no longer an engineering detail, but the variable that determines whether an infrastructure can function, how much it consumes, and how much it can grow.
to understand why the data centre refrigeration it has become one of the most relevant engineering challenges of the decade means looking at the physical limits of air and the paths the sector is taking today to keep temperatures under control.
Heat is the true byproduct of computation
Every watt powering a server eventually turns into heat. It’s a law of physics even before it’s a design problem: the electrical energy that enters a data centre doesn’t vanish, it is dissipated as heat What needs to be removed continuously and reliably. As long as consumption was limited, this task was manageable with traditional air conditioning systems, but the situation has changed radically with the growth of high-intensity computing.
The figures help to illustrate the scale of the phenomenon. Globally, data centres consumed around 415 TWh of electricity in 2024 – just over 1.5% of the world’s electricity – and projections indicate this figure will double to 945 TWh by 2030. Italy is also experiencing this acceleration: in 2025, installed IT capacity reached 609 MW, with the Milan area alone accounting for around 68% of the national total and set to exceed the 1 GW threshold by 2028. More computing means more energy, and more energy means more heat to dissipate.
Power density has exploded
The engine of this change has a precise name: power density, which is how many kilowatts a single cabinet, the so-called Rack. For years, a rack in a traditional data centre has operated between 5 and 15 kW, an amount of heat easily managed by pushing cool air through the equipment.
AI applications have overturned these numbers. A single CPUCentral Processing Unit, the system's main processortraditional ones absorb around 150 or 200 watts, while GPUsGraphics Processing Unit, a processor specialised in parallel computationused for AI have gone from around 400 watts a few years ago to 700 watts and more in recent generations, with the most advanced chips exceeding 1,000 or 1,200 watts each. Concentrating many of these processors in the same rack easily leads to 100 kW racks, and the latest generation configurations reach 132 kW with prospects pointing towards even higher densities within a few years.
To set the proportions:
- Traditional data centre: from 5 to 15 kW per rack
- Today's high-performance configurationsFrom 60 to over 100 kW per rack
- Cutting-edge AI systems: up to 132 kW per rack, with subsequent generations aiming for even higher values
Why air alone is no longer enough
Air cooling has worked for decades because the thermal load was compatible with the properties of the medium. However, air is a poor conductor of heat, and to remove high powers, it is necessary to move enormous volumes at ever-increasing speeds. Beyond a certain threshold, roughly between 20 and 35 kW per rack, simply increasing airflow stops being feasible: the fans consume too much energy, difficult-to-control hot spots form, and overall efficiency deteriorates.
The energy cost of this approach is far from marginal. In traditional data centres, air cooling alone can consume up to 40% of total electricity, a weight that is directly reflected in the most commonly used indicator for measuring efficiency, the PUE (Power Usage EffectivenessA traditional system generally sits between 1.4 and 1.6, meaning that for every watt used for calculation, 0.4 to 0.6 are consumed just for the periphery, while the most efficient operators manage to get down to 1.1. With very high-density racks, no amount of air optimisation is enough to reach those levels.
Liquid cooling, the industry's answer
La solution which the industry converges on today is going from gas to liquid. The principle is intuitive: liquids transport heat much more efficiently than gases, and bringing the refrigerant close – or even into contact – with hot components allows for the removal of power levels unimaginable for air. However, there isn't a single technology, but rather a range of approaches that differ in efficiency, complexity, and cost.
Direct-to-chip cooling
In this approach, a cold plate, Cold plate, it is mounted directly on top of the processor. A coolant, often water or a water and glycol mixture, circulates inside it, which absorbs heat right at the point where it's generated. It is the most widespread technology today because it offers an excellent compromise: it removes heat at the source by reducing thermal resistance and, at the same time, integrates with relative simplicity into existing structures, managing racks of around 80 or 120 kW.
Immersion cooling
Here servers, or entire components, are immersed in a dielectric fluid, which is non-conductive, which absorbs heat over the entire surface of the electronics. In Single-phase systems the fluid remains liquid and circulates towards a heat exchanger, while in the two-phase systems evaporates and then condenses, obtaining even more efficient heat transfer. Immersion ensures Very uniform cooling and extremely high densities, up to 100 or 250 kW and beyond per rack, but requires dedicated facilities and more complex maintenance management.
Rear door exchangersRDHx)
A middle way is represented by the Rear rack heat exchangers, the so-called rear door heat exchanger. They cool the outgoing air by passing it through a battery leak, before it re-enters the environment. They are often adopted as an adaptation solution, useful for increasing the density of air-born structures without needing to redesign them from scratch.
In practice, these approaches coexist. Many AI configurations today adopt a hybrid model, in which most of the heat, around the 70 or 85%, comes liquid-cooled, while a residual portion is left to the air for less critical components. Liquid cooling, in short, does not eliminate air cooling, but redefines its role within the system.
A challenge that goes beyond temperature
Refrigeration is the process of lowering the temperature of a substance or space below its surroundings by removing heat. This is commonly achieved using a refrigeration cycle, which typically involves a refrigerant fluid that undergoes phase changes to absorb heat from the desired space and release it elsewhere. The engineering challenge of the decade it’s not an exaggeration, because the problem isn’t just about chip temperature. Other factors come into play at least three more dimensions, all strictly interconnected.
- The first is energyReducing cooling consumption means improving PUE and containing operational costs which, on these scales, become enormous.
- The second is water, because many traditional plants rely on evaporation to dispose of heat, with significant water consumption that represents an increasingly pressing environmental and social issue.
- The third is the normative dimension, which in recent years has entered the debate forcefully.
In Europe, the Energy Efficiency Directive (EEDimposes on data centres with an IT power rating of 500 kW or more communicate indicators annually like PUE, WUE (Water Usage Effectiveness), the energy reuse factor and the share of energy from renewable sources. For larger plants, above one megawatt, there is also a growing Waste heat recovery drive, for example by feeding it into district heating networks. Heat, from being a problem to be disposed of, thus begins to be seen also as a resource to be valued.
The real question for those designing a data centre today is no longer just how much computing power it can house, but how much heat it is able to remove efficiently, reliably, and sustainably.
From thermal issues to precision components
This all has a concrete and often underestimated consequence: the liquid cooling circuit has become a mechanical system as critical as the electronic part. Pumps, manifolds, fittings, quick connectors, valves, and cold plates must ensure a perfect seal, corrosion resistance e Consistent reliability over time, because a cooling leak or failure puts extremely valuable hardware at risk and can interrupt service in a very short time.
It is precisely the ground in which the precision turning finds its role. The production of small metal components manufactured with tight tolerances, in materials chosen for their resilience to the circuit environment, is an integral part of the reliability of these systems.
Would you like to assess the feasibility of a precision component for high-reliability applications?
Get in touch!