The AI era has pushed data centre cooling into the spotlight, but the real challenge is no longer removing heat from today’s hardware. It is designing infrastructure that can cope with processors, power architectures and thermal loads that have not yet been invented, while ensuring facilities built today remain viable for decades to come.
The discussion around AI infrastructure is often framed as though cooling has suddenly become the industry’s biggest problem. Conference agendas are filled with debates about liquid cooling, direct-to-chip technologies and immersion systems, while every new generation of processors seems to arrive with another dramatic increase in power consumption. It creates the impression that the industry is scrambling to solve an entirely new challenge.
The reality is rather different. Cooling has not suddenly become difficult because of artificial intelligence. Heat densities have been rising steadily for years and the engineering principles behind removing that heat have changed very little. What AI has done is accelerate a trend until conventional air cooling has finally reached its practical limits inside the server itself. The conversation has become louder because the industry has crossed a threshold that many engineers knew was coming.
That distinction is important because it changes where the real challenge lies. Selecting liquid cooling instead of air is increasingly becoming an engineering decision rather than a strategic one. The far more difficult question is how operators design facilities expected to remain operational for thirty years when the hardware they will eventually support may bear little resemblance to anything available today.
Paul Aguilar, Cooling Director at Legrand Data Center Solutions, believes the industry has reached an important point in its evolution, although not because cooling itself has suddenly become more complicated. “The conversation has not really changed because densities have increased continuously over the last fifteen or twenty years,” he says. “The reason there is now a perceived change is because we have reached the limits of cooling with air at the server level. The only way the industry can meet the demand created by higher densities is to transition towards liquid cooling architectures.”
His observation strips away much of the hype that now surrounds AI infrastructure. Liquid cooling is not replacing air because it is fashionable or because vendors have discovered a new technology. It is replacing air because physics has gradually closed off every other option.
Cooling is governed by physics not marketing
Few sectors experience technological change at the pace now seen across AI infrastructure. Every year processors become faster, accelerators become more powerful and rack densities climb higher. Cooling, however, does not enjoy the same freedom to reinvent itself because it remains governed by physical laws that have not changed.
That is why Aguilar is sceptical whenever cooling is presented as an area waiting for some revolutionary breakthrough. “The beauty of cooling is that it is governed by the laws of physics,” he adds. “Whether you call it plumbing, mechanical services or cooling, the fundamental principles are very simple. If you immerse a server in liquid or pass liquid through it, you are still transferring heat. You cannot destroy energy. All you can do is move it somewhere else.”
It is a useful reminder that, despite the pace of innovation elsewhere, cooling engineers remain constrained by constants. A given volume of water removes a known amount of heat. Heat exchangers transfer thermal energy according to established principles. Pumps, valves and pipework continue to perform the same fundamental task regardless of how sophisticated the processors become.
Innovation therefore comes less from rewriting the laws of thermodynamics than from applying them more intelligently. Rear-door heat exchangers, direct-to-chip cooling and immersion systems all represent different methods of solving the same engineering problem rather than entirely different problems.
That perspective also helps explain why experienced operators are often less excited by marketing claims than newcomers to the sector. The technologies continue to evolve, but the engineering discipline itself remains remarkably consistent.
Designing for hardware that does not yet exist
The greatest uncertainty facing the industry is not how to cool today’s AI clusters. It is how to build facilities capable of supporting several generations of technology that nobody has yet seen.
Data centres are commonly designed with operational lifespans of between twenty-five and thirty years. During that period servers may be replaced numerous times, processors will become significantly more powerful and electrical architectures are already beginning to move from traditional AC distribution towards higher-capacity DC designs. Every one of those changes has implications for cooling.
“When you consider that a data centre is built with an expected lifespan of twenty-five to thirty years, you are designing for technology you understand today, but you do not know what comes next,” Aguilar explains. “It is very difficult to future-proof a data centre for something you simply do not know. The challenge is understanding today’s load while building enough flexibility to accommodate whatever comes afterwards.”
That uncertainty has become even more pronounced as AI hardware evolves at extraordinary speed. Nvidia’s annual product roadmap has almost become shorthand for an industry where each generation significantly increases compute performance, while simultaneously placing greater demands on power delivery and thermal management. The cooling system installed today therefore has to accommodate processors that may not appear until years after the building itself has opened.
For cooling designers, flexibility has become every bit as valuable as outright performance. Infrastructure that can adapt to changing heat loads over successive hardware refreshes is proving considerably more valuable than systems optimised around a single generation of servers.
Legrand has seen that progression repeatedly. Rear-door heat exchangers originally installed to support cabinets dissipating only a few kilowatts have remained in service as customers refreshed their IT equipment, with some now supporting heat loads well beyond sixty kilowatts without replacing the cooling architecture itself.
“What is interesting about rear-door cooling is that it stays close to the heat source and adapts to whatever is installed inside that cabinet,” Aguilar continues. “We have installations where the cooling infrastructure has remained constant while the IT equipment has been refreshed several times. The racks have evolved from five or seven kilowatts to more than sixty kilowatts, but the cooling architecture has remained in place.”
The implication is significant. Futureproofing may never be perfect, but designing adaptable infrastructure offers operators a far better chance of accommodating whatever comes next.
The boundaries inside the data centre are disappearing
The rapid growth of AI is also changing relationships across the supply chain. For years, IT vendors focused primarily on processors, servers and networking; while cooling and electrical infrastructure were treated almost as separate disciplines. Those distinctions are becoming increasingly blurred because one can no longer evolve independently of the other.
“The white space and the supporting infrastructure used to operate almost independently,” Aguilar says. “Today those lines are much more blurred. Companies developing AI hardware work much more closely with organisations designing the mechanical and electrical infrastructure because their products depend on those environments performing exactly as intended.”
This closer collaboration reflects a broader reality. Cooling can no longer be considered something that happens after compute decisions have been made. Thermal management, electrical distribution and server architecture are becoming increasingly interdependent, particularly as rack densities continue to rise.
The transition towards higher-voltage DC power architectures illustrates that convergence. While discussions often focus on electrical efficiency, those changes also influence the heat generated within each rack, reinforcing the need for cooling systems that remain adaptable rather than fixed around today’s assumptions.
Heat is becoming part of the energy equation
One of the less discussed consequences of higher-density liquid cooling is that it changes the quality of the heat leaving the data centre. Historically, chilled-water systems supplied relatively cold water and received it back only slightly warmer, limiting opportunities for useful heat recovery. Modern liquid cooling operates at much higher temperatures, creating genuine opportunities to reuse thermal energy rather than simply rejecting it into the atmosphere.
“We used to supply water at around twelve or thirteen degrees and receive it back at perhaps twenty degrees,” Aguilar explains. “Today we might supply cooling water at thirty-five to fourty degrees and see it return at fifty-five or sixty degrees. That creates genuine opportunities to use the heat rather than simply reject it.”
Countries across Scandinavia have demonstrated the potential of district heating for years, but higher return temperatures are making similar approaches increasingly practical elsewhere. Universities, hospitals and industrial campuses are emerging as particularly attractive locations because they combine significant computing requirements with consistent local demand for heating.
Legrand is already collaborating with organisations that are using waste heat to provide heating for student accommodation, demonstrating how cooling infrastructure can be integrated into a local energy ecosystem rather than simply discharging heat into the atmosphere, as is common with many conventional cooling systems.
The significance extends beyond sustainability. Waste heat is gradually shifting from being an unavoidable by-product to becoming another resource capable of improving the economics of data centre investment.
Challenging the bubble
Predictions of an AI bubble frequently dominate headlines, drawing comparisons with previous technology booms and suggesting infrastructure investment may eventually slow. Aguilar believes those comparisons risk overlooking a more fundamental trend. “I am not convinced the bubble is going to burst,” he says. “AI has highlighted a whole series of challenges around sovereign data, and that is creating entirely new categories of infrastructure. Countries are becoming much more concerned about where data resides and how it is protected. That will generate additional demand for sovereign data centres.”
Some AI application companies will inevitably disappear as markets mature, just as countless internet businesses failed during the dot-com era. Infrastructure follows a different logic. Regardless of which software platforms ultimately succeed, the underlying demand for compute continues to grow, and every additional watt consumed still becomes heat that must be removed.
The more immediate challenge is whether the industry can deliver infrastructure quickly enough. “I think the biggest challenge for the whole industry is the supply chain,” Aguilar says. “It is about how we work more closely with customers and with our suppliers to make sure we can meet demand. As electrical architectures move from AC towards DC, heat loads will continue to increase. The challenge is not whether demand exists. The challenge is how we keep pace with it.“
It is a refreshingly pragmatic conclusion because it shifts attention away from speculation and back towards execution. AI will continue to evolve, processors will continue to become more powerful and cooling technologies will continue to improve, but none of those developments alters the fundamental challenge facing designers today.
The future of cooling will not be defined by finding a miracle technology that somehow overturns the laws of physics. It will be defined by building enough flexibility into today’s infrastructure to accommodate tomorrow’s uncertainty. That is a far more difficult engineering problem because nobody yet knows exactly how much heat the next generation of AI hardware will produce. What the industry does know, however, is that every previous prediction has underestimated where compute was heading, and history suggests betting against higher heat densities would be an increasingly dangerous assumption.
