As organisations deploy increasingly powerful AI and high-performance computing workloads, the challenge is no longer simply how to power more infrastructure. It is how to support dramatically higher computational density while making responsible use of energy and resources.
That challenge starts at the chip. For years, sustainability discussions around data centres focused largely on power. Operators worked to improve efficiency, increase renewable energy use and reduce waste across their facilities.
Now, the massive buildout of AI is changing the conversation.
Today’s AI processors deliver extraordinary performance, but they also generate unprecedented amounts of heat. The more power pushed into a rack, the greater the demand placed on the cooling systems that support it.
Without effective heat removal, performance suffers. Operators may be forced to reduce density, limit processor performance or accept lower utilisation from infrastructure designed to do more. In each case, valuable resources are consumed without delivering their full potential, increasing cost.
The scale of the challenge is growing quickly. The International Energy Agency estimates that data centres consumed around 415 terawatt-hours of electricity in 2024 and projects this figure could reach around 945 terawatt-hours by 2030, with AI a major driver.
That growth makes efficiency increasingly important. Expanding AI infrastructure is one challenge; expanding it sustainably is another.
Power and cooling: inseparable challenges
Every watt consumed by an AI compute infrastructure ultimately leaves the facility as heat. That simple reality means cooling can no longer be treated after-the-fact as a supporting technology operating quietly in the background. It has become a strategic part of how AI data centres are designed, operated and scaled.
Traditional air cooling continues to play an important role across many environments, but AI workloads are pushing it towards its practical limits. As rack densities rise, facilities require more airflow, more fan power, and more space dedicated to thermal management.
In some cases, the limiting factor is no longer available compute hardware or even available power. It is the ability of the facility to remove heat efficiently. This situation is called stranded power and many data centres suffer from the limitations of their cooling, not access to power.
This creates a challenge that extends beyond engineering. Every watt spent moving heat is energy not being used to generate computational value. As workloads grow, operators need cooling strategies that make better use of constrained power resources while avoiding additional operational complexity.
Securing power capacity is only part of the equation, making the most of that power is equally important.
Rethinking how heat is removed
This is why liquid cooling is attracting increasing attention.
Because liquids transfer heat more effectively than air, and two-phase systems can absorb more heat through vaporisation, liquid cooling can remove heat efficiently at or near its source and support significantly higher compute densities than traditional air-based systems. Furthermore, water-based systems introduce their own considerations around infrastructure, maintenance, and resource usage.
In regions where water availability is under pressure, those considerations become particularly important. Two-phase direct-to-chip cooling offers a different model.
The technology places a cold plate directly on the processor, where a dielectric fluid absorbs heat through controlled boiling. As the fluid changes from liquid to vapour, it efficiently carries heat away from the chip before condensing back into liquid and returning through a sealed loop.
Rather than relying solely on a rise in fluid temperature to carry heat away, two-phase cooling uses the fluid’s phase change – vaporisation – to absorb substantial amounts of thermal energy. This enables high heat loads to be removed at a relatively stable operating temperature and, in many applications, with lower fluid flow rates than comparable single-phase systems. When a dielectric working fluid is used, it also provides electrical isolation around sensitive electronic components, significantly reducing the risk of electrical damage and corrosion associated with fluid leaks. The result is a cooling architecture designed specifically for the demands of high-density AI infrastructure.
Turning a challenge into an opportunity
More efficient cooling delivers clear operational benefits by reducing cooling overhead and directing a greater share of available power to compute. In AI factories, more power to compute translates directly into greater productive capacity and stronger economics.
While more power to compute drives the business imperative, every watt used for compute eventually becomes heat, creating additional heat load for the cooling loop to manage. Traditionally, that heat has been viewed as a by-product to be removed and discarded. The opportunity is to reuse all this heat in ways that reduce the overall energy requirements around the data centre locale.
Recovered heat can support nearby buildings, district heating networks and other local applications, allowing data centres to contribute energy as well as consume it. Instead of functioning solely as large power users, facilities can become part of wider local energy ecosystems.
This reflects a broader shift in how AI infrastructure can be viewed moving forward. Sustainability is no longer simply about reducing environmental impact, it should be increasingly about making better use of every resource entering the facility and capturing greater value from what leaves it.
That shift is already underway as operators evaluate cooling technologies capable of supporting higher densities, lower overheads and more efficient resource use.
AI infrastructure will continue to expand. The real question is whether that growth will be constrained by legacy cooling assumptions or enabled by technologies that make more efficient use of power, space, water, and compute infrastructure. Data centre operators that rethink cooling as a strategic part of infrastructure design will be best placed to respond to that question. They can deliver the capacity of AI demands to improve efficiency, reduce resource pressure and support more sustainable long-term growth.
