Geoeconomics & Supply Chains · Horizon 2026–2030

As AI Makes Data Centres Hotter

Who Pays the Price for Keeping Them Cool?

By: Muskan Gupta
Published: 14 September 2026
Horizon: 2026–2030
Pillar: Geoeconomics & Supply Chains
Anchors: New Delhi
Reading time: 8 minutes

Where the heat goes: Google's AI builds in Visakhapatnam and northern Finland · Map: ALTARA Map Studio, Natural Earth data

The AI boom is usually measured in chips, models, computing power and electricity. The race is to build bigger models, faster and more powerful systems. But heat, a physical problem, is becoming more difficult to ignore.

Computation is just a way of a processor to convert electrical energy into heat due to resistance. As artificial intelligence systems become more powerful, and operators pack more computing capacity into individual racks, more heat has to be removed. Cooling, once a support function, is becoming central to AI infrastructure. The International Energy Agency (IEA) said data centers consumed 17% more electricity in 2025 than in 2024, with electricity use at AI-focused data centers jumping 50% in the same period. The IEA’s Energy and AI report projects that total data-centre electricity demand will more than double from 2024 levels to around 945 TWh by 2030. The numbers reflect the magnitude of the expansion, but they do not reflect the infrastructure to support it. The AI boom is turning into a race to control heat.

The More Powerful the Chip, the Harder It Is to Keep It Cool

Fundamentally, a data center is a place where electrical energy is converted into computation, and ultimately heat. This is always the case. What has changed is the density of computation.”

AI workloads are highly dependent on specialised processors such as GPUs and AI accelerators; it is a specialised processor, such as graphics processing units (GPUs) designed to handle the massive amounts of calculations required to train and run AI models. Where once the AI infrastructure was about distributing relatively small computing loads across a huge number of traditional servers, today it is about packing vast amounts of computing power into high-density racks. That's a different thermal challenge. The issue is not just that a data center has more machines. This is because individual racks are now capable of generating much more heat than older generations of infrastructure were designed to handle.

For example, Microsoft claims its AI-centric Fairwater infrastructure can support about 140 kW per rack and 1,360 kW per row through liquid-based cooling. The principle is similar to that of a car radiator, in which a liquid absorbs heat generated in the engine and carries it away to release it elsewhere. In a data center, the liquid removes heat from processors and other high-performance computing parts. Google has also been collaborating with the Open Compute Project on infrastructure designed to scale rack power from around 100 kW to 1 MW. The importance of that trajectory is that nearly all the electricity powering computing gear eventually has to leave the facility as heat.

Thermal management can be a significant part of the infrastructure equation, with up to 33-40% of data-centre energy use going to cooling at some facilities, says the Department of Energy’s ARPA-E. So the challenge is less to squeeze more computing equipment into a building than to create an environment where that equipment can run constantly without overheating.

Air Was Once Enough. AI Is Changing the Equation

Fans have traditionally been used to cool the data-centre equipment by blowing air over the hot components, and the systems cool the heated air before recirculating it.

But as heat gets more concentrated in smaller spaces, it gets harder to move enough air over high-power processors. The higher the heat density, the less efficient the conventional air cooling. That’s driving interest in liquid cooling, because liquids transfer heat more efficiently than air and can remove it more directly from processors at scale.

Systems may use immersion and facility level liquid cooling and can circulate coolant through cold plates attached to chips. Microsoft says its latest AI-optimized designs use direct-to-chip cooling in closed loops, while NVIDIA’s high-density rack systems are increasingly designed around liquid cooling. But liquid cooling is not a panacea, it shifts the resource equation rather than eliminates it.

Closed-Loop Cooling Could Change the Water Equation

Mechanical or closed-loop systems can reduce water use but may increase electricity demand. Microsoft says its newer chip-level cooling designs are intended to limit that trade-off. Operators must balance water, power and cost. There is no one cooling technology for all sites.

The response from the industry is not limited to the choice between traditional air cooling and water-intensive cooling. One significant development is the use of closed loop liquid cooling.

Instead of a continuous use of fresh water, these systems utilise a predetermined volume of liquid that is circulated through the cooling infrastructure. A closed loop cooling system has coolant moving in a continuous circuit. It absorbs heat from the processors, processes it to a heat exchanger, and then passes it into another cooling system. The coolant then flows back to the processors to pick up heat again. Since the liquid is recycled rather than constantly replaced this design can greatly reduce water usage where it eliminates evaporative cooling.

Microsoft says its newer AI data-centre designs use this approach and can avoid more than 125 million liters of annual water consumption per data center compared to designs based on its FY2024 average withdrawal water use efficiency (WUE). The water in the cooling loop is initially filled and then continually recirculated, not evaporated during normal operation, the company says. That’s not to say liquid-cooled AI infrastructure doesn’t have an environmental footprint.

Still, the plant will need power. The cooling equipment still needs to be manufactured and maintained. Somewhere the heat still has to be rejected. And the infrastructure up front can be more complex and expensive. But it is a good point: The future of data-centre sustainability might be less about choosing one cooling technology and more about designing the entire system around local conditions.

The Geography of AI is Changing

That’s where cooling becomes a strategic, rather than engineering, problem. It is not just about finding somewhere with land and power to put a data center. Increasingly its location has to answer another question:

Will the surrounding environment be able to support the heat, water and power needs of the facility? Climate matters. Cooler climates can reduce the strain on cooling systems. A recent example of this logic is Google’s decision to expand its AI infrastructure in Finland. The tech giant announced it would invest €13 billion in three new data centers in northern Finland, citing the region’s cold climate and stable electricity supply as an advantage for energy-intensive infrastructure.

Water availability is also important. So does the grid. Even a place with plenty of renewable electricity can run into water limits. A site with abundant water may have little grid capacity. A naturally cooler climate may be distant from major hubs of connectivity or customers.

What this means is that AI infrastructure is becoming increasingly geographic.

The location can impact the economics of the digital services ultimately delivered from that location.

The Cost Does Not Stay Inside the Data Centre

Who’s footing the bill to keep AI cool? Tech companies also help pay for some of the expense with cooling equipment, power infrastructure and facility construction. While electricity use increases operating costs, some of the burdens can be moved outside the data-centre boundary.

Large facilities may require new electricity generation and transmission infrastructure, roads and water connections. These demands can be highly sensitive in water-stressed regions. When a data center uses resources that nearby communities also depend on, cooling becomes a resource allocation challenge, not just a technical one.

India's AI Ambitions Bring the Question Home

India's ambition is to be an important place for digital infrastructure and AI. Global technology firms along with Indian enterprises are making efforts to improve the availability of data centres by developing numerous infrastructures with an emphasis on AI.

The project that Google has unveiled in Visakhapatnam is a vivid example of that. The company announced its new plan that is going to cost approximately $15 billion with the capacity of 1 GW.

However, the project also shows how ambitions in building data-centres can meet problems connected with resources. According to the Reuters report dated August 2026, environmentalists and locals have raised concerns related to water availability and possible impact on the nearby ecosystem. Moreover, the report speaks about the existing pressure on the city’s water supply while Google claims it has integrated certain measures such as air-cooling technologies.

The initiatives that India is intending to take are connected with choosing state and location for data centres, types of cooling systems, electricity generation methods, and water demand management. Brookfield estimated, according to the latest release dated July 2026, that approximately 6.5 GW of AI data-centre capacity would be launched within the next five years in India.

From Power Usage to Resource Usage

Power Usage Effectiveness (PUE) is a ratio of the electricity used in a data center to run IT equipment to the electricity used to run supporting systems such as cooling. A broader view of energy, water, carbon and climate is needed in the AI era. The Green Grid’s Data Center Resource Effectiveness framework considers energy efficiency, water use, water stress and climate in a holistic manner. A low-water facility could be nonviable if it significantly increases electricity use, while an efficient site could stress a water-stressed area.

The AI infrastructure race is often described as a race for better chips, bigger models and more compute power. But its constraints are increasingly physical: enough electricity to power chips, grid capacity to deliver it, cooling capacity to remove the heat, and local infrastructure to support the facility. “Data centers should be measured on efficiency and reliability, not just the number of GPUs crammed in.

Who Really Pays the Price?

There’s no single answer. Technology companies are building ever larger facilities, electricity systems are facing ever higher demand, and water systems might be strained in resource-constrained areas. Costs could also be indirectly passed on to consumers while communities could be affected where water, electricity or land are contested.

It’s not a question of whether we should stop AI, but who should pay for its expansion. Power, land, connectivity and economics are not the only things to consider from the beginning, cooling is too.

As data centers become hotter, the sustainability of AI will depend not only on computing power but on how intelligently the resources behind it are managed. Ultimately, the question is simple: where does the heat go?