AI’s next challenge is not just chips. It’s electricity, water and land

Every AI query, image, video or reasoning task requires computing power. That computing takes place in data centres packed with specialised chips and servers, which consume electricity and generate heat. Keeping those systems running therefore requires not just more power, but also increasingly sophisticated cooling infrastructure.

The scale of that demand is growing rapidly. The International Energy Agency (IEA) says global data-centre electricity demand increased 17% in 2025, while electricity consumption from AI-focused data centres rose 50%. That makes AI increasingly an energy and infrastructure story, as much as a technology story.

The environmental impact will depend on how quickly AI adoption grows, how much more efficient the technology becomes and where the electricity and other resources needed to power it come from.

AI driving new data-centre construction boom

Data centres already account for around 1.5% of global electricity consumption, according to Our World in Data. But their impact is particularly significant in regions where large clusters of facilities are being built.

The US is at the centre of the current expansion. A Washington Post analysis published on September 28 found that the US has nearly 100 hyperscale data centres, with around 120 under construction and another 460 planned. It defines hyperscale facilities in this context as sites consuming at least 100 megawatts — roughly the electricity demand of 80,000 homes.

These facilities require much more than a building filled with servers. Developers need large amounts of land, fibre connections, electricity, substations and cooling infrastructure.

That is pushing some projects into rural areas, where land can be cheaper and large sites are easier to develop.

The physical footprint can therefore extend well beyond the data centre itself. New transmission lines, roads, substations and power infrastructure may also be required to support these facilities.

This is changing the geography of the AI boom. What began largely as a technology-industry story is increasingly affecting rural communities and local energy systems.

Electricity becoming the central climate question

AI models require computing power both during training and after deployment. As AI applications become more sophisticated, the amount of computing required for individual tasks can also increase.

The IEA says some newer AI applications can consume hundreds or even thousands of times more energy per query than simple text generation.

That creates a difficult equation: AI adoption increases computing demand, which increases data-centre demand, which in turn increases electricity consumption.

Efficiency improvements can reduce the amount of electricity required for individual computations. But if the number and complexity of AI tasks grow faster than efficiency improves, overall electricity demand can still rise.

Where that electricity comes from will determine much of AI’s climate impact. Data centres connected to grids that rely heavily on fossil fuels can result in higher greenhouse-gas emissions. Where additional electricity comes from renewable or nuclear generation, the carbon impact can be lower.

Technology companies are therefore investing heavily in clean-energy procurement. Google said in its 2025 Environmental Report that its data-centre electricity demand increased 27%, while its data-centre energy emissions fell 12% in 2024. The company also said it had contracted more than 8 GW of clean-energy generation.

Microsoft’s 2025 sustainability report said the company had contracted 34 GW of new renewable energy across 24 countries.

However, clean-energy contracts do not necessarily mean that a particular data centre is powered entirely by renewable electricity at every moment. Data centres generally remain connected to electricity grids containing multiple sources.

The climate impact therefore depends on the wider power system as well as individual corporate commitments.

Water and land becoming part of AI infrastructure debate

Electricity is not the only resource under pressure. AI servers generate enormous amounts of heat and require cooling systems to keep equipment operating safely. Some cooling technologies use significant amounts of water, creating potential conflicts in regions where data centres compete with agriculture, households and ecosystems.

Microsoft says its newer direct-to-chip cooling technology can save more than 125 million litres of water per facility each year.

Google, meanwhile, reported that it replenished 4.5 billion gallons of water in 2024 and increased the share of freshwater consumption it replenished from 18% in 2023 to 64%.

But water use is particularly dependent on location. A data centre built in a water-rich region presents a different challenge from one built in a drought-prone area. That is why governments are increasingly looking at data-centre water consumption alongside electricity use.

Reuters reported on September 21 that the European Commission had proposed rules requiring data centres in Europe to disclose information about their energy and water efficiency.

Land use is another emerging issue, particularly in the US. Large facilities can require substantial parcels of land, while associated infrastructure can involve forest clearing, farmland conversion and new power infrastructure.

Can AI’s efficiency gains offset its environmental cost?

The environmental impact of AI is not entirely one-directional. AI itself could help reduce energy consumption and emissions in other parts of the economy. The IEA says AI could improve electricity-grid management, energy forecasting and the integration of renewable power.

Google has also reported potential emissions reductions from AI-enabled products. Its 2025 environmental report said five products, including Green Light and fuel-efficient routing in Google Maps, enabled an estimated 26 million tonnes of greenhouse-gas-equivalent emissions reductions.

The technology industry’s own infrastructure is also becoming more efficient.

Google says its Ironwood AI processor is nearly 30 times more power efficient than its first Cloud TPU from 2018. Microsoft is investing in direct-to-chip and liquid cooling technologies, while companies across the sector are exploring nuclear, geothermal, solar and other sources of electricity.

But efficiency alone may not settle the issue. If AI applications become dramatically more popular and computationally intensive, improvements in the efficiency of individual chips or data centres may be offset by the sheer increase in demand.

The IEA has already warned that AI’s future energy demand is difficult to predict because all three factors — efficiency, adoption and increasingly energy-intensive applications — are changing rapidly.

For now, it is becoming clearer that the AI race is also a race to build power, cooling, land and other physical infrastructure. As that infrastructure expands, the environmental footprint of AI will increasingly be determined not only by what the technology can do, but by how the world powers it.