AI data centres are testing the limits of power and water systems
- Editorial Team SDG9
- 2 hours ago
- 6 min read
Published on 31 July 2026 at 03:44 GMT
By Editorial Team SDG9
Â
Artificial intelligence data centres are becoming a new source of concentrated demand for electricity, cooling and grid capacity. The pressure does not arise from artificial intelligence alone, since data centres also support cloud services, streaming, finance and everyday digital activity. Yet the rapid deployment of specialised chips for training and operating AI models is raising the power density of facilities and accelerating construction in locations where electricity networks and water supplies may already be constrained.
The issue is no longer limited to the environmental performance of individual buildings. Large clusters can require new substations, transmission lines, power plants and water infrastructure. Their arrival may affect household bills, industrial access to electricity, local planning and the timing of climate commitments. The central question is whether data-centre growth can be aligned with energy and water security, rather than treated as a separate technical matter.
Demand is growing faster than the wider power system
The International Energy Agency projects that global data-centre electricity consumption could more than double to around 945 terawatt-hours by 2030 in its base case, just under 3 per cent of global electricity use. This is a projection rather than a guaranteed outcome, and the range depends on AI adoption, hardware efficiency, utilisation rates and the pace of new construction. Even so, the expected annual growth is far faster than electricity demand in most other sectors.
The national and local effects can be much larger than the global percentage suggests. The United States Department of Energy, drawing on research by Lawrence Berkeley National Laboratory, reported that data centres used approximately 4.4 per cent of United States electricity in 2023 and could account for an indicative 6.7 to 12 per cent by 2028. Such growth can be especially difficult where projects are concentrated in a few metropolitan regions and compete for the same grid connections.
Renewable-energy contracts can reduce the annual carbon footprint associated with this demand, but procurement claims require careful interpretation. A company may buy enough renewable electricity over a year to match its consumption while its facilities continue drawing power from a fossil-intensive grid during particular hours. Hourly carbon-free energy matching, grid location and the additionality of new generation provide a clearer picture than annual certificates alone.
The IEA expects renewables to meet nearly half of the additional global electricity needed for data centres to 2030 in its base case. Natural gas and coal are also expected to supply part of the increase, while nuclear generation becomes more significant later in the decade. The result is that AI expansion can support new clean-energy investment while still prolonging fossil generation if grid planning, storage and flexible demand do not advance at the same pace.
Cooling creates a second resource question
Electricity is only part of the footprint. Servers convert most of the electricity they consume into heat, which must be removed continuously. Some facilities use air-based systems, while others use chilled water, evaporative cooling, direct-to-chip liquid cooling or immersion systems. The resulting water footprint of AIÂ varies greatly with climate, design, operating temperature, local water source and the electricity generation supplying the site.
Water impacts are both direct and indirect. Direct consumption may occur when water evaporates from cooling towers. Indirect consumption occurs at power stations and across electricity supply chains. A system that reduces on-site water use may consume more electricity, while a highly efficient evaporative system may place greater pressure on a water-stressed catchment. This makes a single global estimate less useful than transparent, site-specific reporting.
The choice of cooling technology is also changing as AI chips become more powerful. Direct liquid cooling can remove heat more efficiently from densely packed processors and may enable higher operating temperatures, reducing the need for mechanical refrigeration. Closed-loop systems can limit routine water withdrawals inside the facility, but the final heat-rejection system still matters. Dry coolers reduce water consumption but may require more energy or larger equipment during hot weather. Hybrid systems can switch between modes according to temperature and water availability.
Location therefore becomes an environmental decision. Building a water-intensive facility in a cool, water-abundant region is different from placing it in a drought-prone area with competing agricultural and household needs. Local authorities need information about expected withdrawals, consumption, wastewater, seasonal peaks and emergency operating conditions before approving projects. Communities also need clarity about whether public infrastructure costs will be recovered from developers or shifted to other users.
Efficiency gains are real, but rebound effects remain
Processor design, software optimisation and better facility management can reduce the energy required for a given task. Specialised accelerators can perform some AI operations more efficiently than general-purpose hardware. Higher server utilisation, model compression, smaller task-specific models and scheduling work for periods of cleaner electricity can also reduce impacts. These measures form part of energy-efficient AI computing, but their benefits depend on how quickly total demand grows.
Historically, efficiency improvements have helped digital services expand without electricity use rising at the same rate. AI may weaken that pattern because companies are deploying much larger computing systems and using them for more products. Cheaper or faster computation can stimulate additional demand, a rebound effect that offsets part of the efficiency gain. Environmental assessment must therefore measure absolute electricity, water and emissions as well as efficiency per calculation.
Common indicators such as power usage effectiveness show how much facility energy is used beyond the computing equipment. Water usage effectiveness measures water consumption relative to IT energy. Both are useful, but neither reveals the full impact without information on energy sources, local water stress, hardware utilisation and the useful work delivered. A low facility overhead does not make an underused or carbon-intensive computing system sustainable.
Reporting is moving from voluntary claims towards public rules
Stronger environmental reporting for data centres is essential because corporate disclosures are often aggregated across countries and can obscure local impacts. The European Union has begun establishing a common reporting framework under the Energy Efficiency Directive. Commission Delegated Regulation (EU) 2024/1364Â sets out information and performance indicators for data centres, creating a basis for greater comparability even as the system continues to develop.
Effective disclosure should distinguish electricity consumption, contracted renewable supply, hourly grid emissions, back-up generation, water withdrawal and water consumption. It should identify the source and catchment of water, explain the cooling method and report performance during drought or extreme heat. Independent verification and facility-level data would make it harder to shift impacts between locations or present efficiency ratios without absolute totals.
Planning authorities and regulators may also need to assess cumulative impacts. Several efficient facilities can still overwhelm a regional grid or watershed when built together. Connection agreements, water permits and land-use approvals can require demand forecasts, efficiency standards, heat-reuse studies and contingency plans. In some locations, operators could provide flexibility by reducing non-urgent computing during grid stress, but this requires technical capability and enforceable arrangements rather than broad promises.
The test is whether infrastructure decisions match public objectives
The expansion of AI infrastructure connects directly with SDG 9 (industry, innovation and infrastructure)Â because the goal concerns resilient infrastructure and sustainable industrial development. The connection is not an assumption that all digital growth is beneficial. It is a requirement to judge whether innovation is supported by energy and water systems that remain reliable, affordable and environmentally responsible.
Renewable procurement, efficient processors and alternative cooling can materially reduce impacts. None is sufficient by itself. Annual renewable matching can conceal difficult hours, efficient chips can encourage greater use, and low-water cooling can increase electricity demand. The strongest approach combines clean electricity supply, efficient hardware and software, water-sensitive siting, flexible operations and public reporting that reflects absolute as well as relative performance.
AI data centres need not inevitably conflict with climate and water-security objectives. Avoiding that conflict, however, requires decisions before facilities are built, not only technical adjustments afterwards. Grid investment, generation planning, cooling design, water allocation and disclosure rules must be considered together. Without that coordination, the environmental costs of artificial intelligence may be transferred to electricity consumers, water users and communities that receive few of its promised benefits.
Further information:
• International Energy Agency, Its Energy and AI analysis provides global projections for data-centre electricity demand and the expected energy-supply mix.
• Lawrence Berkeley National Laboratory, Its research examines data-centre energy use, cooling systems and the factors that determine workload-level water consumption.
• European Union, Commission Delegated Regulation (EU) 2024/1364 establishes the first phase of a common sustainability reporting scheme for data centres.