Why Carbon Alone Cannot Explain AI’s Environmental Impact

By 2030, global data centres could consume 945 terawatt-hours (TWh) of electricity annually, use 9.3 trillion litres of water and be associated with a land footprint exceeding 14,500 square kilometres.

According to a new report from the United Nations University Institute for Water, Environment and Health (UNU-INWEH), those figures suggest that sustainability discussions around AI infrastructure may need to account for a broader range of environmental impacts than carbon emissions alone.

Traditionally, sustainability discussions around data centres have largely focused on carbon emissions, renewable energy procurement, and energy efficiency.

The report argued that as AI infrastructure expands globally, its environmental footprint should be evaluated through three interconnected lenses: carbon, water, and land. The report’s central observation is that these impacts do not always move together, creating trade-offs that become more visible as infrastructure grows.

AI Infrastructure is Becoming a Significant Consumer

The scale of AI infrastructure continues to expand.

According to the report, global data centres could consume 945TWh of electricity annually by 2030, equivalent to almost 3% of projected global electricity demand. That is nearly triple the combined annual electricity consumption of Pakistan, Bangladesh, and Nigeria, countries with a combined population exceeding 650 million people.

The associated environmental footprint extends beyond electricity consumption.

UNU researchers estimate that by 2030, global data centre electricity use could be linked to:

• 399 million tonnes of carbon emissions
• 9.3 trillion litres of water consumption
• 14,500 square kilometres of land footprint

To illustrate the scale, the report noted that if global data centres were treated as a country, their estimated 448TWh of electricity consumption in 2025 would have made them the world’s 11th-largest electricity consumer, ranking between France and Saudi Arabia.

These figures highlight the extent to which digital infrastructure now operates at a scale where broader resource implications become increasingly relevant.

Carbon Tells Only Part of the Story

One of the report’s most significant findings is that environmental impacts do not necessarily move in the same direction.

Much of the technology sector’s sustainability reporting has focused on reducing carbon emissions. The report argues that this can obscure other environmental pressures.

The authors, Dr Miriam Aczel, Professor Kaveh Madani, Professor Tshilidzi Marwala, and Dr Mir Matin, point to examples where lower-carbon energy sources may require substantially greater water consumption or land use. Switching from coal to bioenergy, for example, can reduce carbon emissions by around 70%, while increasing water requirements more than thirtyfold and land requirements by around one hundredfold.

The implication is not that decarbonisation is undesirable. Rather, it highlights how infrastructure decisions increasingly involve trade-offs across multiple environmental objectives.

As the report notes, “low-carbon” does not automatically mean “low-water” or “low-land”.

For operators and policymakers, that creates a more complex sustainability discussion than one centred on emissions alone.

Inference is Becoming the Dominant Workload

The report also challenges another common assumption.

Much public discussion around AI infrastructure has focused on the energy required to train large models. According to UNU, this increasingly misses where most energy consumption occurs.

Once deployed, AI systems spend the majority of their operational life serving user requests. The report estimates that 80% to 90% of total AI energy use is associated with inference rather than training.

ChatGPT alone is estimated to process around 2.5 billion prompts per day, consuming approximately 383 gigawatt-hours of electricity annually.

The observation aligns with developments already visible across the infrastructure sector. Demand for inference environments continues to grow as organisations move AI systems into production, placing greater emphasis on the operational phase of AI rather than model development alone.

The report, therefore, suggests that environmental discussions may need to focus as much on how AI is used as how it is trained.

Data Centres Are Becoming Part of Wider Systems

The report repeatedly frames data centres not simply as compute facilities, but as consumers of energy, water, and land.

That perspective mirrors developments already visible elsewhere in the industry.

Recent European initiatives increasingly position data centres as participants within wider infrastructure systems through grid flexibility, storage, demand response, and energy coordination. As explored in Techerati’s recent analysis, the relationship between digital infrastructure and energy systems is becoming more interconnected.

The UNU report extends that perspective further. It suggests that energy, water, and land should increasingly be considered as interconnected resource systems rather than separate sustainability categories.

Sustainability Reporting May Be Entering a New Phase

Perhaps the report’s most important contribution is the question it raises about measurement.

Sustainability reporting within digital infrastructure has largely centred on emissions reduction, renewable energy procurement and energy efficiency metrics. Those measures remain important.

The report argued that they may not provide a complete picture of AI infrastructure’s environmental footprint as systems scale. Water consumption, land impacts, infrastructure siting, and lifecycle effects are becoming more relevant parts of the discussion.

This arrives alongside wider regulatory developments already taking place in Europe. Recent reporting requirements under the Energy Efficiency Directive, as well as discussions around data centre performance standards, reflect growing interest in how infrastructure impacts are measured, compared, and governed.

The issue is not whether current reporting is inaccurate, but whether the industry’s most visible metrics continue to capture the full range of environmental considerations associated with large-scale AI deployment.

Local Impacts Are Becoming More Visible

The report also highlights how the benefits and costs of AI infrastructure are not always distributed evenly.

In Ireland, data centres accounted for 21% of total metered electricity demand in 2023, leading grid operators to pause new approvals around Dublin until 2028.

The report also points to examples in Mexico and Uruguay where data centre expansion has coincided with water stress and drought conditions.

These examples are not presented as arguments against AI infrastructure growth. Rather, they illustrate how local infrastructure planning, resource availability and community impacts are becoming increasingly important considerations as capacity expands.

Sustainability is a Measurement Challenge

The report ultimately frames AI sustainability as a governance issue rather than a purely technical one.

Its proposed framework for a “responsible AI ecosystem” is built around six principles: transparency, efficiency by design, environmental justice, lifecycle responsibility, global cooperation, and sustainable use.

The broader contribution of the report, however, may be the challenge it presents to how sustainability itself is measured.

As AI infrastructure expands, carbon, water, and land are becoming harder to assess in isolation. Decisions that improve performance in one area may create pressures elsewhere.

The report’s central argument is not that AI infrastructure should stop growing. It is that infrastructure decisions increasingly involve trade-offs across multiple environmental systems. As those trade-offs become more visible, the frameworks used to measure sustainability may need to evolve alongside the infrastructure they are designed to assess.