Earth Day 2026: Can the world afford AI’s growth without rethinking sustainability

Artificial intelligence has become part of the world’s everyday machinery. It writes emails, recommends what to watch, flags fraud, designs drugs and increasingly helps companies make decisions at speed. It feels weightless. Frictionless.

It is neither.

On Earth Day, the question is not whether AI is useful. That debate is settled. The more pressing one is harder to answer and far less visible. What does it take to run intelligence at this scale, and can the planet afford it?

The infrastructure behind the illusion

AI often feels like software. In reality, it is infrastructure.

Every prompt, prediction or recommendation runs through data centres filled with servers that consume electricity constantly and generate heat that must be managed. Cooling those systems requires water, often in vast quantities.

According to the International Energy Agency, data centres accounted for a rising share of global electricity demand, with AI estimated to contribute between 15% and 20% of that usage in 2024. The same estimates put carbon emissions from data centres at around 182 million tonnes.

Water is harder to track but no less significant. Global data centre water use reached about 560 billion litres in 2023, though the share attributable to AI remains unclear.

That gap in clarity is part of the problem. The technology is scaling faster than the data that explains its impact.

A footprint that is still being counted

Researchers trying to isolate AI’s environmental cost are working with incomplete information. Even so, the early estimates are striking.

Alex de Vries-Gao at the University of Amsterdam has suggested that AI-related emissions could range between 32 million and nearly 80 million tonnes of carbon dioxide each year. His calculations also place annual water consumption between 312 billion and 764 billion litres.

Those numbers are not definitive. They are, however, directionally important. They suggest that AI’s footprint is larger than many assumed and still growing.

What complicates matters is that AI’s consumption is not limited to what users see. The visible part, asking a question or generating an image, is only a fraction of the total load.

Training models is far more resource intensive. It involves running complex computations repeatedly over long periods. Researchers say this phase is often excluded from public estimates, which can make the overall impact appear smaller than it is.

The small numbers that add up

Some technology companies have tried to explain AI’s resource use in relatable terms. Estimates suggest that a single query might use a small amount of electricity and a few drops of water.

On its own, that sounds negligible.

Scaled across millions of users, it is not.

This is where AI differs from many traditional industries. Its marginal cost per interaction is low, but its total demand grows rapidly with adoption. Every additional user, every new application, adds to the load on already stretched infrastructure.

That is why researchers are less interested in individual usage and more focused on aggregate impact.

A growing but uneven burden

Studies are beginning to quantify what that growth could mean.

Research published in Environmental Research Letters estimates that expanding AI adoption across the United States could add about 896,000 tonnes of carbon dioxide emissions annually. That represents a small share of total emissions, around 0.02%, but it is not insignificant.

The same study suggests that energy demand linked to AI could rise by up to 12 petajoules each year, roughly equivalent to the electricity use of 300,000 households.

These are not system-breaking numbers. But they point to a trajectory. As AI becomes embedded across industries, its environmental cost will continue to rise unless efficiency improves.

Not all AI is equal

One of the more nuanced aspects of this debate is geography.

Where AI runs has a direct impact on how green it is. Data centres powered by renewable energy have a lower carbon footprint than those dependent on fossil fuels. Facilities in cooler climates require less water for cooling.

That means the same AI system can have very different environmental costs depending on where it is deployed.

It also means that decisions about infrastructure, location and energy sourcing are not just operational choices. They are climate decisions.

The transparency gap

Despite the scale of investment in AI, transparency around its environmental impact remains limited.

Companies report overall emissions and energy use, but rarely break down how much is tied specifically to AI workloads. That makes it difficult to compare systems, track progress or hold companies accountable.

De Vries-Gao and other researchers have called for more detailed disclosures, arguing that without them, it is impossible to build an accurate picture of AI’s footprint.

There is also a practical argument for transparency. Better data can help companies identify inefficiencies, reduce costs and improve performance.

At the moment, much of the debate relies on estimates rather than precise measurement.

The other side of the ledger

Focusing only on AI’s environmental cost tells an incomplete story.

The same technology is being used to address some of the planet’s most persistent challenges. In research published in Science, scientists at the University of Washington used AI to design an enzyme capable of breaking down certain plastics more efficiently.

The work involved generating and testing hundreds of potential protein structures before identifying one that could degrade materials such as PET. The implication is clear. AI can accelerate scientific discovery in ways that were previously not possible.

Applications extend beyond plastics. AI is being used to optimise energy systems, improve climate modelling and streamline supply chains. Each of these has the potential to reduce emissions elsewhere in the economy.

The question is not whether AI can help. It is whether its benefits can outpace its costs.

Efficiency becomes the next frontier

As awareness grows, so does the pressure to make AI more efficient.

Researchers and industry leaders are increasingly focused on reducing the energy intensity of models, improving hardware performance and designing data centres that use less water.

The study in Environmental Research Letters highlights this shift, noting the importance of integrating sustainability into AI development as adoption accelerates.

There are early signs of progress. Companies are exploring new cooling techniques, investing in renewable energy and optimising workloads to reduce waste.

But these efforts are uneven and, in many cases, still at an early stage.

The trade-off that defines the moment

AI sits in a space that is both promising and uncomfortable.

It offers tools that could help address climate change, yet it also adds to the demand for resources that drive that change. It improves efficiency in some areas while increasing consumption in others.

That tension is not unique to AI, but it is more visible here because of the speed and scale of its adoption.

On Earth Day, the conversation often centres on visible forms of environmental impact. AI’s footprint is less tangible, hidden in servers and supply chains.

That does not make it any less real.

The next phase of AI will not be defined only by what it can do, but by how it does it.

Regulators are likely to push for clearer reporting. Investors are paying closer attention to sustainability metrics. Companies are beginning to recognise that efficiency is not just an environmental issue, but a business one.

The central question remains unresolved. Can AI continue to scale without placing unsustainable pressure on energy and water systems?

The answer will depend on choices made now. Where data centres are built. How they are powered. How models are designed. And how transparent companies are about the cost of running them.

AI is not going away. Its environmental impact, however, is still up for negotiation.