What’s the Environmental Cost of AI?

Photo: OCED.AI , Everyone must understand the environmental costs of AI

Introduction

AI is everywhere now. From a kid to a grandma everyone is using it. From writing emails to telling us what all can be cooked with the vegetables left in the kitchen, to recommending movies to being part of an image generating trend in the Gen Z social media, AI is omnipresent. But behind every query, every generated image, every chatbot response is an enormous, energy-hungry machine that almost nobody talks about.

The true cost of AI isn’t measured only in dollars or computing hours. It's measured in electricity, carbon and water. As the AI industry expand at breakneck speed and places unprecedented pressure on electricity and water, two of the most vital resources. The solutions being offered most notably nuclear energy raise serious questions of their own.

So, the question this blog asks is simple – is the rapid growth of artificial intelligence creating a new environmental crisis with the alternatives proposed?

The Environmental Impact: Power and Water

The computational power required to train generative AI models- which often has billions of model parameters- demands a huge amount of electricity leading to increased carbon emissions and significant pressure on the electric grid. But the impact doesn’t stop once the training of these models end. Deploying these models in the real world and then fine tuning them throughout to enhance the performance draws large amounts of energy long after a model has been developed.

According to the International Energy Agency’s Energy and AI special report, the global data centres’ electricity consumptions have reached 415 terawatt hours. To put this in easy terms: - a typical AI- focused data centre consumes as much electricity energy as 100,000 average households and the largest ones today, some of which are still under the construction, will consume 20 times as much energy.

This is not just limited to energy. Water is another hidden cost. A great deal of water is needed to cool the hardware used to train, deploy and fine-tune the generative AI models, which can strain municipal water supplies and disrupt local ecosystems as well.

The Cost is Growing Rapidly

This is not a static problem- it’s accelerating rapidly. The IEA projects that AI data centres’ electricity consumption is set to double to around 945-terawatt hours by 2030 showing how rapidly the demand is increasing. Industry analysts estimate that a single hyper scale AI data centre- large facilities designed to handle massive data and computing tasks- can demand 300 to 500 megawatts of electricity which is comparable to the consumption of a mid-sized city. Multiply that across dozens of facilities under construction, energy supply becomes less of an operating expense and more of a key limiting factor for the growth of AI infrastructure.

Global investment in data centres has nearly doubled since 2022 and reached half a trillion dollars in 2024. The pace of growth is extraordinary and so is the environmental pressure that is growing with it.

Photo: China US Focus, Environmental AI Governance: U.S. and China Have Different Roads to developing Green AI Systems

The Possible Solution: - Nuclear Energy

This challenge is recognized by the world’s biggest tech companies, and they have chosen to arrive at what they believe is the answer- nuclear power. Instead, of relying solely on renewable energy both Microsoft and Amazon are now securing direct relationship via agreements with nuclear power generations Through this, they will begin to operate like long term energy planners rather than pure technology companies.

The reason for this shift is clear. Modern AI systems run continuously and require a round the clock power. Natural resources like solar or the wind energy remain essential parts of the energy generation, but they cannot satisfy or guarantee the steady output required by massive computing clusters without additional stable electricity generation resource.

The move to nuclear energy is concrete which can clearly be gaged by Microsoft’s involvement in restarting the former Three Mile Island Unit 1 reactor now known as the Crane Clean Energy Center. This is significant because it reflects a direct attempt to secure a stable electricity generation resource for AI data centers.  Meanwhile, Amazon’s strategy emphasizes direct control over energy supply. Its acquisition of the Cumulus Data Center campus from Talen Energy provides direct access to electricity generated by the Susquehanna nuclear facility. This “behind the meter” model allows Amazon to directly use power from the source avoiding transmission bottlenecks and delays.

The IEA itself acknowledges that nuclear energy has a role to play alongside renewable energy, natural gas and advanced geothermal energy in meeting the energy requirements of data centres through 2035.

The Problem with the Solution

Nuclear energy is not however, an escape. Replacing one set of environmental concerns with another is a trade-off.

The most obvious issue here is the radioactive waste. Nuclear reactors generate spent fuel that remains hazardous for thousands of years and no country in the world has a fully operational permanent repository for high-level nuclear waste. Therefore, the problem at hand is no longer theoretical. Then, addressing the question of water: - nuclear power plants are among the most water-intensive forms of electricity generation. It relies on large volumes of water for cooling. The very resource that AI’s data centres are already depleting would be further strained by a large-scale nuclear buildout.

In addition to these two points, there comes the issue of time taken in building the nuclear power plant; which takes years. Therefore, restarting an existing reactor is a faster approach to reliable carbon-free generation of electricity which is why Microsoft has chosen to restart the Three Mile Island reactor rather than commissioning a new one. But even restarting a new plant takes years to come online and become functional. The IEA notes that the first small modular reactors are not expected to be operational until around 2030, meanwhile the AI data centres are being commissioned in months.

Renewable source of energy faces their own limitations too. The IEA projects that half of all data centres’ growing electricity demand by 2035 will be met by renewables but as the industry acknowledges- solar and wind energies alone cannot alone provide the 24/7 uninterrupted power that AI infrastructure demands.

Conclusion

AI is here to stay, so the question now is- how do we build it while keeping the environmental cost in mind. Switching to nuclear power is not a solution but a trade-off of carbon emissions with radioactive waste which also would double the pressure on the water resource as well and renewable energy sources are not sufficient to match the demands. AI required genuine sustainability more than a cleaner source. It requires the industry to be honest with itself regarding its pace of growth and not ask how AI is powered but whether every model is even necessary? Responsible scaling, transparency and serious investment in efficiency must become central to AI development as the technology itself. This also means prioritising fewer but more efficient systems reducing unnecessary computational demands and designing infrastructure that minimises resource use. The algorithm might be getting smarter but so is the requirement of the industry development.

[This post has been authored by Mrinalini Yadav, a third-year law student at JGLS]