Blog
The Environmental Cost of AI
The rapid growth of artificial intelligence is raising serious questions about energy, water, and carbon, and the answers so far are not reassuring.
Artificial intelligence is now embedded in search engines, email clients, workplace tools, creative software, and personal assistants. The pace of adoption has been remarkable. But every query, every generated image, every summarised document runs on physical hardware in a physical building, drawing real electricity and real water. The environmental cost of that infrastructure is climbing, and it deserves more scrutiny than it tends to receive.
This is not an argument against AI. It is an attempt to lay out what we know, what the trends suggest, and what levers exist to push the industry in a better direction.
How much energy does AI use?
Training a large language model is computationally expensive. Estimates vary depending on the model, the hardware, and how efficiency is measured, but researchers have placed the energy required to train a single frontier model in the range of tens of gigawatt-hours. To put that in more human terms, a single training run can consume as much electricity as several hundred average homes use in an entire year.
Inference, the process of running a trained model to answer questions or generate output, uses less energy per operation. But inference happens billions of times a day across millions of users, and the cumulative demand is enormous. Some analysts estimate that a single AI-powered search query uses roughly ten times the energy of a conventional web search. As AI features are added to more products, the aggregate inference load is growing faster than the training load.
The International Energy Agency projected that data centre electricity consumption could more than double between 2022 and 2026, with AI workloads accounting for a significant share of that growth. In some regions, new data centres are already straining the electrical grid. Ireland, where data centres now account for roughly a fifth of national electricity consumption, has had to restrict new connections. Parts of the Netherlands and several US states are facing similar pressures.
The water question
Data centres generate heat, and that heat needs to be managed. Many facilities use evaporative cooling, which consumes large quantities of water. A single large data centre can use several million litres of water per day, depending on its cooling design and the local climate.
Research published in 2023 estimated that training GPT-3 alone consumed around 700,000 litres of fresh water for cooling. More recent, larger models are likely to have consumed more, though exact figures are often not disclosed. Microsoft's 2024 environmental report showed a 34% increase in water consumption between 2021 and 2023, a period that coincided with its aggressive expansion of AI infrastructure. Google reported a similar trajectory.
This is not a trivial concern. Many data centres are located in regions already experiencing water stress. The competition between industrial cooling and local water needs is a real tension, one that will intensify as more facilities are built.
Carbon emissions are going up, not down
Both Microsoft and Google made prominent climate commitments in the early 2020s. Microsoft pledged to be carbon negative by 2030. Google aimed to run on 24/7 carbon-free energy by the same year. Both companies have since reported that their carbon emissions are moving in the wrong direction.
Microsoft's 2024 sustainability report showed a 29% increase in overall carbon emissions compared to 2020, with data centre construction and AI workloads cited as primary drivers. Google reported a 48% increase in emissions over a comparable period. In both cases, the companies acknowledged that AI-related infrastructure was a significant factor.
Some companies have quietly revised their climate timelines or softened their language around targets. The tension is structural: the commercial incentive to build and deploy AI as fast as possible is at odds with the time and investment required to decarbonise the energy supply for that infrastructure. Renewable energy procurement is growing, but not fast enough to offset the pace of new demand.
Jevons' paradox and the efficiency trap
There is a common assumption that as AI hardware becomes more efficient, the environmental impact will naturally decrease. This is worth examining carefully, because the historical pattern suggests otherwise.
Jevons' paradox, first observed in the context of coal consumption in the 19th century, describes the phenomenon where improvements in efficiency lead to increased total consumption rather than decreased consumption. More efficient engines made coal cheaper to use, which led to more uses for coal, which led to more coal being burned overall.
The same dynamic is visible in computing. Chips are more energy-efficient than ever. Models are becoming more efficient to train and run. But the response has been to train more models, run them more often, embed them in more products, and serve more users. The net effect is higher total energy consumption, not lower. Efficiency gains are necessary but not sufficient. Without constraints or incentives that operate at the level of total consumption, efficiency improvements may simply accelerate growth.
What regulations exist?
The regulatory landscape is still sparse, but some frameworks are emerging. The EU AI Act, which began entering into force in 2024, includes provisions that require high-risk AI systems to report on energy consumption and environmental impact. These provisions are modest in scope, and enforcement mechanisms are still being developed, but they represent a starting point.
In the United States, regulation has been slower. Some states have introduced data centre energy reporting requirements, and there is growing legislative interest in the environmental footprint of large-scale computing, but no comprehensive federal framework exists. China has introduced some data centre efficiency standards, though transparency around compliance is limited.
At the international level, the topic has appeared in G7 and G20 discussions, but binding commitments remain rare. The gap between the speed of AI deployment and the speed of environmental governance is wide.
What major companies are doing
To their credit, most large technology companies are investing in renewable energy. Microsoft, Google, Amazon, and Meta have all made substantial commitments to renewable energy procurement and have signed power purchase agreements for wind and solar projects. Some are exploring nuclear energy, including small modular reactors, as a potential source of baseload power for data centres.
These efforts are real, but they are also incomplete. Renewable energy certificates, which allow a company to claim renewable energy usage even when the electrons powering their facilities come from the grid, can overstate the actual environmental benefit. The distinction between "matched" renewables (certificates purchased elsewhere) and directly co-located clean energy is important.
Water stewardship programmes exist at several major companies, including commitments to replenish more water than they consume. The methodologies for measuring water replenishment vary, and independent verification is not always available. There is progress, but it is hard to assess precisely how much.
What users and buyers can push for
Individual users have limited direct leverage over the energy mix of a data centre. But collective choices, purchasing decisions, and vocal demand for transparency can influence corporate behaviour. A few areas where pressure can be productive:
Transparency is the foundation. Users and enterprise buyers can ask providers to disclose energy consumption per service, water usage by facility, and carbon emissions with clear methodology. Vague commitments to "sustainability" are less useful than auditable numbers.
Procurement decisions matter. When organisations choose cloud providers, AI tools, or workspace platforms, environmental track record can be a factor in the evaluation. Some companies are beginning to include sustainability criteria in their vendor selection processes.
Supporting standards and regulation is also productive. Industry standards for AI energy reporting, if widely adopted, would make it easier to compare providers and hold them accountable. Users who care about this issue can advocate for stronger reporting requirements.
Choosing tools with care is worth considering. Not every task needs a frontier model. Smaller, more efficient models can handle many use cases with a fraction of the energy cost. AI assistants that work within a defined scope, drawing on your own files rather than the entire internet, can be both more useful and less resource-intensive.
What Fabric is doing
We should be transparent about our own position. Fabric is a small company compared to the hyperscale cloud providers, and our environmental footprint is proportionally smaller. But we believe that every company in the AI space has a responsibility to take this seriously.
Every Fabric subscription plants mangrove trees through restoration projects in Kenya and Brazil. Mangroves are among the most effective natural carbon sinks, sequestering carbon at rates several times higher than terrestrial forests. We are working towards becoming 100% carbon and water neutral by 2030.
We do not claim this solves the problem. Tree planting is not a substitute for reducing emissions at the source. But we think it is better than doing nothing, and we think building environmental costs into the business model, rather than treating them as an afterthought, is the right approach.
Our broader philosophy is that your tools should work efficiently on your behalf. Fabric's AI works across your files without sending them elsewhere for training, and its cloud streaming approach means files are delivered on demand rather than duplicated across devices, which reduces storage infrastructure overhead.
The bigger picture
The environmental cost of AI is not a reason to stop building or using AI. The technology has real benefits, and some of its applications, from climate modelling to energy grid optimisation, may help address environmental challenges directly.
But the current trajectory is concerning. Emissions are rising. Water consumption is increasing. Power grids are under strain. The gap between corporate climate pledges and corporate climate performance is growing. And the market incentives currently favour speed and scale over environmental responsibility.
The path forward involves several things happening at once: genuine investment in clean energy (not just certificates), transparent reporting of environmental metrics, regulatory frameworks that keep pace with deployment, business models that internalise environmental costs, and informed users who factor sustainability into their choices.
This is a shared problem. It belongs to the companies building AI, the governments regulating it, and the people using it. The least we can do is be honest about the costs.
You can read more about how Fabric approaches sustainability and how we compare in the most sustainable AI apps and most sustainable cloud storage categories.
Frequently asked questions
How much electricity does training a large AI model use?
Training a frontier language model can consume tens of gigawatt-hours of electricity, roughly equivalent to the annual consumption of several hundred average homes. The exact figure depends on the model size, hardware efficiency, and training duration.
Why do data centres use so much water?
Many data centres use evaporative cooling systems to manage the heat generated by thousands of servers. This process consumes large quantities of water. A single large facility can use several million litres per day, depending on the cooling technology and local climate conditions.
Have tech companies met their climate targets?
Several major companies have reported emissions increases that move them further from their stated targets. Microsoft reported a 29% carbon emission increase between 2020 and 2024, and Google reported a 48% increase over a comparable period, with both citing AI infrastructure as a contributing factor.
What is Jevons' paradox and how does it apply to AI?
Jevons' paradox describes the observation that efficiency improvements can lead to higher total resource consumption rather than lower. In AI, more efficient chips and models have enabled more widespread deployment, resulting in higher aggregate energy use despite per-operation efficiency gains.
Does the EU regulate AI's environmental impact?
The EU AI Act includes provisions requiring certain AI systems to report on energy consumption and environmental impact. These requirements are still being implemented and are limited in scope, but they represent one of the first regulatory attempts to address AI's environmental footprint.
What can individual users do about AI's environmental cost?
Users can choose providers that are transparent about their environmental metrics, favour tools that use appropriately sized models for the task, support regulatory efforts around energy and emissions reporting, and include sustainability criteria when evaluating cloud storage and AI tools.
Is renewable energy enough to offset AI's energy demands?
Renewable energy procurement is growing, but it is not keeping pace with the growth in data centre energy demand. The distinction between renewable energy certificates and directly co-located clean energy matters. Renewables are necessary but not sufficient on their own without addressing total demand growth.
Does Fabric offset its environmental impact?
Fabric plants mangrove trees with every subscription through projects in Kenya and Brazil, and is working towards 100% carbon and water neutrality by 2030. You can read more about Fabric's environmental commitments on the environment page.
How does cloud streaming reduce environmental impact?
Cloud streaming delivers files on demand rather than syncing full copies to every device. This reduces the total storage infrastructure required, which in turn reduces the energy and cooling demands associated with maintaining duplicate data across multiple locations.
