Blog
AI's environmental blind spot
The industry building the future runs on energy, water, and silence about both.
There is a question the AI industry prefers not to answer, and it is a simple one: what does all of this cost the planet?
Not in dollars. In kilowatt-hours. In litres of water. In carbon. The cost that doesn't appear on a balance sheet but shows up in strained power grids, drained aquifers, and a warming atmosphere.
We build AI products at Fabric. We use large language models every day. Our AI assistant helps people think, write, and organise their work. We are not writing this from the outside. We are writing it because we think the people building AI have a responsibility to talk about what it takes to run it.
What training a model costs
A single training run for a large language model can consume as much electricity as several hundred homes use in an entire year. That number is not a projection or a worst case. It is a rough consensus drawn from the limited data that AI labs have published, and from independent researchers who have tried to fill in the gaps.
The word "limited" is doing a lot of work in that sentence. Most AI companies do not publish detailed breakdowns of the energy consumed during training. Some publish headline carbon figures. Fewer publish methodology. Almost none publish water usage data. The result is an industry that talks constantly about capability benchmarks and almost never about environmental ones.
Training is only part of the picture. Once a model is trained, it has to be served. Every query, every conversation, every generated image runs on hardware in a data centre somewhere. Inference, the process of running a trained model to produce output, is less energy-intensive per operation than training, but it happens billions of times a day across the industry. At scale, inference energy use dwarfs training energy use.
The water nobody mentions
Data centres get hot. The processors inside them generate enormous amounts of heat, and that heat has to go somewhere. In many facilities, the answer is water. Evaporative cooling systems pull in fresh water, use it to absorb heat, and release it as vapour. The water is gone.
A single large data centre can use millions of litres of water per day. In arid regions, where many data centres are built because of cheap land and favourable tax conditions, this puts direct pressure on local water supplies. Some municipalities have raised concerns. A few have pushed back. Most have not.
The AI boom has accelerated data centre construction at a pace that energy grids were not designed to handle. In parts of the United States, Ireland, and the Netherlands, new data centre projects have been delayed or blocked because the local grid cannot supply enough power. Utilities are restarting retired gas plants and, in some cases, coal plants to meet demand. The irony is not subtle: an industry that positions itself as the future is, in some places, being powered by the energy infrastructure of the past.
The transparency gap
Several major technology companies made carbon-neutral or net-zero pledges in the years before the current AI boom. Those pledges were made when their primary workloads were search, email, and cloud storage. Since then, AI has changed the calculus. Training and serving large models requires significantly more energy per unit of compute than traditional workloads. Some companies that were on track to meet their climate targets have publicly acknowledged that AI has pushed those targets further away. Others have quietly stopped reporting against them.
This is not a small discrepancy. Microsoft reported a 29% increase in carbon emissions between 2020 and 2024, attributing much of the rise to data centre expansion for AI. Google reported a 48% increase in greenhouse gas emissions over a similar period, citing AI-related energy demand. These are companies with some of the most sophisticated sustainability programmes in the world. If they are struggling, smaller companies with fewer resources and less public scrutiny are unlikely to be doing better.
The problem is not that AI uses energy. Everything uses energy. The problem is that the industry has, on the whole, chosen not to talk about it with any specificity. Press releases announce new models with benchmark scores and capability lists. The energy and water consumed to produce those results almost never appears in the same announcement. The environmental cost is treated as someone else's problem, or as a problem that will solve itself through future efficiency gains.
Efficiency is not a solution by itself
It is true that hardware is getting more efficient. Each new generation of GPU and each new chip architecture does more work per watt than the last. But efficiency gains have historically been consumed by scale. When processors get more efficient, the response is to build larger models, run more queries, and serve more users. This is Jevons' paradox, and it has held remarkably steady across the history of computing.
Relying on efficiency gains to solve the environmental cost of AI is like relying on fuel-efficient engines to solve the environmental cost of driving while simultaneously building more roads and selling more cars. Efficiency is necessary, but it is not sufficient. It has to be paired with transparency, measurement, and deliberate action.
What we think the industry should do
We are not in a position to tell other companies how to run their operations. But we can say what we think responsible practice looks like.
First, publish the numbers. Energy consumed in training. Energy consumed in inference. Water used for cooling. Carbon emitted, with methodology. Not as a glossy sustainability report released once a year, but as a routine part of how AI companies talk about their work. If the industry can publish benchmark scores to four decimal places, it can publish energy consumption data.
Second, offset meaningfully. Carbon credits are not all created equal. Some represent verified, measurable, lasting carbon removal. Others are closer to accounting fictions. The difference matters, and companies should be transparent about which kind they are buying.
Third, invest in reduction, not just offsets. Use renewable energy where possible. Choose data centre locations with access to clean power. Design systems that minimise unnecessary computation. Offsets are a bridge, not a destination.
Fourth, account for water. Carbon gets most of the attention in environmental reporting, but water use is a growing concern, particularly in regions facing drought or water stress. AI companies should measure and report their water consumption alongside their carbon emissions.
What Fabric is doing
We are a small company compared to the organisations training frontier models. Our direct environmental footprint is correspondingly smaller. But "smaller" is not the same as "zero," and we do not think scale is an excuse for inaction.
Every Fabric subscription plants mangrove trees on the coasts of Kenya and Brazil. This is not a vague promise. The trees are geotagged, photographed, and independently audited. Our planting sites are monitored for three years. We work with GoodAPI, which connects us to verified planting projects, and Veritree, which handles on-the-ground verification in Kenya. You can see the details on our environment page.
We chose mangroves for specific reasons. They store carbon at rates significantly higher than most terrestrial forests. They protect coastlines from erosion and storm damage. They support local livelihoods through fishing and sustainable harvesting. And they support biodiversity in some of the most ecologically important coastal ecosystems on Earth. Planting mangroves is not a symbolic gesture. It is a measurable, verifiable contribution to carbon sequestration and ecosystem health.
We are working towards making Fabric 100% carbon and water neutral by 2030. That is a goal, not a claim. We have not reached it yet. But we have a plan, we have partners, and we are measuring our progress. The details of that plan, including our current status, are published on our environment page.
Why this matters for the tools we build
There is a connection between environmental responsibility and the kind of product we want to build. Fabric is designed to help people organise their thinking, find what they need, and work together without unnecessary friction. We want to build tools that are useful without being wasteful. That principle applies to how our product works and to how our company operates.
We do not think caring about the environment and building AI products are in tension. We think they are complementary. The companies that take their environmental impact seriously are, in our experience, the same companies that think carefully about what they build and why. Rigour in one area tends to produce rigour in others.
The AI industry is growing fast. The decisions being made now, about infrastructure, energy sources, transparency, and accountability, will shape the environmental impact of this technology for decades. We think those decisions deserve more attention than they are getting.
If you are interested in how we approach this at Fabric, our environment page has the full picture: our partners, our planting projects, our targets, and our progress. We have also published comparisons of how different categories of software stack up on sustainability, covering AI apps, cloud storage, note-taking tools, productivity apps, and eco-friendly AI tools. Whether you are a founder building a company, part of a startup team, or someone who wants their tools to reflect their values, we think these are worth reading.
We build AI because we believe it can help people do better work. We also believe that building it responsibly is not optional. It is the minimum.
Frequently asked questions
How much energy does training a large AI model use?
A single training run for a large language model can use as much electricity as several hundred homes consume in a year. The exact figure varies depending on the model size, the hardware used, and the duration of training, but the energy demands are substantial by any measure.
Why do data centres use so much water?
Data centres house thousands of processors that generate significant heat. Many facilities use evaporative cooling systems, which consume fresh water to absorb and dissipate that heat. A single large data centre can use millions of litres of water daily, which raises concerns in water-stressed regions.
Are AI companies transparent about their environmental impact?
Most AI companies publish limited environmental data. Some provide headline carbon figures in annual sustainability reports, but detailed breakdowns of energy use, water consumption, and methodology are rare. The industry lags behind other sectors in environmental disclosure.
What is Fabric doing to reduce its environmental footprint?
Every Fabric subscription plants mangrove trees on the coasts of Kenya and Brazil through verified partners. The trees are geotagged, photographed, and independently audited, with sites monitored for three years. Fabric is working towards 100% carbon and water neutrality by 2030. Full details are on our environment page.
Why does Fabric plant mangrove trees specifically?
Mangroves store carbon at rates significantly higher than most terrestrial forests. They also protect coastlines from erosion, support local livelihoods, and sustain biodiversity in critical coastal ecosystems. They offer environmental, social, and ecological benefits that make them particularly effective for restoration projects.
Who verifies Fabric's tree planting?
Fabric works with GoodAPI, which connects to verified planting projects, and Veritree, which provides on-the-ground verification in Kenya. Trees are geotagged and photographed, and planting sites undergo independent audits and three years of monitoring.
Will AI's energy consumption keep growing?
Current trends suggest yes. While hardware efficiency improves with each generation, those gains are typically consumed by larger models, more users, and more queries. Without deliberate action on transparency, renewable energy, and accountability, the environmental cost of AI is likely to increase.
What can individuals do about AI's environmental impact?
Choosing products and services from companies that are transparent about their environmental practices is one step. Supporting organisations that measure, report, and offset their impact helps create incentives for the broader industry. You can learn more about how Fabric approaches this on our environment page.
Is carbon neutrality the same as zero emissions?
Carbon neutrality means that a company offsets its emissions through verified carbon removal or reduction projects, bringing its net emissions to zero. It does not mean the company produces no emissions at all. The quality and verifiability of offsets matter significantly, which is why Fabric works with audited, geotagged planting projects.
Does using Fabric's AI features contribute to environmental impact?
Any use of AI involves computation, which requires energy. Fabric's AI assistant and agents run on cloud infrastructure with an associated environmental footprint. Our goal is to offset that impact fully through our tree planting programme and our path to carbon and water neutrality by 2030.
