The Environmental Cost of Training Large AI Models
AI training and inference now draw serious electricity and water — here is what it costs and what UK data centres mean for your bills.
Every time you type a prompt into ChatGPT or Claude, something happens you never see. A server rack somewhere spins up, draws power from the grid, and gets hot enough that it needs constant cooling. Multiply that by billions of queries a day, and the environmental bill for AI is no longer a footnote. UK households are starting to feel it too, through higher electricity demand near new data centres and questions Ofgem is only just beginning to answer.
How Much Energy Does Training a Large Model Actually Use
Training a frontier AI model is not a one-off cost. It is months of thousands of GPUs running flat out. Researchers at the University of Massachusetts Amherst estimated that training a single large language model can emit over 280,000 kilograms of CO2 — roughly the same as five cars driven for their entire lifetimes.
That figure has only grown. Newer frontier models use ten times more compute than models from 2022. When I looked into this for DigiTech readers, the pattern was clear: every leap in AI capability comes with a matching leap in power draw.
Training is just the start. Once a model launches, it answers millions of queries daily. That ongoing “inference” cost, spread across a model’s lifetime, often dwarfs the original training bill.
Data Centres and the UK’s Power Grid
The UK has become a magnet for AI infrastructure. Data centre electricity demand in Britain is forecast to roughly triple by 2030, according to National Grid ESO planning documents. Some new facilities near London and Slough already draw as much power as a small city.
This matters because grid capacity is not infinite. Housing developers in west London have reported delays connecting new homes to the grid, partly because data centres have already reserved available capacity. Nobody expected AI to compete with new housing for electricity. It does now.
Ofgem classified data centres as critical national infrastructure in 2024, which fast-tracks their grid connections. Critics say this puts AI ahead of hospitals and schools in the connection queue. That is a genuinely uncomfortable trade-off.
Water Use — The Hidden Cost Nobody Talks About
Cooling is the part people forget. Data centres do not just need electricity, they need water — a lot of it — to stop servers overheating. Google reported its data centres consumed over 6 billion gallons of water in 2023 alone, up sharply year on year.
Microsoft’s water use jumped 34% in the year it deployed ChatGPT-linked infrastructure at scale. A single conversation with a chatbot, spread across a session of maybe 20-50 prompts, can use half a litre of water for cooling. That’s not nothing when you add it up across a billion users.
UK data centres mostly use closed-loop or air cooling systems, which need less water than the US facilities in hot, dry states. Still, several new sites planned for the south east sit in areas already flagged as water-stressed by the Environment Agency.
Carbon Footprint Per Query — What a Single Prompt Costs
Numbers here vary wildly depending on the model and hardware, but a rough estimate from MIT Technology Review puts a single ChatGPT query at around 3 watt-hours of electricity — roughly ten times a Google search.
Image generation costs more. Producing one AI image can use as much energy as fully charging a smartphone. Video generation, the newest frontier, costs more again.
Scale that to daily usage. OpenAI has said ChatGPT handles over a billion messages a day. Even at conservative per-query estimates, that adds up to a genuinely large chunk of national grid demand somewhere in the world, every single day.
Are AI Companies Doing Anything About It
Some are trying. Anthropic has committed to matching its energy use with renewable purchases and publishes environmental reporting alongside its model releases. Google and Microsoft both claim carbon-neutral operations through offsets and renewable power purchase agreements, though critics argue offsets paper over the real problem rather than solving it.
Microsoft signed a deal to restart part of the Three Mile Island nuclear plant specifically to power AI data centres. That’s how seriously the biggest players are taking the power question — restarting nuclear reactors, not just buying wind credits.
Newer, smaller models are part of the answer too. Chip efficiency gains and smaller specialised models can cut inference costs by 90% compared with running everything through a giant general-purpose model. Falls apart fast if demand keeps outpacing efficiency gains, though — and so far, it has.
The UK’s Own AI Data Centre Boom
The government’s AI Growth Zones initiative, announced in 2025, is fast-tracking planning permission for new data centres across the UK. Sites in Wales, the north east and Scotland have been earmarked, partly to spread the grid burden away from the overloaded south east.
Investment figures are eye-watering. Combined pledges from US tech firms and UK partners for AI infrastructure now exceed £40 billion. That builds jobs and tax revenue. It also builds permanent new electricity demand that has to come from somewhere.
UK investors keep asking about this because it cuts both ways — energy and infrastructure companies stand to benefit from AI’s power appetite, even as consumer electricity bills face upward pressure from the same source.
Can AI Actually Help Fight Climate Change Too
Here’s the twist. AI is also being used to cut emissions elsewhere. DeepMind’s work on data centre cooling optimisation cut Google’s own cooling energy use by 40%. AI models are helping grid operators balance renewable supply more efficiently, reducing waste from wind and solar.
Climate scientists use AI to model weather patterns and predict extreme events with more accuracy, which helps infrastructure planning. It’s a genuinely mixed picture — the technology causing part of the problem is also solving pieces of it elsewhere.
Whether the net effect is positive or negative depends entirely on how fast AI’s own footprint grows relative to the savings it enables. Right now, nobody has a definitive answer.
How AI Stacks Up Against Other Energy-Hungry Industries
Perspective matters here. The International Energy Agency estimates data centres — AI included — account for around 1.5% of global electricity use today. Aviation, by comparison, accounts for roughly 2-3% of global CO2 emissions on its own.
What makes AI different is the growth curve. Aviation’s emissions grow slowly, tracked against passenger numbers that rise a few percent a year. Data centre electricity demand tied to AI is forecast to double or even triple within five years, according to the IEA’s 2025 Electricity report. Nothing else in the economy is scaling that fast.
Bitcoin mining faced similar criticism a few years back, and that comparison isn’t accidental. Both industries run enormous banks of specialised hardware around the clock. The difference is that Bitcoin’s energy use is roughly fixed by its mining difficulty, while AI’s appetite grows with every new model generation.
Cement production and steel manufacturing still dwarf AI’s current footprint in absolute terms. Ugly comparison, but a useful one — AI isn’t yet the biggest industrial emitter. It’s the fastest-growing one, and that’s the number that worries grid planners.
What Efficiency Gains Are Actually Happening in Hardware
Chip design is where most of the real progress sits. Nvidia’s newest Blackwell GPUs deliver roughly 25 times more energy efficiency per AI query than the Hopper generation that trained GPT-4, according to Nvidia’s own published benchmarks.
That’s not marketing spin dressed up as green credentials — it’s a genuine architectural leap, driven by smaller transistor nodes and smarter memory handling. Google’s custom TPU chips have followed a similar trajectory, cutting inference energy costs generation over generation.
Model design has shifted too. Techniques like quantisation — running models at lower numerical precision — and distillation, where a smaller model learns to mimic a larger one, can cut energy use for a given task by 70% or more without a meaningful quality drop.
The catch, and there’s always a catch, is Jevons paradox. Every time AI gets cheaper to run, more people run more of it. Efficiency gains keep getting eaten by demand growth, which is exactly why total energy use keeps climbing even as per-query costs fall.
Small Steps That Actually Reduce Your AI Footprint
You can’t control how a data centre is cooled, but you do have some influence over which tools you reach for. Smaller, task-specific models genuinely use less energy per query than giant general-purpose ones — using a lightweight model for a simple summarisation task, rather than the biggest frontier model available, is a real reduction, not just a feel-good gesture.
Batching requests helps too. Running one longer, well-structured prompt instead of five back-and-forth follow-ups reduces the total compute spent reaching the same answer. It’s a small thing. It adds up at billion-user scale.
None of this replaces the need for policy-level change — grid investment, renewable procurement, transparent reporting. But it’s a reasonable place to start if the footprint genuinely bothers you.
What This Means for You
If you use AI tools daily, your personal footprint from that use is small compared with things like flying or driving. But the aggregate picture matters for UK energy policy, electricity prices, and where new housing and infrastructure can be built.
Watch for two things over the next year: Ofgem’s grid connection reforms, and whether AI companies start publishing per-query energy figures the way food products carry calorie counts. Transparency here is still miles behind where it needs to be.
For UK investors, energy infrastructure, grid technology and nuclear power all look like quiet beneficiaries of the AI boom — regardless of which chatbot wins the popularity contest.
This article is for educational purposes only and does not constitute financial advice. Cryptocurrency investments involve significant risk. Always do your own research.
Stay ahead of the market
Join our community of nearly 5,000 across YouTube, LinkedIn, X, and Facebook — weekly crypto, AI, and digital lifestyle insights every Thursday. No spam. Unsubscribe any time.
Partner picks
Build a smarter digital stack
Explore curated AI, automation, wealth, and creator tools selected for practical value, transparent pricing, and clear use cases.
Disclosure: some links may be affiliate links. DigitechLifestyle may earn a commission at no additional cost to you.
