The Environmental Cost of Training Large AI Models

Training GPT-4 is estimated to have consumed around 50 gigawatt-hours of electricity — enough to power approximately 5,000 UK households for a full year, used up in a single model training run. It generated an estimated 500 tonnes of CO₂ in direct emissions. Gemini Ultra, Claude 3 Opus, and Llama 3 405B each carried comparable footprints. The AI industry’s energy consumption has quietly become one of the fastest-growing contributors to tech sector emissions worldwide, and the numbers are getting harder to ignore.

The UK has committed to net-zero by 2050, with interim targets that require meaningful reductions in grid emissions through this decade. AI data centres are being built across England and Scotland at speed. Microsoft is expanding in London. Google has a £1 billion data centre under construction in Hertfordshire. These facilities draw power from the same grid the government is trying to clean up. The timing matters more than the industry currently admits.

This is not an argument against AI development. The technology is too useful and too embedded to make a credible case for stopping it. But understanding the true environmental cost — the energy, the water, the carbon — is the first step to holding companies accountable and making informed choices about which services and providers to support.

How Much Energy Training Actually Takes

Training a large language model requires running billions of calculations simultaneously across thousands of specialised GPU chips, continuously, for weeks or months. A single Nvidia A100 GPU draws between 250 and 400 watts of power. A cluster of 10,000 A100s running for three months consumes roughly 240,000 kilowatt-hours — before cooling overhead is added. Real-world data centre efficiency ratios typically add 30 to 50% to that figure.

Estimates for specific models are inherently imprecise because companies rarely disclose full training details. Researchers at the University of Massachusetts Amherst published influential work in 2019 showing that training a large transformer model produced roughly the same carbon emissions as five cars over their entire lifetimes. Models have grown substantially larger since then. GPT-2 had 1.5 billion parameters. GPT-4 is estimated to have around 1.8 trillion in a mixture-of-experts configuration. The scaling law cuts both ways — more capability, more cost.

Anthropic, OpenAI, and Google DeepMind have not published verified training energy figures for their frontier models. The estimates that circulate — 50 GWh for GPT-4, 20 to 30 GWh for Claude 2, similar ranges for Gemini — come from independent researchers reverse-engineering compute requirements from published architecture details. These are educated estimates, not confirmed numbers. That opacity is itself a problem, and one that deserves more attention from regulators.

The Water Problem Nobody Talks About

Energy consumption gets the headlines, but water usage may be more immediately alarming. Data centres require enormous cooling capacity to prevent GPU clusters from overheating. The dominant cooling method uses evaporative towers — water is pumped across hot surfaces and evaporated to dissipate heat. That water does not come back. It enters the atmosphere, not the water system.

Microsoft’s own environmental report for 2022 disclosed that the company’s global water consumption increased by 34% that year, reaching 6.4 million cubic metres. The company attributed a significant portion of this increase to AI research and development. A 2023 paper from researchers at University of California Riverside estimated that training GPT-3 required approximately 700,000 litres of fresh water — equivalent to cooling 370 MW of compute for the duration of the run. One conversation with ChatGPT uses roughly 500 millilitres when accounting for both training amortisation and inference cooling.

In the UK, water stress is a growing practical constraint. Parts of the South East of England are already classified as seriously water stressed by the Environment Agency. Planning applications for new data centre sites in these regions are beginning to face water availability scrutiny — a friction point that barely existed three years ago. The water footprint of AI is not an abstract concern; it is starting to appear in infrastructure planning decisions at the local authority level.

UK Data Centres and the National Grid

The UK is one of Europe’s largest data centre markets, with heavy concentration in London and surrounding counties. The Greater London Authority estimated in 2023 that data centres accounted for roughly 20% of London’s total electricity demand. AI’s rapid expansion pushed that figure higher through 2024 and 2025, though no authoritative updated estimate had been published as of early 2026.

National Grid ESO has identified data centre load growth as one of the primary demand-side challenges for grid planning through 2030. The UK’s electricity grid is becoming cleaner — renewable generation exceeded 50% of total output for the first time in 2023 — but peak demand periods still rely on gas peaker plants. When large data centres draw heavy load during peak hours, they can push gas plants to run that would otherwise stay offline. The marginal electricity consumed at peak is rarely the clean electricity the certificates suggest.

Several large AI-adjacent data centre projects have already faced planning delays in the UK due to grid connection backlogs. National Grid Electricity Transmission had a queue of over 700 GW of generation and storage projects awaiting connection approvals as of 2025. New large-load data centres face similar queues. Demand that cannot connect to the grid cleanly often ends up partially served by on-site diesel generators — an outcome substantially worse from an emissions standpoint than simply waiting for a clean connection.

What the Numbers Look Like in Practice

When I looked into concrete comparisons for UK context, several figures stood out. The International Energy Agency estimated in 2024 that global data centre electricity consumption was between 240 and 340 terawatt-hours annually — roughly the total electricity consumption of France in a year. AI accelerated that growth trajectory sharply, with the IEA projecting that data centre consumption could double by 2026 under central forecasts.

ChatGPT alone, serving around 100 million active users daily during 2024, was estimated to consume roughly 564 megawatt-hours of electricity per day — nearly ten times the daily consumption of Google Search at comparable traffic levels. That 564 MWh would power approximately 20,000 average UK homes for one day. Multiply that across Gemini, Claude, Copilot, and dozens of smaller models, and the aggregate consumption becomes extremely large at very high speed.

For UK businesses integrating AI tools into daily workflows, this energy consumption is partially allocated to their Scope 3 emissions — indirect emissions from purchased services. Under the UK’s Streamlined Energy and Carbon Reporting requirements, larger organisations must already account for material Scope 3 sources. AI service consumption may eventually become a disclosed line item for significant users. Procurement teams at larger firms should start thinking about this before it becomes mandatory, not after.

What AI Companies Actually Say

The major AI labs and their cloud infrastructure hosts have made ambitious renewable energy commitments. Microsoft pledged to be carbon negative by 2030 and to remove all historical carbon emissions by 2050. Google has promised to operate on 24/7 carbon-free energy by 2030. Amazon Web Services committed to 100% renewable energy by 2025 — a target it claims to have met through renewable energy certificate purchasing.

The gap between commitment and reality is real and measurable. Google’s own environmental reports show greenhouse gas emissions increased 48% between 2019 and 2023, partly attributed to data centre energy growth driven by AI infrastructure investment. Microsoft reported a 29% emissions increase over the same period. Renewable energy certificates — where a company pays for renewable generation elsewhere on the grid to offset fossil fuel consumption at their actual location — are a legitimate accounting tool. They do not mean the data centre itself runs on clean power at the moment of consumption.

24/7 carbon-free energy, matching renewable supply to demand hour by hour at the physical location of the facility, is a meaningfully higher bar. As of 2026, Google is furthest along on this metric, reaching around 64% on a 24/7 basis globally. Microsoft and Amazon lag considerably. For UK sites specifically, the South East — where the densest concentration of large data centres sits — has more constrained access to local renewable sources than Scotland or Wales. Location matters in ways the headline commitments tend to obscure.

The Inference Problem Is Growing Faster Than Training

Training a model happens once — or a small number of times during a development cycle. Inference — running the model to respond to user queries — happens billions of times every day. Until recently, inference was assumed to be the smaller energy cost relative to training. That assumption is changing fast as AI is integrated into search, productivity software, customer service, and creative tools across every sector.

Research from MIT CSAIL published in 2023 found that inference accounts for roughly 60 to 70% of total AI energy consumption across the industry — surpassing training when aggregated at scale. Every ChatGPT response, every Copilot code suggestion, every AI-generated image draws inference compute. The aggregate is growing at a rate that training efficiency gains cannot fully offset. Making models faster and cheaper to run attracts more users, which drives more total compute — a demand-rebound effect that undermines some efficiency improvements.

Smaller, more efficient models help. The growth of small language models — 7 billion parameters running on a laptop rather than one trillion parameters in a data centre — is a genuine step toward sustainable deployment at the edge. But these smaller models serve different use cases, and frontier model query volumes continue to grow. There is no single technical fix. The inference problem requires both efficiency improvements and honest demand-side accounting.

What This Means for You

For individual users, the practical impact is indirect but real. The AI services you use daily carry environmental costs that are largely invisible in the interface. That does not mean you should stop using them — the same is true of most digital services — but it is worth factoring in when choosing between providers. Companies that publish verified, location-specific renewable energy data are doing meaningfully more than those that rely primarily on certificate purchasing to meet their targets.

For UK businesses, the Scope 3 implications of AI tool adoption are worth tracking now rather than later. HMRC, OFGEM, and Companies House requirements around carbon reporting are tightening. If AI service consumption becomes a material emissions source — and for data-heavy operations it can — having an audit trail matters. Ask your AI providers for energy and emissions data, favour those who publish it, and treat environmental performance as a real factor in procurement decisions.

The broader issue is accountability. The AI industry is expanding faster than its environmental reporting frameworks. Independent researchers, investigative journalists, and regulators are working to close that transparency gap. The data they surface is already changing how some governments approach planning approvals and grid access for data centres. Pay attention to those developments — they are likely to affect AI service costs and availability in the UK within the next few years.

This article is for educational purposes only and does not constitute financial advice. Cryptocurrency investments involve significant risk. Always do your own research.