How much energy does training a large AI model consume?

Training GPT-3 consumed an estimated 1,287 MWh. Larger models require substantially more. The exact figures depend on hardware, training duration, and data center efficiency.

Does the EU AI Act require environmental impact reporting?

Not directly, but Article 11 and Annex IV require documentation of computational resources for high-risk systems. The CSRD separately requires companies to report AI-related emissions under ESRS E1.

How can organisations reduce AI carbon emissions?

Select low-carbon grid regions, use model distillation and quantization, implement carbon-aware scheduling, evaluate smaller models, and track AI-specific energy consumption.

What is PUE and why does it matter for AI?

Power Usage Effectiveness (PUE) measures data center energy efficiency. The EU Taxonomy requires PUE of 1.3 or better for new facilities. AI workloads are among the most energy-intensive data center operations.

Does the EU Taxonomy apply to AI activities?

The EU Taxonomy applies to data center operations hosting AI workloads. AI companies reporting under CSRD must assess whether their compute infrastructure meets Taxonomy environmental criteria.

Quick answer

AI systems generate significant environmental impact through energy-intensive training and inference, water consumption for cooling, and hardware lifecycle waste, with reporting obligations emerging under the EU Corporate Sustainability Reporting Directive (CSRD) and the EU AI Act's energy efficiency transparency requirements.

Updated June 2026 · MmowW AI Compliance

Environmental Risk of AI: Carbon Footprint, E-Waste, and Sustainability

The Environmental Cost of AI

Training a single large language model can consume energy equivalent to hundreds of households' annual consumption and generate carbon emissions measured in hundreds of tonnes of CO2 equivalent. GPT-3 training was estimated at 1,287 MWh of electricity and 552 tonnes CO2e (Patterson et al., 2021). Subsequent models are orders of magnitude larger. Inference costs, while smaller per query, aggregate to significant environmental impact when models serve millions of daily requests.

Beyond carbon, AI infrastructure consumes substantial water for data center cooling. A 2023 study estimated that GPT-3 training consumed approximately 700,000 liters of fresh water for cooling alone (Li et al., 2023). As AI workloads grow, competition for water resources in data center locations becomes an environmental justice concern.

Environmental Impact Dimensions

Impact Category	AI Contribution	Measurement Challenge
Energy consumption	Training: thousands of MWh per large model. Inference: growing share of global data center energy	Provider-specific data often unavailable
Carbon emissions	Direct (Scope 1/2) and embodied (Scope 3) emissions from compute	Grid carbon intensity varies by location and time
Water consumption	Cooling water for data centers, especially in arid regions	Varies dramatically by cooling technology
E-waste	GPU and TPU hardware with 3-5 year replacement cycles	Limited recycling infrastructure for specialised AI chips
Mineral extraction	Rare earth elements in AI hardware manufacturing	Complex supply chains obscure origin

Regulatory Framework

The EU AI Act includes environmental considerations in Recital 27, noting that AI systems should be developed in an environmentally sustainable manner. Article 11 and Annex IV require technical documentation of high-risk AI systems to include information about the computational resources used, which enables environmental impact estimation. For GPAI models with systemic risk, Article 55(1)(a) requires model evaluations that may include energy efficiency assessments.

The Corporate Sustainability Reporting Directive (CSRD, Directive 2022/2464) requires large companies and listed SMEs to report on environmental impacts under the European Sustainability Reporting Standards (ESRS). ESRS E1 (Climate change) requires disclosure of Scope 1, 2, and 3 greenhouse gas emissions, which includes emissions from AI training and deployment. ESRS E5 (Resource use and circular economy) is relevant to AI hardware lifecycle impacts.

EU Taxonomy and Green AI

The EU Taxonomy Regulation (Regulation 2020/852) establishes criteria for environmentally sustainable economic activities. Data center operators seeking Taxonomy alignment must meet energy efficiency thresholds: a Power Usage Effectiveness (PUE) target of 1.3 for new facilities from 2025 and reuse of waste heat where feasible. AI workloads hosted in non-Taxonomy-aligned data centers may affect the sustainability reporting of downstream users.

Practical Mitigation Measures

Select cloud regions with low-carbon electricity grids for AI training workloads
Use model distillation and quantization to reduce inference compute requirements
Implement carbon-aware scheduling that shifts non-urgent training to periods of high renewable energy availability
Track and report AI-specific energy consumption separately from general IT workloads
Evaluate smaller, task-specific models before defaulting to large general-purpose models
Establish hardware lifecycle management policies that maximize GPU utilization and enable responsible recycling

Measuring AI Carbon Footprint

Tools for measuring AI environmental impact include ML CO2 Impact (Lacoste et al., 2019), CodeCarbon, and cloud provider sustainability dashboards. Organisations should measure both training-time and inference-time energy consumption, apply region-specific carbon intensity factors, and include embodied carbon from hardware manufacturing. Reporting should follow the GHG Protocol and align with CSRD/ESRS requirements.

The Rebound Effect

AI can improve energy efficiency in other sectors (smart grids, logistics optimization, building management), but these efficiency gains may be offset by increased AI usage. This Jevons paradox means that net environmental impact assessment must consider both the direct cost of AI compute and the indirect savings from AI-optimized processes. Organisations claiming environmental benefits from AI deployment should quantify both sides of the equation.

Check your AI compliance readiness — free.

Take the Readiness Check 3 minutes · 10 questions · no signup required

This article is for informational purposes only and does not constitute legal advice. Regulatory requirements change frequently — verify current rules with official sources. Built by Sawai Gyoseishoshi Office, Hiroshima, Japan.