This article quantifies the energy and carbon footprint of LLM training in 2025, based on the latest public industry and research estimates. Training a frontier-scale large language model can consume several gigawatt-hours (GWh) of electricity and lock in thousands of tonnes of CO2-equivalent if powered by fossil-heavy grids. Yet, when deployed carefully, the same models can help optimise energy systems, industrial processes, and building operations. At Energy Solutions, we model AI training pipelines alongside renewables, storage, and electrification investments to understand when AI's climate impact is a problem—and when it becomes part of the solution.
Key benchmarks (public estimates):
What You'll Learn
- AI Training Basics: Where the Energy Goes
- Energy Use & Emissions Benchmarks in 2025
- Regional Grid Mix and Data Center Location
- Case Studies: Cloud LLM, On-Prem Cluster, and Greenfield Build
- Global Perspective: US vs EU vs Asia Data Center Strategies
- Mitigation Levers: Efficiency, Scheduling, and Procurement
- Devil's Advocate: Why Green AI Is Hard in Practice
- Outlook to 2030: Scenarios for AI's Energy Footprint
- FAQ: Counting, Reporting, and Reducing AI Emissions
AI Training Basics: Where the Energy Goes
Training large language models involves:
- Compute: GPUs/TPUs or custom accelerators consuming megawatts over weeks or months.
- Cooling & overhead: chillers, pumps, fans, and power distribution losses captured by the PUE (Power Usage Effectiveness) metric.
- Embodied emissions: manufacturing servers and networking gear, often amortised over multiple training runs.
Energy Use & Emissions Benchmarks in 2025
Public reporting on LLM training energy varies by hardware, run duration, and accounting boundaries (compute-only vs full facility energy). One widely cited estimate indicates that training OpenAI's GPT-3 consumed approximately 1,287 MWh of electricity and generated roughly 552 metric tons of CO2e for the training phase (source). That scale is enough to power about 120 average US homes for a year (order-of-magnitude comparison).
Inference is a separate (and often larger) operational footprint at scale. Recent measurements for Google's Gemini serving stack report around 0.24 Wh (0.00024 kWh) per typical query, including infrastructure overhead (source). Data center efficiency matters: a PUE of 1.12 implies about 12% additional energy overhead beyond IT equipment (source).
Indicative Training Energy Use for Different Model Scales (2025–2026)
| Model Class (Illustrative) | Parameter Count | Training Compute (FLOPs) | Training Energy | CO2e on 400 g/kWh Grid |
|---|---|---|---|---|
| Foundation LLM — Tier 1 | 100–200B | 1–3 × 1023 | 2–6 GWh | 800–2,400 tCO2e |
| Foundation LLM — Tier 2 | 20–60B | 1–5 × 1022 | 0.4–1.5 GWh | 160–600 tCO2e |
| Domain / Enterprise Model | 5–20B | 1–5 × 1021 | 0.05–0.3 GWh | 20–120 tCO2e |
| Fine-Tuning / Instruction Tuning | Base model reused | 1019–1020 | 1–20 MWh | 0.4–8 tCO2e |
Estimates assume modern accelerators at 30–40% utilisation and PUE ≈ 1.2–1.4. Actual projects vary widely.
Approximate Energy Use by Model Class
Regional Grid Mix and Data Center Location
The same training run can have radically different climate impact depending on grid carbon intensity.
Illustrative LLM Training Emissions by Region (Single 3 GWh Run)
| Region / Grid Profile | Grid Intensity (kg CO2/kWh) | Emissions for 3 GWh Run | Notes |
|---|---|---|---|
| Hydro / Nuclear Heavy | 50–80 | 150–240 tCO2e | Nordics, Québec, parts of France. |
| Average OECD Grid | 300–400 | 900–1,200 tCO2e | Mix of gas, coal, renewables. |
| Coal Heavy | 650–900 | 1,950–2,700 tCO2e | Regions with high unabated coal share. |
Same 3 GWh Training Run on Different Grids
Case Studies: Cloud LLM, On-Prem Cluster, and Greenfield Build
Case Study 1 — Cloud LLM on Mixed Grid
- Setup: Frontier-scale model trained on a large public cloud region with average PUE 1.3.
- Energy: 5 GWh over multi-week run.
- Emissions: ≈2,000 tCO2e on 400 g/kWh grid.
- Mitigation: 100% renewable energy certificate (REC) retirements reduce the market-based footprint, but do not change physical emissions in the short term.
Case Study 2 — On-Prem Cluster with Aggressive Efficiency
- Setup: Enterprise data center retrofitted with liquid cooling, PUE ≈ 1.15.
- Energy: 1.2 GWh to train a 20B-parameter model.
- Emissions: ≈360 tCO2e on a 300 g/kWh grid, reduced to a lower market-based footprint via a regional PPA contract.
Case Study 3 — Greenfield Renewable-Powered Campus
- Setup: A new site adjacent to a wind farm plus batteries, backed by a long-term PPA.
- Energy: 3 GWh across multiple annual training cycles.
- Effective emissions: approach zero on an annual basis, but require high upfront capital investment and careful grid planning.
Global Perspective: US vs EU vs Asia Data Center Strategies
Regions are approaching AI energy and emissions differently:
- United States: rapid AI cluster build-out in states with low-cost electricity (gas, wind, solar), with growing emphasis on tying federal incentives to sustainability outcomes.
- European Union & UK: stricter rules on data-center efficiency, energy disclosure, and alignment with national climate targets.
- East Asia: a mix of very large data centers operating on still carbon-intensive grids, alongside rising investment in local renewables and cross-border PPAs.
Mitigation Levers: Efficiency, Scheduling, and Procurement
Simplified LLM Training Emissions Breakdown
Key levers to reduce AI training footprint include:
- Model & algorithmic efficiency: better architectures, sparsity, and curriculum learning reduce FLOPs per unit of model quality.
- Hardware & PUE: high-efficiency accelerators and liquid cooling reduce kWh per FLOP.
- Temporal & spatial shifting: scheduling heavy runs when and where renewable output is high.
- Procurement: long-term PPAs, on-site solar plus storage, or participation in grid-interactive data-center programmes.
Devil's Advocate: Why Green AI Is Hard in Practice
Even with strong sustainability goals, several factors complicate reducing the training footprint:
- Speed-to-market pressure: model teams often prefer the fastest available capacity rather than the cleanest options.
- Metric uncertainty: the lack of standardised methods for counting FLOPs, energy, and emissions makes project-to-project comparisons difficult.
- Demand growth: efficiency gains can be offset by larger models and more experiments, creating a rebound effect.
Outlook to 2030: Scenarios for AI's Energy Footprint
By 2030, most realistic scenarios suggest that:
- global data-center electricity consumption will continue to grow, but LLM training will remain a subset of the total, albeit highly concentrated in specific locations.
- regulatory and investor pressure will drive towards mandatory disclosure of energy use and emissions for major AI projects.
- leading projects will turn large-scale model training into planned green loads, tied to dedicated wind and solar assets instead of acting as unplanned demand on carbon-intensive grids.
For enterprises, the core message through 2030 is to treat AI training as a major infrastructure investment—not a transient compute job—and to integrate it into company-wide energy and sustainability strategy.
Tools to quantify and reduce your footprint
- Business Carbon Footprint Calculator (scope-level baseline for IT and operations).
- Solar Calculator (estimate on-site generation potential for data center loads).
- AI Energy Advisor (practical steps to cut AI workload energy use).