AI
AI Tool Environmental Impact Assessment 2025: Computing Resource Consumption and Carbon Emissions
A single query to a large language model like GPT-4 can consume **2.9 watt-hours** of electricity, compared to roughly 0.003 watt-hours for a standard Google…
A single query to a large language model like GPT-4 can consume 2.9 watt-hours of electricity, compared to roughly 0.003 watt-hours for a standard Google search — a factor of nearly 1,000x. According to the International Energy Agency (IEA) World Energy Outlook 2024, data centers, AI, and cryptocurrency consumed an estimated 460 terawatt-hours (TWh) of electricity in 2022, representing roughly 2% of global demand. The IEA projects this figure could double to 1,000 TWh by 2026, putting AI training and inference squarely in the crosshairs of global carbon accounting. This assessment breaks down the specific computing resource consumption and carbon emissions of the top AI tools — ChatGPT, Claude, Gemini, DeepSeek, and Grok — using the latest benchmark data from the 2025 MLPerf Power v1.0 suite and independent academic audits. You will find per-tool wattage, training carbon debt, inference efficiency, and a clear rating on whether the provider offsets its impact.
Training Energy and Carbon Debt
The training phase remains the single largest energy spike for any AI model. A single training run for a frontier model can emit between 500 and 10,000 metric tons of CO₂-equivalent (tCO₂e), depending on the hardware, cooling, and grid carbon intensity.
GPT-4 training footprint
OpenAI has not published exact training energy figures, but independent analysis by Luccioni et al. (2024, Estimating the Carbon Footprint of Generative AI) estimates that GPT-4’s training consumed approximately 50,000 megawatt-hours (MWh) of electricity. Using the average US grid carbon intensity of 0.4 kg CO₂e per kWh, this yields roughly 20,000 tCO₂e — equivalent to the annual emissions of 4,400 gasoline-powered cars. OpenAI has stated it purchases renewable energy certificates (RECs) to offset this, but RECs do not guarantee additionality.
Claude 3 and Gemini Ultra training
Anthropic’s Claude 3 Opus, based on its published system card, used 12,000 MWh for the final training run, producing an estimated 4,800 tCO₂e at US average grid mix. Google’s Gemini Ultra, leveraging Google’s 64% carbon-free energy portfolio (2023), trained on 25,000 MWh but emitted only 2,100 tCO₂e due to lower-carbon data center locations. DeepSeek-V2, a Chinese model, trained on 8,000 MWh but used a coal-heavy grid (0.8 kg CO₂e/kWh), resulting in 6,400 tCO₂e — the highest per-MWh carbon debt in the 2025 cohort.
Inference Efficiency Benchmarks
Inference — the per-query processing — dominates long-term operational emissions once a model is deployed. The MLPerf Power v1.0 (2025) benchmark measures wattage per inference across hardware configurations.
Wattage per query
On equivalent hardware (NVIDIA H100 SXM), the Gemini 1.5 Pro inference engine achieves 0.45 watt-hours per query, the lowest in the group. GPT-4 Turbo uses 0.52 Wh/query, while Claude 3 Opus uses 0.61 Wh/query. Grok-2 (xAI) uses 0.58 Wh/query. DeepSeek-V2 records 0.72 Wh/query, partly due to less optimized kernel fusion on the same GPU architecture. At 10 million queries per day, the difference between Gemini and DeepSeek equates to 2,700 MWh per year — roughly the annual consumption of 250 US homes.
Latency vs. energy tradeoff
Lower latency often means higher energy per token. Claude 3 Haiku (the fastest Anthropic model) uses 0.38 Wh/query but has a 40% higher energy-per-token rate than Claude 3 Opus due to smaller batch sizes. Gemini Nano, designed for on-device inference, uses only 0.08 Wh/query — a 90% reduction from cloud inference — but handles only text classification and short generation tasks. For heavy multimodal tasks, cloud inference remains unavoidable.
Data Center Water and Hardware Waste
Energy is only one dimension. Water consumption for cooling and e-waste from GPU turnover add hidden environmental costs.
Water usage effectiveness (WUE)
Google reported a global average WUE of 0.3 L/kWh for its data centers in 2024. OpenAI, using Microsoft Azure data centers, averaged 0.49 L/kWh. Anthropic uses a mix of AWS (0.41 L/kWh) and GCP (0.3 L/kWh). DeepSeek’s data centers, located in water-stressed northern China, record WUE of 1.2 L/kWh due to evaporative cooling. At 50,000 MWh annual inference load, DeepSeek consumes 60 million liters of water per year — enough for a small city of 1,500 people.
GPU lifespan and e-waste
Training clusters replace GPUs every 2–3 years. The 2024 United Nations Global E-waste Monitor reports that data center GPU e-waste reached 1.2 million metric tons in 2023, with AI accelerators representing 35% of that total. Each H100 GPU contains 0.3 kg of conflict minerals (tantalum, tin, tungsten, gold). The industry has no standardized recycling program for retired AI hardware.
Carbon Offset and Renewable Energy Claims
Providers differ sharply in how they account for and mitigate emissions.
Scope 2 and Scope 3 reporting
Google (Gemini) reports Scope 2 market-based emissions of 0 tCO₂e for its global data centers, backed by 64% carbon-free energy on an hourly basis. Microsoft (OpenAI) claims 100% renewable energy matching on an annual basis, but hourly matching remains below 50%. Anthropic publishes a voluntary carbon disclosure with 80% Scope 2 coverage. DeepSeek and xAI (Grok) do not publish Scope 3 (supply chain) emissions. For cross-border data transfers, some international teams use secure access channels like NordVPN secure access to route traffic through lower-carbon data center regions.
Additionality and carbon credits
OpenAI has purchased 1.5 million tCO₂e in voluntary carbon credits since 2023, but CarbonPlan (2024) found that 60% of those credits were from avoided-deforestation projects with low additionality. Google uses only high-additionality renewable energy PPAs (power purchase agreements). Anthropic has committed to a 50% absolute emission reduction by 2030 (from a 2023 baseline) without relying on offsets.
Per-Tool Environmental Rating (2025)
We assign a composite score based on training efficiency, inference efficiency, water use, and offset quality. Scale: A (best) to F (worst).
| Tool | Training Score | Inference Score | Water Score | Offset Score | Overall |
|---|---|---|---|---|---|
| Gemini 1.5 Pro | A | A | A | A | A |
| GPT-4 Turbo | C | B | B | C | B- |
| Claude 3 Opus | B | C | B | B | B |
| Grok-2 | D | C | C | D | C |
| DeepSeek-V2 | F | D | F | F | F |
Gemini leads due to Google’s hourly carbon-free energy matching and efficient TPU hardware. DeepSeek lags due to coal-heavy grid and no published offset program. GPT-4 Turbo suffers from opaque training data and low-quality carbon credits.
Regulatory and Future Trends
Governments are beginning to mandate disclosure.
EU AI Act and energy reporting
The EU AI Act (effective August 2025) requires high-risk AI systems to report training energy consumption in MWh and estimated tCO₂e. The US Executive Order on AI (2023) directs the Federal Energy Regulatory Commission to develop a data center energy efficiency standard by 2026. These regulations will force all five providers to publish standardized metrics.
Hardware efficiency gains
NVIDIA’s B200 GPU, shipping in Q2 2025, claims a 2.5x improvement in performance-per-watt over H100. Google’s TPU v6 (Trillium) achieves 4.7x energy efficiency improvement per training epoch. If adopted broadly, these chips could reduce per-query inference energy by 60% by 2027, even as query volume grows.
FAQ
Q1: How much CO₂ does a single ChatGPT query emit?
A single GPT-4 query emits approximately 4.32 grams of CO₂e based on 0.52 Wh/query and the US average grid carbon intensity of 0.4 kg CO₂e/kWh. That is about 1/100th of the emissions from a single Google search (0.2 g) but 1,000x more than a simple database lookup. Over 100 million daily queries, that totals 432 metric tons of CO₂e per day — roughly the annual emissions of 95 US cars.
Q2: Which AI tool has the lowest total carbon footprint?
Gemini 1.5 Pro has the lowest total carbon footprint among frontier models, with a training debt of 2,100 tCO₂e and inference at 0.45 Wh/query. Google’s hourly carbon-free energy matching across 64% of its data center hours further reduces operational emissions. On-device models like Gemini Nano (0.08 Wh/query) have the absolute lowest per-query footprint but cannot perform complex multimodal tasks.
Q3: Will AI energy consumption double by 2026?
The IEA World Energy Outlook 2024 projects that total electricity consumption by data centers, AI, and cryptocurrency could reach 1,000 TWh by 2026, up from 460 TWh in 2022. This represents a doubling in four years. However, hardware efficiency gains from next-generation GPUs and TPUs may slow the rate of increase. If all new data centers are powered by 100% renewable energy, the carbon impact could remain flat despite the growth in compute.
References
- International Energy Agency. 2024. World Energy Outlook 2024.
- Luccioni, A.S., et al. 2024. Estimating the Carbon Footprint of Generative AI. arXiv:2401.04561.
- MLCommons. 2025. MLPerf Power v1.0 Benchmark Results.
- United Nations University. 2024. Global E-Waste Monitor 2024.
- European Commission. 2025. EU AI Act: Energy Reporting Requirements for High-Risk AI Systems.