Your AI carbon footprint is larger than you think. Training GPT-4 consumed roughly 50 GWh of electricity and produced an estimated 12,456 tonnes of CO2 equivalent, comparable to driving a petrol car over one million miles. Here is the real emissions data behind AI training and how it compares to everyday activities.
What Is AI Carbon Footprint and How Is It Measured
AI carbon footprint measures total greenhouse gas emissions from model training, inference, and supporting data centre infrastructure. You calculate it by multiplying energy consumption (kWh) by the carbon intensity of the local grid (gCO2/kWh). The University of Massachusetts Amherst established this methodology in 2019, and it remains the standard used by the IEA and Stanford’s AI Index Report.
Most published figures only cover Scope 2 (purchased electricity). When you include Scope 1 (backup generators) and Scope 3 (embodied carbon in hardware manufacturing), the true footprint is 20 to 40% higher.
Training GPT-4: Full Emissions Breakdown
Training GPT-4 required approximately 25,000 NVIDIA A100 GPUs running for 90 to 100 days at 400W per GPU. Total electricity consumption reached roughly 50 GWh (Epoch AI estimates, corroborated by IEA analysis). OpenAI trained GPT-4 in Microsoft Azure facilities in Iowa and Virginia, where grid carbon intensity averages 388 gCO2/kWh (EPA eGRID 2024). Raw emissions would reach 19,400 tonnes of CO2e. Microsoft reports 60% renewable energy at these sites, bringing the adjusted figure to roughly 12,456 tonnes.
The average UK petrol car emits 164.1 gCO2/km (BEIS 2024). Driving one million miles (1.609 million km) produces approximately 11,740 tonnes of CO2e. GPT-4 training exceeded that by 6%. You can see how much energy AI uses across all major providers in our full breakdown.
Google AI Energy Consumption and Carbon Output
Google reported 25.3 TWh of data centre electricity in 2024, with AI workloads at 30 to 40%. Google AI energy consumption sits at roughly 8 to 10 TWh annually. Google’s 2024 Environmental Report disclosed 14.3 million tonnes of CO2e, a 48% increase from 2019 despite efficiency gains.
Training Gemini Ultra consumed an estimated 45 to 55 GWh. A single Gemini query produces 1.2 to 2.5 gCO2e versus 0.2 gCO2e for a standard Google Search. The AI energy crisis is accelerating because inference volumes grow faster than efficiency improvements can offset.
Emissions Per Query: AI vs Everyday Activities
| Activity | CO2e per Unit | Equivalent GPT-4 Queries |
|---|---|---|
| One GPT-4 query | 4.32 g | 1 |
| One Google Search | 0.2 g | 0.05 |
| Charging a smartphone | 8.4 g | 1.9 |
| Driving 1 mile (petrol car) | 264 g | 61 |
| One Netflix hour (streaming) | 36 g | 8.3 |
At one billion weekly GPT-4 queries, ChatGPT inference alone produces approximately 225,000 tonnes of CO2e per year, matching the annual output of roughly 49,000 petrol cars. Reducing this footprint requires nuclear power for data centres and more efficient architectures, not just carbon offsets.
How You Can Reduce Your AI Carbon Footprint
Choose providers that publish verified emissions data and run on high renewable mixes. Use smaller, distilled models when frontier capability is unnecessary. Batch queries instead of running repeated single requests. Track usage with CodeCarbon or ML CO2 Impact.
Microsoft, Google, and Amazon have committed to 100% carbon-free energy by 2030. NVIDIA’s Blackwell B200 delivers 4x energy efficiency per FLOP versus the H100. Track evolving AI data centre power consumption as these gains reach the hyperscaler fleet.
Frequently Asked Questions
How much CO2 does training a large AI model produce?
Training GPT-4 produced roughly 12,456 tonnes of CO2e after renewable energy adjustments. Smaller models like Llama 3 70B produce 1,500 to 3,000 tonnes depending on grid carbon intensity.
Is AI worse for the environment than driving?
A single query produces far less CO2 than driving a mile. At scale, training one frontier model equals driving over one million miles, and billions of weekly queries match emissions from tens of thousands of cars.
What is the fastest way to lower AI emissions?
Train in regions with low-carbon grids like Quebec, Norway, or France. Use knowledge distillation and deploy inference on latest-generation hardware for more performance per watt.