Carbon Footprint of AI

10xmore energyper query

Global avgAI chatbot vs Google search

This figure is a global average from peer-reviewed research. Actual values vary by country due to differences in energy grids, farming practices, and supply chains. Use our calculators for country-specific estimates.

Key Takeaway

Artificial intelligence is one of the fastest-growing sources of technology-related carbon emissions. Training a single large language model like GPT-4 is estimated to produce roughly 5,000 tonnes of CO2. Each ChatGPT query consumes approximately 10 times the energy of a standard Google search. As AI adoption accelerates across industries, data centre energy demand is projected to double or triple by 2030.

In this article

Training vs Inference: Two Phases of AI Emissions
Energy Per Query: AI vs Traditional Search
The Data Centre Boom and Grid Pressure
How to Reduce AI's Carbon Footprint
Frequently Asked Questions

Training vs Inference: Two Phases of AI Emissions

AI emissions fall into two phases. Training is the process of building a model by processing vast datasets on thousands of specialised GPU chips over weeks or months. Training GPT-3 consumed an estimated 1,287 MWh of electricity and produced roughly 552 tonnes of CO2, according to research from the University of Massachusetts Amherst. GPT-4 and newer models are estimated to have required 5-10 times more compute, placing training emissions in the range of 3,000-10,000 tonnes CO2. Inference is when the trained model responds to user queries. While each individual query uses far less energy than training, the cumulative impact of inference is enormous due to scale: ChatGPT alone handles hundreds of millions of queries daily. Industry estimates suggest that inference accounts for 60-90% of total AI compute over a model's lifetime, meaning the ongoing cost of serving AI to users far exceeds the one-off training cost.

Energy Per Query: AI vs Traditional Search

The International Energy Agency estimates that a standard Google search consumes approximately 0.3 Wh of electricity, producing about 0.2 g CO2. A ChatGPT query, by contrast, consumes roughly 3 Wh — approximately 10 times more energy. This difference arises because AI inference requires running billions of parameters through specialised GPU hardware that draws significantly more power than traditional search indexing. Google's own AI-powered search summaries (AI Overviews) fall somewhere between, using roughly 3-5 times the energy of a standard search. As AI becomes integrated into more products — email, office suites, photo editing, coding assistants — the aggregate energy impact grows substantially. Goldman Sachs estimates that data centre power demand from AI could increase by 160% by 2030, requiring an additional 47 GW of generation capacity — equivalent to adding several medium-sized countries' electricity consumption.

The Data Centre Boom and Grid Pressure

AI's energy demands are driving an unprecedented boom in data centre construction. Major technology companies including Microsoft, Google, Amazon, and Meta have collectively announced over $200 billion in data centre investments. Microsoft signed a deal to restart a unit at Three Mile Island nuclear plant specifically to power AI data centres. Google's total emissions rose 48% between 2019 and 2023, with data centre energy being the primary driver — undermining the company's net-zero pledges. In Ireland, data centres already consume roughly 20% of the national electricity supply, creating tension with residential and industrial users. The water footprint is also significant: data centres require cooling, and a single large facility can consume 1-5 million litres of water daily. Training GPT-3 alone consumed an estimated 700,000 litres of freshwater for cooling. These pressures are forcing a reckoning between AI ambitions and environmental commitments.

How to Reduce AI's Carbon Footprint

Efforts to reduce AI emissions span hardware, software, and infrastructure. More efficient chip architectures like Google's TPU v5 and NVIDIA's H100 deliver significantly more computation per watt than predecessors. Model distillation and quantisation can create smaller, faster models that deliver 80-90% of a large model's capability at a fraction of the energy cost. Locating data centres in regions with clean electricity — such as Nordic countries with abundant hydroelectric and geothermal power — can dramatically reduce per-query emissions. Running training jobs during periods of peak renewable generation further improves the energy mix. On the user side, choosing smaller or more efficient AI models when a full frontier model is unnecessary reduces energy consumption. Using AI selectively rather than for every trivial query helps control aggregate demand. Regulatory frameworks requiring transparency in AI energy reporting and carbon disclosure are emerging in the EU and being considered elsewhere.

Frequently Asked Questions

How much CO2 does a ChatGPT query produce?

A single ChatGPT query consumes roughly 3 Wh of electricity and produces approximately 1-3 g of CO2 depending on the grid powering the data centre. This is roughly 10 times the energy of a standard Google search but still small in absolute terms — equivalent to running a LED bulb for about 10 minutes.

How much energy did training GPT-4 use?

Exact figures have not been publicly disclosed, but estimates based on known hardware and training duration suggest GPT-4 consumed 30,000-80,000 MWh of electricity during training, producing roughly 3,000-10,000 tonnes of CO2. This is equivalent to the annual electricity use of 3,000-8,000 average US homes.

Is AI making climate change worse?

AI's growing energy demands are putting upward pressure on emissions, particularly as data centre construction outpaces renewable energy buildout. Google, Microsoft, and Amazon have all seen emissions rise partly due to AI. However, AI also enables climate solutions such as optimising energy grids, improving weather forecasting, and accelerating materials science for clean energy.

What can I do to reduce AI's environmental impact?

Use AI tools selectively rather than for every task. Choose smaller, more efficient models when possible. Support companies that power AI with renewable energy and disclose their emissions transparently. Advocate for regulatory frameworks requiring AI energy and carbon reporting.