The Carbon Cost of Training an AI Model

Every time you generate an image or ask an AI to write you a cover letter, somewhere a data center is drawing power — a lot of it. The environmental impact of AI is real, growing, and poorly understood by most users.

Training vs. inference

There are two distinct phases of an AI model's carbon life. Training — the one-time process of building the model — is enormously energy-intensive. GPT-3's training run was estimated to emit roughly 550 tons of CO₂, equivalent to about 60 round-trip flights between New York and London. GPT-4 and similar frontier models are believed to have training costs orders of magnitude higher, though companies don't publish exact figures.

Inference — actually using the model — is cheaper per query but adds up at scale. Every ChatGPT query uses roughly 10x the energy of a Google search. With hundreds of millions of daily queries, the cumulative impact is significant.

Data center water consumption

Less discussed than energy: AI data centers consume enormous amounts of water for cooling. Microsoft reported that its water consumption increased by 34% in 2022, largely attributable to AI workloads. A 2023 study estimated that training GPT-3 required roughly 700,000 liters of fresh water.

The renewable energy picture

Major cloud providers (Google, Microsoft, Amazon) have made commitments to match their energy consumption with renewables. Google claims to match 100% of its electricity with renewable energy purchases. However, "matching" isn't the same as "running on" — renewable energy certificates can be purchased from sources disconnected from actual consumption. The grid the data center actually draws from matters, and many regions still rely heavily on natural gas peaker plants during high-demand periods.

What can users do?

The individual user's impact is real but modest compared to the systemic level. The more tractable interventions are at the infrastructure level: demanding that AI companies publish actual energy and emissions data, choosing providers with verifiable (not just claimed) renewable commitments, and supporting policy that requires data center energy transparency.

Quantifying the footprint

The energy consumption of AI is difficult to measure precisely, but the orders of magnitude are striking. Training GPT-3 was estimated to consume around 1,300 MWh of electricity — equivalent to the annual consumption of approximately 120 US households. GPT-4 training figures remain undisclosed, but credible estimates suggest the energy cost was substantially higher. For context, a single Google search uses roughly 0.3 watt-hours; a single GPT-4 query uses approximately 10 times more.

The aggregate scale is what matters: hundreds of millions of AI queries per day across all providers adds up to a significant and growing fraction of global data centre electricity consumption. The International Energy Agency projected that data centres could account for 4-6% of global electricity demand by 2030, with AI workloads being a major driver of that growth.

Water consumption: the invisible cost

Less discussed than energy but equally significant is water consumption. Data centres use water for cooling — either directly through evaporative cooling towers or indirectly through the water used to generate the electricity they consume. Microsoft disclosed that its global water consumption increased by 34% in the year it launched heavy AI workloads, primarily attributable to training large models.

In water-stressed regions, this creates direct competition with agricultural and residential needs. The geography of data centres — often located in regions chosen for cheap land and energy, which sometimes correlates with water scarcity — amplifies this concern. The environmental accounting of AI must include water alongside carbon to be complete.

The efficiency research frontier

The AI research community has recognised this challenge and is actively working on efficiency. Model distillation compresses large models into smaller ones that retain most capability with a fraction of the parameters — and therefore a fraction of the inference cost. Quantisation reduces the numerical precision used to represent model weights, shrinking memory and compute requirements significantly.

Hardware advances are also compressing the energy cost per computation. Specialised AI chips (Google's TPUs, Nvidia's latest GPU generations, custom silicon from Apple and Amazon) perform AI inference far more efficiently than general-purpose processors. The combination of algorithmic efficiency and hardware improvement means the energy cost per useful AI output is falling even as aggregate consumption rises.

Using AI to fight climate change

The relationship between AI and environmental sustainability is not one-directional. AI is increasingly a tool for accelerating climate solutions. Google DeepMind's AlphaFold has revolutionised protein structure prediction, dramatically accelerating research into biological carbon capture, enzyme engineering, and novel materials for solar cells and batteries.

AI-powered grid management is enabling higher penetration of intermittent renewable energy by predicting supply and demand with enough accuracy to balance the grid without fossil fuel backup. Smart building systems using AI optimisation are reducing commercial building energy consumption by 10-20%. The question is whether these efficiency gains and clean energy innovations outpace the resource consumption of the AI systems enabling them — a race that will define much of the next decade's environmental trajectory.