The Big Picture

Every time you ask an AI assistant a question, generate an image, or run a code completion, you're consuming electricity. Not much on its own — roughly 0.24 watt-hours for a typical text prompt, according to Google's 2025 Gemini study, about the same as watching nine seconds of television. But multiply that by billions of queries per day, add the training runs for frontier models, the cooling infrastructure, the networking gear, and the manufacturing of specialized hardware, and the numbers become staggering.

“Every time you ask an AI assistant a question, generate an image, or run a code completion, you're consuming electricity.”

The International Energy Agency (IEA) estimates that data centers consumed approximately 415 terawatt-hours (TWh) of electricity in 2024 — roughly 1.5% of global electricity demand. Their projection for 2030: 945 TWh, with AI identified as the primary driver of that growth. To put that in perspective, 945 TWh is more than the entire electricity consumption of Germany and France combined.

This post breaks down what we actually know about AI's environmental footprint — the energy, the water, the hardware, and the policy responses taking shape around the world.

The Energy Picture: Training vs. Inference

AI's energy consumption happens at two main stages: training (building the model) and inference (using it). The narrative has long focused on training — the headline-grabbing numbers about GPT-4's multi-month training runs consuming tens of gigawatt-hours. But the reality is shifting.

Training: Still Significant, But Better Understood

Training large models remains energy-intensive. Early analyses of frontier models showed rapid historical growth in training compute, though recent trends have moderated as efficiency gains, model architecture improvements, and hardware advances have taken effect. The key insight is that training is a one-time cost — a model is trained once and used millions or billions of times.

Inference: The Hidden Majority

What matters more for the long-term environmental footprint is inference — the ongoing energy cost of actually running AI systems in production. As AI becomes embedded in search engines, productivity tools, customer service, code editors, and creative workflows, inference workloads are growing far faster than training.

The numbers vary dramatically by task:

Sources: ACM FAccT 2024 benchmark study; Google 2025 Gemini environmental study; University of Michigan Llama 3.1 measurements.

The range is enormous — three orders of magnitude between a simple classification and an inefficient image generation. This means that system design choices matter enormously. A search engine that replaces a 10-millisecond database lookup with a 405B-parameter model call is making a very different environmental bet than one that uses a small, distilled model for the same task.

The Efficiency Paradox

Here's the good news: efficiency is improving rapidly. Google reported that software and infrastructure improvements reduced energy use by a factor of 33× and carbon emissions by a factor of 44× for a typical Gemini prompt over a single year. Hardware efficiency continues to improve with each generation of accelerators. Model distillation, quantization, pruning, and speculative decoding all reduce the energy cost per query.

The bad news: Jevons paradox applies. As AI becomes cheaper and more efficient, it gets used more. The total energy consumption is rising even as per-query energy falls, because the volume of queries is growing exponentially.

The Water Footprint: An Often-Overlooked Cost

Data centers generate enormous amounts of heat. Cooling that heat requires water — either directly through evaporative cooling towers, or indirectly through the water used in electricity generation.

The numbers are sobering:

  • Global AI water withdrawal could reach 4.2–6.6 billion cubic meters by 2027, according to research by Li and colleagues published in Communications of the ACM (2025).
  • Training GPT-3 in Microsoft's US data centers consumed an estimated 700,000 liters of onsite water and 5.4 million liters total when including offsite electricity-related water use.
  • A 2025 Reuters analysis, citing Verisk Maplecroft data, found that an average mid-sized data center uses about 1.4 million liters of water per day for cooling.
  • The same analysis warned that Phoenix would experience a 32% increase in annual water stress if currently planned data centers come online.

Location matters enormously. Research on US data centers found that one-fifth of servers' direct water footprint came from moderately to highly water-stressed watersheds, while nearly half of servers were fully or partially powered by plants located in water-stressed regions.

A 2025 study published in Nature Sustainability projected that US AI servers alone could require 731 to 1,125 million cubic meters of water annually by 2030, with 24 to 44 million metric tons of carbon emissions — and concluded that current growth trajectories are unlikely to meet net-zero targets without policy intervention.

Hardware and E-Waste: The Physical Side

AI's environmental footprint isn't just electricity and water. The hardware itself has a lifecycle:

Manufacturing. Specialized AI accelerators (GPUs, TPUs, NPUs) require complex supply chains involving critical minerals whose extraction and processing carry significant environmental and social costs. Semiconductor fabrication is itself water-intensive, requiring large quantities of ultrapure water.

E-waste. A 2024 study in Nature Computational Science estimated that generative AI could add between 1.2 and 5 million tonnes of e-waste by 2030. In the study's higher-end scenarios, this would represent up to 12% of projected global e-waste. The good news: the same study found that circular-economy strategies along the generative AI value chain could reduce AI-related e-waste by 16–86%.

For context, the Global E-waste Monitor 2024 reported that the world generated 62 million tonnes of e-waste in 2022, projected to rise to 82 million tonnes by 2030.

The Corporate Reality Check

The major AI companies are caught between ambitious climate commitments and the reality of AI infrastructure buildout. Google's 2024 environmental report revealed that its total greenhouse gas emissions increased 13% year-over-year in 2023, primarily because of increased data center energy consumption and supply-chain emissions. This came despite Google's long-standing claim of being carbon-neutral since 2007 and its 2020 goal of achieving net-zero emissions across its operations by 2030.

Microsoft, Amazon, and Meta face similar tensions. All have announced renewable energy procurement, efficiency improvements, and water stewardship programs. But the scale of new data center construction — and the fact that renewable energy isn't available 24/7 everywhere — means that the carbon footprint of AI depends heavily on where and when computing happens.

Policy Responses: A Fragmented Landscape

Governments are beginning to respond to AI's environmental footprint, though the policy landscape remains fragmented.

United States. The Artificial Intelligence Environmental Impacts Act of 2024 (S. 3732) would require a federal study on AI's environmental impacts and establish a voluntary reporting system. At the state level, at least 27 states are considering or have passed legislation related to data center development. California, Ohio, and Utah were the first to require data center developers to bear the costs of new energy infrastructure. Some states are also pushing for water-use reporting requirements — an issue not addressed at the federal level.

European Union. The EU's Energy Efficiency Directive introduced reporting obligations for large data centers. The European Commission maintains a database collecting information on the energy performance and water footprint of data centers across member states. National AI strategies in France, Germany, and Italy all include references to sustainability, energy efficiency, and environmental applications of AI.

The gap. What's missing is a standardized, mandatory framework for measuring and reporting AI's environmental impacts. Without consistent metrics, comparisons between companies and systems are difficult, and regulators lack the data needed for evidence-based policy.

The Counterargument: AI as Climate Solution

It would be incomplete to discuss AI's environmental footprint without acknowledging its potential as a climate solution. AI is being used for:

  • Weather forecasting. Systems like Huawei's Pangu-Weather have demonstrated strong medium-range forecasting performance relative to leading numerical weather prediction systems.
  • Grid optimization. AI can optimize renewable energy integration, demand forecasting, and grid balancing.
  • Traffic optimization. Google's Green Light project uses AI to optimize traffic signal timing, reducing stop-and-go traffic and associated emissions.
  • Climate research. AI accelerates climate modeling, materials science for batteries and solar panels, and carbon capture research.

Whether AI produces net environmental benefits at scale remains an open question. The answer depends on deployment choices, rebound effects, the pace of grid decarbonization, and whether efficiency gains outpace demand growth.

What This Means

AI's environmental footprint is real, growing, and unevenly distributed. A single image generation can consume 11 Wh — equivalent to half a smartphone charge. A single large-model training run can consume millions of liters of water. And the aggregate trajectory points toward data centers consuming nearly 1,000 TWh annually by 2030.

But the picture isn't static. Efficiency is improving rapidly — 33× in a single year for one major system. Hardware is getting more efficient. Policy is beginning to catch up. And AI itself is being deployed to solve environmental problems.

The key takeaway: transparency matters. We need standardized, mandatory reporting of energy use, water use, carbon emissions, and e-waste across the AI industry. We need to know what we're measuring before we can manage it. And we need to make conscious choices about when to use a 405B-parameter model versus a distilled alternative — because the environmental difference between those two choices is a factor of 60×.

The intelligence revolution is real. But it shouldn't come at the cost of the planet we're trying to understand and protect.

---

Sources: IEA Electricity 2024 report; ACM FAccT 2024 benchmark; Google 2025 Gemini environmental study; University of Michigan Llama 3.1 measurements; Communications of the ACM (2025) — Li et al.; Nature Computational Science (2024); Nature Sustainability (2025); Global E-waste Monitor 2024; Reuters/Verisk Maplecroft (2025); EU Energy Efficiency Directive; S. 3732 (US Congress).