The Environmental Cost of AI

Every time you ask an AI model a question, a server somewhere draws electricity, generates heat, and consumes water to cool itself back down. Multiply that by hundreds of millions of daily queries, add the months-long training runs that created the model in the first place, and factor in the mining, manufacturing, and shipping of the hardware — and you begin to see the outline of an environmental footprint the industry would rather not discuss.

The cost is large. It is growing. And it is distributed in a pattern that should by now be familiar: the benefits flow to wealthy companies and their users, while the environmental burden falls on communities with the least power to refuse it.

The Energy Bill

Start with electricity, the most visible cost.

Training a single frontier AI model is extraordinarily energy-intensive. Nvidia has estimated that training GPT-4 consumed approximately 50 gigawatt-hours of electricity — roughly what a small city of 40,000 people uses in an entire year, compressed into a few months of continuous computation (de Vries, 2023). Each new generation of model tends to be larger than the last, and training costs have scaled accordingly.

But training is a one-time event. Inference — the ongoing process of actually serving responses to users — runs continuously, and its aggregate energy consumption is catching up. A single ChatGPT query uses roughly ten times the electricity of a standard Google search (IEA, 2024). At hundreds of millions of queries per day, the arithmetic is relentless. Some researchers estimate that inference energy will surpass training energy within the next few years as AI products reach mass adoption.

The macro numbers capture the scale. The International Energy Agency projects that global data center electricity consumption could reach 1,000 terawatt-hours by 2026. That is roughly equivalent to Japan's entire electricity consumption — the world's fifth-largest economy, 125 million people, every factory and home and bullet train (IEA, 2024). AI is not the only driver of data center growth, but it is the fastest-growing one, and it is reshaping energy markets in real time.

In the United States, the surge in data center demand is outstripping renewable energy supply. New natural gas plants are being built specifically to power AI infrastructure. The US Energy Information Administration projects that power-sector carbon emissions, which had been declining for over a decade, will flatten or rise through the late 2020s as data centers absorb renewable capacity that would otherwise have displaced fossil fuels (EIA, 2025). The AI industry's electricity appetite is not just large — it is actively slowing decarbonisation.

The Water Nobody Talks About

Electricity generates heat. Thousands of GPUs running inference around the clock produce enormous thermal loads, and the most common method of managing that heat is evaporative cooling — which consumes water, vast quantities of it.

Microsoft reported a 34% increase in water consumption in 2022, attributing much of the rise to its expanding AI infrastructure. Google's water use grew 20% over the same period. Researchers at the University of California, Riverside estimated that a typical conversation with ChatGPT — ten to fifty exchanges — consumes roughly 500 millilitres of water for server cooling alone (Li et al., 2023). Half a litre of water for a chat. Scale that to hundreds of millions of conversations per day and the numbers become extraordinary.

The problem sharpens in water-stressed regions. Data centres in Arizona, central Texas, and parts of the Middle East draw from the same aquifers and municipal systems that serve agriculture and residential users. In The Dalles, Oregon — a town of 15,000 that hosts major Google facilities — water consumption by data centres has become a persistent source of community tension. The water that cools servers does not irrigate crops, fill reservoirs, or flow through taps.

Unlike electricity, water consumption data is difficult to obtain. Companies report it selectively, often buried in sustainability reports with favourable framing. Many data centres do not disclose water usage at all. The opacity is not accidental. It is a strategy.

Carbon: What the Ledgers Leave Out

The headline carbon story is straightforward enough: data centres that run on fossil fuels produce carbon emissions. But the full picture is considerably more complex — and worse — than the headline suggests.

Microsoft's own reporting shows that its total emissions rose 30% between 2020 and 2024, driven largely by the expansion of AI infrastructure. This increase directly threatens the company's high-profile pledge to be carbon-negative by 2030 (Microsoft, 2024). Google's emissions have followed a similar trajectory. The companies that made the most ambitious climate promises are the ones whose AI ambitions are making those promises hardest to keep.

Corporate accounting often obscures the real picture. Many technology companies claim to operate on 100% renewable energy by purchasing Renewable Energy Credits — essentially paying someone else, somewhere else, to generate clean energy. The data centre itself may still run on the local grid, which may be predominantly fossil-fuelled. RECs do not reduce the emissions coming out of a specific power plant; they are an accounting mechanism, not a physical one.

Then there are Scope 3 emissions — the supply chain. Manufacturing a modern GPU involves rare earth mining, chemical processing, precision fabrication at plants like TSMC's facilities in Taiwan (themselves enormous energy consumers), assembly, packaging, and global shipping. These upstream emissions are substantial but routinely excluded from corporate climate reporting. When companies talk about their carbon footprint, they are usually talking about a carefully bounded subset of it.

The Material Footprint

Kate Crawford's Atlas of AI traces the physical supply chain of artificial intelligence from lithium mines in Nevada to cobalt mines in the Democratic Republic of Congo. AI is not a weightless cloud phenomenon. It depends on a global extraction chain: rare earth elements, precious metals, water, sand for concrete, copper for wiring (Crawford, 2021).

The hardware refresh cycle compounds the extraction. The competitive pressure among AI companies drives rapid GPU turnover — Nvidia's H100, released in 2023, was already being supplanted by the B200 architecture roughly two years later. Each generation renders the previous one less competitive, creating powerful incentives to replace hardware on accelerated timescales. The result is significant electronic waste: circuit boards, cooling systems, and server chassis that are difficult and expensive to recycle.

This is a pattern familiar from the consumer electronics industry, but at industrial scale. And unlike smartphones, where right-to-repair movements have gained some traction, the AI hardware lifecycle operates largely outside public scrutiny.

Who Raised the Alarm

These concerns are not new. In 2021, Emily Bender, Timnit Gebru, and their co-authors published "On the Dangers of Stochastic Parrots," which identified the environmental costs of large language models as one of the central risks of the field's trajectory (Bender et al., 2021). The paper's reception — Gebru's departure from Google after the company attempted to suppress it — itself illustrates the industry's discomfort with environmental accountability.

The paper argued that the rush to build ever-larger models imposed environmental costs that were poorly measured, rarely disclosed, and borne by people who had no say in the decision. Four years later, with models orders of magnitude larger and energy consumption orders of magnitude higher, the argument has only grown stronger.

The Redistribution Question

Here is where the environmental story becomes a redistribution story.

The costs are externalised. Communities near data centres bear the noise, the water draw, the grid strain, and the foregone tax revenue from generous abatements. Communities near mines bear the ecological damage of extraction. Communities near fossil fuel plants bear the health consequences of the emissions that power AI inference. These communities are overwhelmingly rural, lower-income, or in the Global South.

The benefits concentrate elsewhere. AI companies and their shareholders capture the value. Users — disproportionately in wealthy countries — enjoy the convenience. The redistribution runs in one direction: environmental costs flow down; financial returns flow up.

Climate change makes this worse. The regions most exposed to the consequences of rising emissions — sub-Saharan Africa, South Asia, small island states — are also the regions that bear the environmental costs of AI hardware supply chains while receiving the fewest benefits from AI products. The injustice is not incidental. It is structural, embedded in where data centres get built, where minerals get mined, and where pollution gets tolerated.

"Green AI" claims rarely withstand scrutiny. Buying carbon offsets is not the same as reducing emissions. Purchasing RECs is not the same as powering a data centre with renewables. Planting trees in one country does not undo the water drawn from an aquifer in another. Much of what passes for corporate environmental responsibility in the AI industry is, plainly, greenwashing — a public relations strategy, not a climate strategy.

What Should Change

If the environmental costs of AI are to be taken seriously, several things need to happen.

Mandatory environmental reporting. Every frontier model training run should disclose its total energy consumption, water use, and carbon emissions. Inference costs should be reported per-query at an aggregate level. These numbers should be audited, not self-reported. The EU AI Act gestures in this direction; it does not go far enough.

Scope 3 accountability. Supply chain emissions — chip manufacturing, hardware production, construction, logistics — must be included in corporate climate reporting. Excluding them is indefensible.

Data centre siting regulations. Communities should have meaningful input into whether and where data centres are built. Environmental impact assessments should account for cumulative effects on water and energy systems, not just the marginal impact of a single facility.

Real renewable energy mandates. Not RECs. Actual power purchase agreements that add new renewable generation to the grid, with temporal and geographic matching — the data centre runs on renewable electricity generated nearby, at the time it is consumed.

Efficiency requirements. Regulators should set minimum efficiency standards for AI inference, just as they set standards for appliances and vehicles. There is enormous variation in how efficiently different models use compute; creating incentives for efficiency would redirect engineering effort toward doing more with less. This is not unprecedented: the US appliance efficiency standards introduced in the 1980s saved an estimated 2.5 billion tonnes of CO2 over three decades. Efficiency standards work when they are set clearly and enforced consistently.

Right-to-repair for AI hardware. Extending the usable life of GPUs and server equipment, rather than discarding functional hardware after two-year refresh cycles, would reduce both e-waste and upstream extraction.

None of this will happen on its own. The market incentives all point in the opposite direction — toward larger models, faster refresh cycles, and externalised costs. Policy must create countervailing pressure. But it is worth noting that some countervailing trends are emerging from within the field itself. The small language model movement, edge AI, and algorithmic efficiency research (exemplified by DeepSeek's efficiency breakthroughs) all reduce environmental costs per unit of AI capability. The question is whether these efficiency gains will be consumed by growing demand — the Jevons paradox applied to compute — or whether they can be channelled toward genuinely lighter infrastructure.

The Bottom Line

AI has a material footprint. It consumes electricity, water, minerals, and land. It generates carbon emissions, electronic waste, and ecological disruption. These costs are real, growing, and systematically borne by communities and countries with the least power to refuse them.

The industry's response so far has been to report selectively, promise ambitiously, and deliver evasively. That is not a climate strategy. It is reputation management.

The question is not whether AI should exist. It is whether its environmental costs should be hidden, externalised, and ignored — or measured, allocated, and governed. The answer determines who pays.

Related

• Compute Is the New Oil
• Data Center Energy
• AI Won't Solve Climate Change
• Hardware Bottleneck
• Atlas of AI
• Stochastic Parrots
• Small Language Models
• Edge AI
• Who Controls Supply Chain

Sources

• Bender, Emily M. et al. "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" FAccT '21, 2021. ACM
• Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Publisher
• de Vries, Alex. "The Growing Energy Footprint of Artificial Intelligence." Joule 7(10), 2023. Cell Press
• International Energy Agency. "Electricity 2024: Analysis and Forecast to 2026." IEA, 2024. IEA
• Li, Pengfei et al. "Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models." arXiv, 2023. arXiv
• Microsoft. "2024 Environmental Sustainability Report." Microsoft, 2024. Microsoft
• US Energy Information Administration. "Annual Energy Outlook 2025." EIA, 2025. EIA