The AI carbon footprint is now large enough to be measured against entire nations, not just companies. Global data centres consumed an estimated 448 terawatt-hours of electricity in 2025, which would rank as the world’s 11th largest national electricity consumer if data centres were a country, according to a United Nations University report published in June 2026.
By 2030, that figure is projected to reach 945 terawatt-hours, nearly tripling the combined annual electricity use of Pakistan, Bangladesh, and Nigeria, a group of countries collectively home to more than 650 million people. The associated water footprint will match the basic annual domestic water needs of all 1.3 billion people in Sub-Saharan Africa, and the land footprint will exceed 14,500 square kilometres, roughly twice the Jakarta metropolitan area.
Separately, peer-reviewed research published in Nature Sustainability in November 2025 modelled the US-specific picture in detail, projecting that AI server deployment could generate an annual water footprint of 731 to 1,125 million cubic metres and additional annual carbon emissions of 24 to 44 million tonnes of CO2-equivalent between 2024 and 2030, depending on the scale of expansion. These are the numbers that define the environmental cost side of the AI equation, and they are large enough to require serious engagement rather than dismissal.
But they are also only half the equation, and the half that dominates public coverage is not the half that determines whether AI’s net effect on the environment is beneficial or harmful. The other half is what AI enables: renewable energy optimisation, grid management, predictive maintenance for clean infrastructure, climate modelling, and emissions reduction applications that could deliver environmental benefits substantially exceeding AI’s own footprint if deployed at scale.
The AI carbon footprint story in 2026 is a genuine paradox, not resolvable by citing only the cost or only the benefit, and understanding both sides with the specificity the numbers deserve is the precondition for making good decisions about it at every level from individual tool choice to national energy policy.
The Cost Side: What Drives the AI Carbon Footprint
The UN University research clarifies an important distinction that most coverage misses: inference, not training, drives 80 to 90 percent of AI’s ongoing energy use. The energy required to train a large AI model is significant but one-time. The energy required to run billions of daily user interactions with that model accumulates continuously and grows with adoption rather than staying fixed.
ChatGPT alone is estimated to process around 2.5 billion prompts per day, translating to roughly 383 gigawatt-hours of electricity a year for a single product. This means the AI carbon footprint is directly coupled to success: the more useful AI becomes and the more it is used, the larger its energy consumption grows, independent of improvements in training efficiency.
The report adds a dimension most environmental coverage ignores entirely: task-level variation. A typical conversational chat query uses around 200 times the energy of basic text classification. Generating a single AI image requires around 1,450 times that baseline. A single short AI-generated video can consume as much electricity as 200,000 spam classifications. Model choice, prompt length, output format and resolution all materially shape the footprint, yet most of these decisions are made invisibly by product defaults the user never sees.
The Rebound Effect That Undermines the AI Carbon Footprint Gains
Google reported in 2025 that it had reduced the median energy consumption per Gemini prompt by a factor of 33 and the associated carbon footprint by a factor of 44 within a single year of deployment optimisation. That efficiency improvement rate, if sustained, changes the long-term energy trajectory substantially. But the UN University report is explicit that efficiency gains alone will not contain growth, invoking the rebound effect, sometimes called the Jevons Paradox: as models become more efficient, they become cheaper and are used more frequently, and without explicit limits on tokens, resolution, or default output length, improvements at the per-query level are easily wiped out by sheer volume growth.
More significantly for the net environmental calculation, the applications that AI enables in energy systems, as covered in detail in LiveAIWire’s report on how AI is accelerating renewable energy deployment, are already delivering quantified benefits: the IEA identified up to 175 gigawatts of additional transmission capacity unlockable through AI grid optimisation, and up to $110 billion in annual power plant cost savings by 2035.
The Local Costs, Distant Benefits Problem Behind the AI Carbon Footprint
The benefits and burdens of AI’s global expansion are highly unequal, and the UN University report documents concrete cases. In Ireland, data centres accounted for 21 percent of total metered electricity in 2023, exceeding all urban households combined, and the national grid operator has paused new approvals around Dublin until 2028.
In Uruguay, plans for a water-intensive data centre coincided with a 2023 drought that depleted Montevideo’s freshwater reserves, making tap water unsafe to drink. Only 32 countries currently host AI-specialised data centres, and more than 90 percent of that capacity is concentrated in just two countries, the United States and China, while more than 150 countries have little or no access to sovereign AI compute.
What Needs to Change to Resolve the Paradox
The resolution of the AI environmental paradox depends on several converging decisions being made right now. Data centre siting and energy procurement matters enormously: the carbon intensity of AI workloads varies dramatically depending on where the data centres are located and what energy sources they use, and the UN University report notes that switching from coal to bioenergy can cut carbon footprint by 70 percent while increasing water footprint more than thirtyfold, meaning low-carbon is not automatically low-water or low-land.
Disclosure requirements matter just as much: insufficient environmental reporting from AI companies means the market cannot currently price environmental cost accurately, which removes one of the mechanisms that would otherwise drive efficiency investment.
Understanding the scale of the investment flowing into AI infrastructure is the starting point for understanding why the demand side of this race is not going to slow down voluntarily. The supply side, renewable energy and grid decarbonisation, needs to keep pace with deployment that the private sector is funding at a scale that public investment in clean energy has not historically matched. The expansion of ambient AI into everyday environments represents a further growth vector for AI energy demand that most environmental analyses have not yet fully incorporated.
The Disclosure Gap That Prevents Honest Accounting
The most important practical problem in the AI carbon footprint debate is not the scale of the impact but the quality of the data. Companies report aggregate data centre energy and water use; they do not, in most cases, separately report what portion of that consumption is driven by AI workloads versus general computing.
This means the uncertainty ranges on AI environmental footprint estimates are wide enough to span genuinely different policy conclusions: at the lower end of the range, AI’s footprint is significant but manageable with existing decarbonisation trajectories; at the upper end, it threatens to undermine Paris Agreement targets, which call for a 53 percent reduction in data centre emissions by 2030.
The UN University report proposes a six-principle governance framework built on transparency, efficiency by design, equity and environmental justice, lifecycle responsibility, global cooperation, and sustainable use. That framework is the standard infrastructure that would allow more precise accounting, but standards cannot substitute for the political will to mandate the disclosure they enable. The gap between what could be measured and what is being disclosed is itself a policy failure, and closing it is the prerequisite for any honest resolution of whether AI’s net environmental effect is what its proponents claim or its critics fear.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.