Artificial intelligence is transforming our world at a dizzying pace, but at what environmental cost? I have been asking myself this question since I discovered the alarming study by the Green IT association, which projects a 6.9-fold increase in the environmental footprint of global AI infrastructure between 2025 and 2030. From 13 to 25 million GPUs, from 174 to 1,169 TWh per year: the figures are staggering. What really struck me is that the electricity consumed by AI-dedicated data centers would grow much faster than the number of processors themselves. The bulk of these impacts would come from electricity consumption, not from the manufacturing of servers. This reality forces us to rethink our approach to AI and its massive deployment. In this article, I propose to explore the concrete data from this study, to understand the true environmental impacts beyond just the carbon footprint, and to discover the recommendations that Green IT addresses to public authorities to contain this exponential growth.
📋 Summary
Twice as many GPUs, seven times more electricity
The Green IT study is based on a life cycle analysis (LCA), a standardized method that adds up the impacts of a product at each stage of its existence. I find this approach particularly relevant because it does not limit itself to just the carbon footprint. The calculation starts from the GPU fleet, meaning the total number of graphics processors in service in data centers. Green IT estimates this fleet at 13 million units by the end of 2025 and nearly 25 million by the end of 2030, based on market studies that anticipate a 190% increase in sales. These chips, originally designed for display, perform most of the calculations used to train and run AI models 💻.
What really surprised me was the increase in electricity consumption. It would rise from 174 TWh per year in 2025 to 1,169 TWh in 2030, nearly seven times more. The cause is the average power drawn by each GPU, which triples from about 700 to over 2,000 watts. The racks, those standardized metal cabinets that house the servers, are also becoming denser thanks to liquid cooling. A high-end rack would thus increase from 198 kW to 1 MW of power.
The number of servers would triple, from 1.9 to 6.2 million, and the surface area of data centers would be multiplied by 6 to house and cool these machines. The first edition of the study, published last year, only anticipated a fourfold increase in GPU electricity consumption. The authors themselves acknowledge: “We can question our physical capacity to reach such volumes.”
The climate represents only a third of the AI footprint
I must admit that this discovery changed my perspective on the environmental impact of AI. The LCA from Green IT evaluates 16 indicators according to the Product Environmental Footprint (PEF) method of the European Commission. The global warming accounts for 35% of the footprint in 2025 and 36% in 2030, which means that 64% of the impacts come from elsewhere 🌍. The eutrophication of freshwater, an excess of nutrients in rivers and lakes that causes algae to proliferate and suffocates wildlife, represents 23% and then 25% of the impacts and mainly results from coal extraction.
The depletion of fossil resources remains stable at 17%, with an additional 2% for metals and minerals. Fine particles account for 15% of the total footprint. Green IT associates them with 3,217 new cases of diseases worldwide in 2025, then 20,639 in 2030. This is an increase of over 540%, highlighting the scale of the health problem related to air pollution.
Frédéric Bordage, founder of GreenIT.fr and co-author of the study, already defended this multicriteria reading last year. He explained that greenhouse gases accounted for only 11% of the impacts of digital technology as a whole. “If a company limits itself to the carbon footprint of its information system, it misses 89% of the impacts,” he pointed out. This perspective really changes the way we should evaluate the real impact of AI.
Electricity far ahead of server manufacturing
On the AI side, the production of the electricity consumed by data centers concentrates 92% of the impacts in 2025 and 96% in 2030. The share of manufacturing decreases from 7% to 4%. According to the report, the increase in consumption “compresses the impacts associated with the manufacturing of servers.” I find this dynamic interesting because it shows that the real problem is not so much the production of machines as their operation ⚡.
Manufacturing remains, however, the primary source of depletion of metals and minerals. At the scale of a server, the main impact of manufacturing is not the GPU but the RAM, regardless of the range, excluding metals and minerals. Manufacturing a single AI server is equivalent to 8 to 14 sustainable annual budgets of an individual, based solely on climate. This is a considerable environmental burden right from production.
This reality pushes us to rethink our energy optimization strategy. If we truly want to reduce the impact of AI, we must focus on the energy source used to power data centers. Renewable energies thus become crucial, as well as the energy efficiency of processors and cooling systems.
CO2 emissions: from 99 to 636 million tons
The greenhouse gas emissions from AI infrastructure would rise from 99 million tons of CO2 equivalent in 2025 to 636 million in 2030, more than six times higher. The association relates them to the sustainable carbon budget, set at 985 kg of CO2 equivalent per year per person on the planet, according to the European Commission. They accounted for 22% of the European Union’s budget in 2025 and would reach 1.4 times this budget in 2030, equivalent to that of 645 million people 🌱.
Green IT attributes this growth to the massive AI offering. AI is offered for free, even imposed in everyday tools, from message summaries to search results. “The offer thus creates the demand that creates the impacts,” writes the association. This observation makes me reflect on our collective responsibility. We are creating exponential demand without really measuring the consequences.
In France, the report recalls the plan of 109 billion euros announced in early 2025, which foresees 35 new data centers dedicated to AI. Projects like Google’s in Châteauroux are already advancing. However, there is no regulation of their environmental impact at the French or European level, according to Frédéric Bordage, who believes that “the European AI Act has completely overlooked the environment.”
Green IT’s recommendations for public authorities
Green IT makes five concrete recommendations to public authorities to contain the impact of AI. The first is to legislate in favor of an “AI sobriety” plan to contain the supply. I believe this is a necessary approach, as without regulation, growth will continue exponentially. The second recommendation is to establish an “AI” labeling system, modeled after the labeling of GMOs, so that consumers know what they are using 📋.
The third recommendation is to create a sector of excellence in frugal AI, meaning AI that consumes less energy and generates fewer impacts. The fourth is to make eco-design of AIs hosted in France mandatory. Finally, the fifth recommendation is to tax, via a bonus-malus system, AI centers powered by fossil fuels. This last measure seems particularly relevant to me as it would create an economic incentive to use renewable energies.
In the United States, the first limits come from the states. New York has instituted a one-year moratorium on new data centers of 50 megawatts and more, the first in the country. This initiative shows that regulation is possible and necessary. I believe we must follow this example in Europe and France.
Understanding the methodology of the Green IT study
The study was conducted voluntarily and without funding by five authors from the Green IT association, which advocates for digital sobriety. It underwent internal critical review but not third-party review. This simplified LCA covers all AI servers in service worldwide by the end of 2025, and then those projected by the end of 2030. The calculations are based on the NegaOctet and Ecoinvent databases 🔬.
The 2030 scenario is unique and based on GPU sales forecasts, without a range. It retains the current electricity mix of the countries hosting AI data centers, primarily the United States and China. Gas or coal power plants built to supply the new sites are not included, nor are direct purchases of renewable electricity. User terminals and networks are excluded from the scope of the study.
I must emphasize that this methodological transparency is important. Green IT acknowledges the limitations of its study, which enhances its credibility. The results are expressed in “sustainable annual budgets,” meaning the amount of impacts an individual can generate in a year without exceeding planetary limits. AI mobilized 101 million in 2025, and it would mobilize 702 million in 2030.
Conclusion
I must be honest: this Green IT study has profoundly impacted me. The figures are alarming, but they also give us an opportunity to act. The environmental footprint of AI could be multiplied by 7 by 2030, but we still have time to change course. What encourages me is that solutions exist: AI sobriety, renewable energies, eco-design, regulation. The real challenge is to implement them quickly and on a large scale. I firmly believe we need to rethink our approach to AI, not to abandon it, but to make it sustainable. The impacts of AI on our society go far beyond just the carbon footprint, and we must take them into account in our decisions.
As a digital professional, I believe we have a collective responsibility. We must ask our cloud service providers and technology partners to commit to energy sobriety. We must also raise awareness among our teams and clients about the environmental issues of AI. Finally, we must support initiatives like those of Green IT that help us understand the true impacts of our technological choices. Digital transformation must be responsible and sustainable.
📝 In Brief
- The environmental footprint of AI infrastructure could be multiplied by 6.9 between 2025 and 2030, according to the Green IT association
- The electricity consumption of AI data centers would rise from 174 TWh to 1,169 TWh per year, an increase of 570%
- The climate represents only 35-36% of the total footprint of AI; freshwater eutrophication and fine particles are also major impacts
- Green IT recommends strict regulation, including an “AI” labeling system, a bonus-malus tax on fossil fuels, and an AI sobriety plan
