AI 2030: CO2 at 636 Mt and Electricity Increased by 7 Times, Projects Green IT

A projection from Green IT anticipates emissions from AI infrastructure to reach 636 million tons of CO2 by 2030, with an electricity consumption of 1,169 TWh. Electricity would account for up to 96% of the impacts, far exceeding the manufacturing of servers. The scenario is based on GPU sales and a defined scope, while recommendations are directed at public authorities as new data centers advance in France and a first moratorium appears in New York.

Emissions and Health: Significant Increases in Magnitude

Greenhouse gas emissions associated with artificial intelligence infrastructure are expected to rise from 99 to 636 million tons of CO2 equivalent between 2025 and 2030. Relative to the European Union's budget, they would reach 1.4 times this level by 2030, equivalent to that of 645 million people. In terms of impact distribution, electricity production consumed by data centers would concentrate 92% of the effects in 2025 and 96% in 2030. The proportion related to manufacturing would decrease from 7% to 4%, with Green IT noting that the increase in electricity consumption reduces the share of impacts attributed to server production. However, production remains the primary cause of depletion of metallic and mineral resources, and for an individual server, RAM constitutes the dominant impact factor for manufacturing outside of metals and minerals. The manufacturing of a single AI server corresponds to 8 to 14 times the annual sustainable budget for an individual regarding climate. In terms of health, the model associates fine particles with 3,217 new cases of diseases in 2025, rising to 20,639 in 2030. Global warming accounts for 35% of the impacts in 2025 and 36% in 2030.

Assumptions and Scope: A Unique and Limited Scenario

The 2030 projection presented is based on a unique scenario, without a range, founded on GPU sales forecasts. It retains the current electricity mix of countries hosting the majority of AI data centers, notably the United States and China. Power plants running on gas or coal that could be built for new sites are not taken into account, nor are direct contracts for purchasing electricity from renewable sources. End-user equipment and network infrastructure are not included in the study's scope. The life cycle assessment covers 16 indicators according to the European Commission's Product Environmental Footprint methodology, with each result compared to planetary boundaries and then weighted for comparability. The reference for the sustainable carbon budget is set at 985 kg CO2e per year per person. The estimate of water scarcity risk, which represents about 1% of the footprint, relies on data considered imperfect. Conducted voluntarily by five members of Green IT, the study underwent internal critical review without external validation and focuses on AI servers in operation at the end of 2025 as well as those projected for the end of 2030, relying on the NegaOctet and Ecoinvent databases.

Infrastructure: GPU, Servers, Surface Area, and Megawatts on the Rise

Annual electricity consumption is expected to rise from 174 TWh in 2025 to 1,169 TWh in 2030, driven by an average power per GPU that would triple, from about 700 to over 2,000 watts. The GPU fleet is estimated to reach 13 million units by the end of 2025 and approach 25 million by the end of 2030, based on market studies anticipating a 190% increase in sales. Meanwhile, the number of servers would triple, from 1.9 to 6.2 million. At the rack level, a high-end rack would increase from 198 kW to 1 MW, in infrastructures made denser by liquid cooling. To accommodate and cool these machines, the surface area of data centers would multiply by 6 by 2030 and would grow faster than just the number of GPUs. The first edition of the study projected a fourfold increase in GPU electricity consumption; the physical capacity to reach such volumes is now being questioned. The calculation includes the fleet of GPUs in operation, as these processors concentrate the majority of training and inference calculations.

Public Policies: Recommendations, National Plans, and Moratorium in New York

Green IT links the increase in impacts to the availability of free or automatically integrated AI in everyday uses and makes five proposals for authorities: adopt a law for an AI sobriety plan, implement an "AI" labeling system, develop a resource-efficient AI excellence sector, make eco-design mandatory for AIs hosted in France, and establish a bonus-malus system for AI centers powered by fossil fuels. A plan of 109 billion euros presented in France in early 2025 envisions the creation of 35 new data centers dedicated to AI, with initiatives already underway, such as Google's project in Châteauroux. According to Frédéric Bordage, there is currently no regulation governing their environmental impact at either the national or European level, and he believes that the European AI Act has sidelined environmental concerns. In the United States, the first restrictive measures are being taken at the state level: New York State implemented a one-year moratorium in July on new data centers of 50 MW and above, a first-of-its-kind measure in the country.

Beyond Carbon: An Impact Footprint Dominated by Four Families of Impacts

The projection is not limited to climate. In addition to global warming, which would account for 35% of the footprint in 2025 and 36% in 2030, three other families stand out: freshwater eutrophication would account for 23% and then 25% of the impacts, primarily linked to coal extraction; fossil resource depletion would remain at 17%, with an additional 2% for metals and minerals; finally, fine particles would represent 15%, with thousands of new disease cases associated according to the model. The risk of water deprivation is estimated at about 1%, based on imperfect data. Frédéric Bordage has long advocated for this multi-criteria perspective: according to him, at the digital scale, greenhouse gases represent only 11% of the impacts, and a company limiting itself to the carbon balance of its information system would overlook 89% of the impacts.