The Boom in Artificial Intelligence and Planetary Boundaries: Is AI Crossing the Line?

Executive Summary

Artificial intelligence is taking hold in organizations at an unprecedented pace, driven by massive investments. This boom faces a physical constraint: it generates additional demand for electricity at the very moment when the electrification of end-use applications is driving the green transition in all other sectors. It also leads to an increase in GHG emissions that threatens global climate goals.

This publication provides a trend-based projection[1] by 2030, the effects of the global increase in computing power on the availability of resources (electricity, water, raw materials) and on the global carbon budget. Based on an assessment of physical limits, it offers recommendations for reconciling AI and sustainability throughout the value chain. Significant points of friction are emerging:

• AI is likely to account for a major share of new global electricity capacity. By 2030, it is projected to consume approximately 815 TWh of electricity—2.1% of global consumption—but, more importantly, 13% of the total increase in electricity demand between 2025 and 2030. It is the share of growth—rather than the share of the total—that is at stake. Ireland offers a striking example: data centers there have gone from accounting for 5% of electricity consumption in 2015 to 23% in 2025, prompting the regulator to impose a connection moratorium in the Dublin region starting in 2021, which will be replaced by a self-generation requirement at the end of 2025.

• These volumes amount to 330 million tCO2e, which is roughly equivalent to France’s annual emissions in 2025. This is 2.5 times the carbon budget that the digital sector must adhere to through 2030 to remain on track to stay below 2°C.

• The immediate obstacle to AI development is not resources—it's the timeline. It takes at least five years to deploy new low-carbon electricity generation capacity. Caught up in this race, digital industry players are turning to whatever is immediately available: installing off-grid gas turbines directly attached to buildings and extending the operation of coal-fired power plants that were scheduled to close.

For France, the issue is not the immediate availability of electricity but the connection. Currently, 18 GW have been requested and partially reserved with RTE for future data centers, nearly 80% of which have been preempted by U.S. hyperscalers. More realistically, RTE’s central scenario projects an increase in demand of 15 to 20 TWh by 2030, representing 3% of national consumption. While this consumption is far from the 92 TWh exported in 2025, AI data center projects could nevertheless eventually come into direct competition with the electrification of French industry, transportation, and heating regarding the speed of grid connection.

• Taiwan, a critical upstream hub, is undermining the value chain. Electricity consumption attributable to AI could rise from 5 to 27 TWh between now and 2030, driven not by data centers but by semiconductor foundries that supply the world with AI chips. These 22 TWh represent nearly half of the projected additional consumption on the island, where 73% of the electricity mix would continue to come from fossil fuels.

• When it comes to water resources, the global average masks acute local pressures. AI accounts for 0.1% of global water consumption, but in the United States, according to the World Resources Institute, two-thirds of the data centers built or under development since 2022 are located in areas experiencing high water stress or severe drought. The risk is twofold: ensuring the operational continuity of these facilities and competition with agricultural and domestic water uses. Droughts in Taiwan are also threatening chip production.

The following courses of action appear necessary to reconcile AI with respect for planetary boundaries:

· From the users' perspective, combining minimal usage with frugal AI: Establish a "just-in-time" culture (AFNOR Spec 2314 standard, DINUM guidelines), and receive training ineco-prompting, limit the rebound effect and prioritize lightweight specialized models rather than oversized general-purpose LLMs for simple tasks.

· For corporate clients, examine usage patterns and prioritize them : Question the relevance of each AI integration, advocate for limiting the native inclusion of these tools in software, focus efforts on a limited number of high-value use cases, and prioritize smaller models when tasks allow.

· For industry stakeholders, maximizing technical efficiency and regional integration: Build data centers in cold-climate regions with a low-carbon electricity mix (Scandinavia, France), that can be connected to the power grid; recover waste heat from servers in high-density facilities; and widely adopt the algorithmic optimization techniques (mixture-of-experts, pruning, quantization) already used to reduce costs.

· From the regulator's perspective, establishing a regulatory framework and harmonized metrics: Enforce rigorous and harmonized carbon and environmental accounting standards, verified by independent, trusted third parties, to improve transparency in the sector and certify calculations of carbon footprints and avoided emissions (using the methodology NZI4IT).

· In terms of energy planning, accelerate the deployment of renewable energy (Renewable Energy) and low-carbon connections: A massive expansion of wind and solar power capacity is essential to avoid relying on fossil fuels as a backup, given the limited pace of nuclear capacity development. Data center operators must be encouraged to directly finance new local renewable energy capacity through long-term power purchase agreements (PPAs).

These levers do not act at the same pace. Policy levers can alter the trajectory by 2030; the expansion of low-carbon capacity, given the time required for deployment, will have most of its impact beyond that date. It is precisely this time lag that makes energy conservation and the prioritization of energy uses essential in the short term.

All modeling assumptions and sources are published in the appendix.