With our partner Salesforce, unify sales, marketing and customer service. Accele your growth!
Jean-Marc Jancovici asks the question without detour: What place can artificial intelligence occupy in a world that will soon or later have to decarbonate? The Shift Project has just published an intermediate report entitled “Artificial intelligence, data, calculations: what infrastructures in a decarbon world?” ” a document open to comments, which seeks to assess the energy and climatic impact of AI. Carbon 4, for its part, publishes an eloquently titled note: “Generative AI … Climate change! »»alerting to the unbearable trajectory of a technology as energy -consuming as it is an Opt.
The figures speak for themselves. In 2021, digital represented 4 % of global CO2 emissionseither the equivalent of emissions from all trucks and heavy utilities. Half comes from the electricity necessary for servers, networks and terminals, the other from the manufacture of components. In France, he weighs 4.4 % of the national carbon footprintwith annual growth of 6 % worldwide and 2 to 4 % in France. Each efficiency gain is immediately erased by the increase in uses.
Generative AI is no exception to the rule. Ten chatgpt requests per day generate 100 kg of co₂ per yearaccording to Carbon 4. In 2022, digital absorbed 10 % of global electricity productioneither Five times more electricity than total global consumption in 1945. A dynamic that questions: is AI an energy priority while other sectors-transport, agriculture, housing, industry-must also absorb limited resources to decarbonate?
In this context, three practical questions are emerging:
- Where are the figures? NVIDIA, which dominates 80 % of the ia flea marketdoes not communicate on the environmental footprint of its processors. TSMC, Openai, Anthropic, Microsoft and Google are hardly more transparent. However, in three years, the emissions of the cloud giants have jumped from 50 %.
- What arbitrations for the limited resources of tomorrow? Faced with competitive needs for electricity, materials and financing, Is AI a rational choice?
- What uses are really justifiable? Should we air transport, advertising and the entertainment industry Capture valuable calculation capacities while other sectors seek to reduce their carbon footprint?
Far from a technological rejection, this analysis invites to a structuring debate: What ends should the computer power of tomorrow be mobilized? Because in a constrained world, each Térawatt Heure must be justified.
An explosion of energy and financial demand
In 2022, the electrical consumption of data centers rose to 460 TWhmore than double the estimates of 2021 (200 TWh). By 2030, projections vary from 700 TWh at 2,100 Twh According to the scenarios. In comparison, the total annual consumption of France was established at 454 TWh in 2020.
The key factor in this increase is the boom in generative AI models, whose training mobilizes considerable resources. The GPT-3 model required 1,287 MWh for his initial training, the equivalent of Several hundred French households fed for a year. GPT-4 and its successors require even more computing power, lengthening the load on data centers over the sides.
This energy consumption translates directly into operational costs. On average, electricity represents 20 to 40 % of expenditure of a data center. A hyperscaler consuming 10 TWh per year spend 1 billion euros in electricityon the basis of an average price of € 100/mWh. With the increase in energy demand, data centers operators negotiate Electricity purchase contracts (PPA) On durations of 10 to 20 years, a strategy that lock prices but engage them on growing volumes.
Generative AI, engine of inflation of technological investments
The cost of infrastructure to support the development of AI reaches new levels. High performance (HPC) data centers require CAPEX (investment expenditure) greater than € 10,000/m²against 3,000 to 5,000 €/m² for a classic center.
In 2023, Microsoft spent $ 10 billion To support Openai and equip its Azure infrastructure in GPU Nvidia H100, whose unit price reaches € 35,000. Alphabet, Meta and Amazon followed similar trajectories, each investing more than $ 30 billion In their cloud and AI infrastructure since 2022.
The cost of semiconductors is another key factor. The demand in Gpu exploded prices:
- In 2020, a 1,000 GPU cluster cost approximately 30 million euros.
- In 2023, an equivalent cluster exceeded 50 million euros.
- By 2026, the rise in models could require clusters worth Over 100 million euros.
THE Cloud giants are today the only ones capable of absorbing these costs. Startups and companies seeking to develop their own models must Rent access to GPUsgenerating a increasing dependence on hyperscalers. Large model training on Microsoft Azure Or Google Cloud can cost up to 10 million eurosde facto excluding many market players.
The rise of data centers: an energy and financial challenge
The energy density of data centers grows quickly. Graphic processors (GPU), essential to the operation of AI models, see their electrical consumption multiplied. The previous generation displayed a Thermal Design Power (TDP) 250 Wagainst Over 1 kW today For the most advanced fleas.
In the United States, the rise in infrastructure pushes operators to diversify their sources of supply. Amazon, Microsoft and Oracle explore modular nuclear solutionswhile Meta announced the construction of three natural gas power plants dedicated to its data centers, totaling 2.3 GW of capacity. In the south-east of the country, 20 additional GW power plants are plannedinvolving programs of 80 MTCO2E per year.
In Europe, the situation is critical. In Irelanddata centers are already consuming more than 20 % of national electricityexceeding urban residential consumption. In Francethe electricity consumption of digital has reached 11 % of national demand in 2022 and could triple by 2040.
The construction of new centers requires massive investments. In 2024, Google announced 1 billion euros in investment in France For new AI data centers. The cost of a 1 GW center, capable of supporting the training of the biggest models, is estimated between 5 and 10 billion euros.
These infrastructures are funded by sovereign funds, venture capital and long-term debtsaccentuating the pressure on technological companies to generate Income via premium services. This economic model pushes to a Aggressive AI monetizationvia paid subscriptions and integration into SaaS offers.
What levers for an AI more financially and energetically sustainable?
Hyperscalers oligopoly strengthens economic asymmetry. Alternatives, such as the European federated cloud (Gaia-X), struggle to emerge for lack of funding. The EU could impose pooling obligations IA infrastructure.
Optimization of algorithms and decentralized training on lighter infrastructure (Edge Computing) remain under-exploited. Initiatives like Tinyml Or the Light Open-Source IA models could reduce pressure on data centers.
The explosion of electric demand pushes the cloud giants to conclude long -term electricity purchase contracts (PPA). Differentiated taxation according to the energy mix used could encourage greater transparency.
Investments in data centers are now motivated by long -term anticipation of profitabilitywithout certainty on the massive adoption of IA services. Financial regulation of the market could avoid a risky overinvestmentsimilar to the dot-com bubble of the 2000s.
AI, a strategic choice or an energy luxury?
The rise of generative artificial intelligence raises a fundamental question: In a world where energy and material resources will be increasingly limited, what priorities will we allocate our calculation capacity?
Should this power be focused on critical uses for ecological transition-optimization of electrical networks, reduction of industrial waste, improvement of the energy efficiency of buildings-or accept that an increasing share is used to improve advertising algorithms, conversational assistants or automated content flows?
If AI is a technological accelerator, it is also a revealer of our company choices. Are we ready to arbitrate rationally between innovation and sobriety, or are we going to let the market dynamic impose an energy-consuming trajectory of which we do not master the costs or the consequences?
