What is it about?
AI is growing fast and its share in energy demand is constantly increasing. Without major changes to smarter AI deployment and energy policies, AI could have a massive share in global energy demand and carbon emissions by 2050.
Featured Image
Photo by Omar:. Lopez-Rincon on Unsplash
Why is it important?
This work provides the first comprehensive, long-term global projection of AI’s environmental footprint through 2050. Prior to this work, most research focused on isolated, case-specific benchmarks, such as the energy cost of training a single model like BERT or GPT-3. These anecdotal figures did not generalize across different model types or the massive scale of global deployment. This paper moves beyond individual models to a systematic, "scenario-based analysis" of the entire global AI sector. The researchers developed a novel modeling framework that is the first to jointly account for three critical dimensions: . AI Deployment Trajectories: Projections of how AI will be adopted and used globally. . Evolving Electricity Systems: How different regions' energy mixes (e.g., coal vs. renewables) will change over time. . Supply-Chain Emissions: The "upstream" carbon cost of manufacturing and transporting AI hardware, like GPUs. A key contribution of the work is the realization that technical efficiency is not a "silver bullet." While hardware and algorithms are becoming more efficient, the study finds that the aggregate demand still increases by roughly an order of magnitude due to the sheer growth of AI workloads. Even under optimistic efficiency scenarios, AI energy demand could remain six times higher than 2024 levels. The study demonstrates that the environmental future of AI depends heavily on deployment strategy and decarbonization policy.
Perspectives
This work highlights the need to analyze the impact of AI adoption by different sectors and also the impact of this on the global energy landscape and Carbon emissions. The growth and adoption trajectories of AI under different scenarios are considered (Baseline: continuation of the current trend, More Smaller: wide use of small and sector specific AI solutions, Fewer Larger: only a few AI systems racing towards AGI), hardware and software efficiency gains, energy policies and targets by different countries. Possible AI adoption and energy policy trajectories are analyzed under different scenarios and the results are striking: Massive Growth in Electricity Demand: By 2050, AI's share of global electricity consumption is projected to rise to between 13% and 47%, depending on growth and deployment patterns. Emissions Peak: Under current trends, AI-related emissions could exceed 8 gigatons of CO2-equivalent annually by mid-century. The "Efficiency Gap": While improvements in hardware and algorithms substantially reduce energy use per operation, total AI energy demand is still expected to increase by an order of magnitude due to the explosive growth in training and inference workloads. Inference vs. Training: The study projects that inference (usage) will eventually surpass training to dominate total computational demand as AI becomes more integrated into daily life. Policy and Strategy Impact: Scenarios that combine the use of fewer, larger models with a transition to renewable energy (Energy Target scenarios) can reduce total emissions by up to 40% compared to business-as-usual paths. Lifecycle Contributions: The research highlights that the carbon footprint is not just from electricity; the manufacturing of GPUs and their international transportation are significant "upstream" contributors to the total footprint. Regional Disparities: The carbon intensity of AI is heavily dependent on location. For instance, training the same model in coal-heavy regions can produce double the emissions compared to regions like Norway or France that use low-carbon energy.
Metin Turkay
Koc Universitesi
Read the Original
This page is a summary of: Scenario-based forecasting of the global energy demand and carbon footprint of artificial intelligence, PLOS One, March 2026, PLOS,
DOI: 10.1371/journal.pone.0343056.
You can read the full text:
Contributors
The following have contributed to this page
