Barclays’ recent report projects that data-centre usage will likely triple due to the rapid AI growth by 2030. The firm’s estimates are notably higher than the current sustainable consensus, raising questions about the potential implications of AI development for global efforts to reduce greenhouse gas emissions.
Barclays’ recent Impact Series report titled “AI revolution: Meeting massive AI infrastructure demands” forecasts that AI growth could triple U.S. data-centre usage by the end of the decade, increasing energy consumption from 150-175 terawatt hours (TWh) in 2023 to 560 TWh, or 13% of current U.S. electricity demand.
The report further suggests that while efforts to improve energy efficiency may help reduce consumption, they are unlikely to fully offset the rapid growth of AI and its associated demands on power usage. Such a surge in infrastructure demands poses challenges for global net-zero targets. Therefore, Barclays emphasises the need for collaboration between tech companies, policymakers, and the energy sector to balance responsible AI advancements with sustainability goals.
Currently, data centres in different corners of the world jointly consume 1.0%-1.5% of global electricity, excluding the data on data centres used for mining cryptocurrencies.
AI requires processing power, storage, and connectivity of data centres to process and store vast amounts of data needed for tasks like machine learning, training models, and running complex algorithms. These centres provide the necessary infrastructure to handle the immense computing power required for AI operations, especially as AI applications rapidly grow in scale and complexity.
Notably, the predictions of future AI power requirements may not be too accurate, as current AI applications are mostly generative ones, while a recent Citi GPS report reveals a significant shift of venture capital funding towards autonomous AI startups developing the game-changing agentic AI technology.
Agentic AI performs tasks autonomously, making decisions and executing actions in real-time. Such activity demands continuous processing and data handling, unlike generative AI, which is often used for specific temporary tasks like generating content based on input, which doesn’t require the same level of ongoing real-time interaction or decision-making, thus generally consuming less computing power compared to agentic systems.
