Recent analysis suggests that energy consumption by artificial intelligence (AI) could rival, and even exceed, that of Bitcoin mining. Projections indicate that AI may account for nearly half of the global electricity usage in data centers by the close of 2025.
These predictions come from Alex de Vries-Gao, a PhD candidate at the Vrije Universiteit Amsterdam’s Institute for Environmental Studies. He is known for his work tracking the electricity consumption and environmental impact of cryptocurrencies on his website Digiconomist. His latest insights into AI’s escalating electricity demands were shared last week in the journal Joule.
According to de Vries-Gao, AI currently consumes about 20% of the electricity used by data centers. Accurate measurements are difficult without major tech companies disclosing specific energy usage for their AI models. His estimates derive from analyzing the supply chain for specialized computer chips utilized in AI. Despite increasing efficiency, researchers, including de Vries-Gao, have noted a rapid rise in energy demand for AI technologies.
“Oh boy, here we go.”
Previously, with alternative cryptocurrencies like Ethereum transitioning to less energy-intensive practices, de Vries-Gao thought the demand for analysis might decrease. However, the emergence of technologies like ChatGPT prompted him to reconsider, expressing concern over the energy usage associated with high-performance AI in competitive markets.
He identifies a prevailing industry trend: the belief that “bigger is better.” This fixation has led major tech companies to continually expand their models, thereby increasing their resource requirements. The subsequent surge in new AI data centers, particularly in the United States, is notable. Energy providers are planning to construct new gas-fired power plants and nuclear facilities to satisfy the increased electricity demand stemming from AI operations, echoing challenges faced by new cryptocurrency mining facilities.
De Vries-Gao also points out similarities with his work on cryptocurrency, noting the difficulties in accurately assessing the energy consumption and environmental implications of these technologies. While many leading tech companies have established climate goals and report greenhouse gas emissions, the specific contributions of AI are often not disclosed.
To analyze these trends, de Vries-Gao employed a “triangulation” method utilizing publicly available hardware details, industry estimates, and insights from corporate earnings calls. For instance, Taiwan Semiconductor Manufacturing Company (TSMC) is expected to more than double its production capacity for AI chips from 2023 to 2024.
His comparisons suggest that last year, AI hardware consumed as much electricity as the entire Netherlands, with estimates indicating a rise to a consumption level equivalent to that of the United Kingdom by the end of 2025, with power needs predicted to reach 23GW.
In a separate report released last week, consulting firm ICF projected a 25% increase in electricity demand across the United States by 2030, primarily driven by AI, traditional data centers, and Bitcoin mining.
Identifying a clear trajectory for AI’s energy consumption and its environmental ramifications remains a challenge. An article in last week’s MIT Technology Review highlighted the complexities, noting that energy usage varies based on numerous factors, including the nature of the AI queries and the availability of renewable energy in local power grids.
It’s a mystery that could be solved if tech companies were more transparent
Greater transparency from tech companies regarding AI in their sustainability reports could help clarify these issues. “The extensive steps required to generate any meaningful data is frankly absurd,” de Vries-Gao stated, underscoring the need for improved clarity in reporting.
Looking ahead, uncertainties persist regarding whether advances in energy efficiency will stabilize electricity demands. A recent statement from DeepSeek indicated its AI model operates at a fraction of the power utilized by Meta’s Llama 3.1, sparking debates about the necessity of high energy consumption for technological progress. The ongoing question remains whether companies will shift towards efficient models or continue with their existing methods that emphasize increased data and computing power.
The transition of Ethereum to a significantly more energy-efficient transaction validation method highlights the potential for dramatic reductions in electricity usage, as illustrated by its 99.988% decrease in consumption. While environmental advocates push for similar changes in other blockchain technologies, certain sectors, notably Bitcoin miners, remain hesitant to pivot away from their current investments and ideological stances.
Concerns about Jevons paradox also loom over AI; even with more efficient models, the overall electricity demand may rise as increased usage of the technology could offset efficiency gains. A thorough measurement and analysis of AI’s impact on energy consumption will be essential to manage these challenges effectively.
