AI and Electricity: A System Dynamics Approach - Explained (2/10) - Schools Of Thought

Welcome to our second series depicting 'AI and Electricity: A System Dynamics Approach'.

After having discussed last week the problem statement, we will today examine existing methods for forecasting electricity usage, identifying their weaknesses and limitations, and describing the genesis of the Schools of Thought of AI.

I hope you enjoy it and look forward to fruitful discussions!

What Are The Existing Forecasts Of Electricity Usage?

The emerging field of AI electricity footprint estimation faces complex challenges, notably lacking comprehensive scenarios and forecasts. Since 2023, a surge in research has revealed divergent modeling methodologies. Additionally, there are varying perceptions of AI’s future (1). These differing perspectives, driven by methodological disagreements and data gaps, are shaping the discipline’s trajectory.

To address these issues, we first examine the strengths and weaknesses of bottom-up and top-down approaches. Second, recognizing the potential unreliability of relying on a single projection for long-term future impacts, we approach insights into the future by simulating multiple scenarios based on fundamental AI Schools of Thought. Understanding and addressing these emerging schools of thought is crucial for reliably assessing and managing AI’s future electricity use (2).

Assessing The Strengths And Weaknesses Of Bottom-up And Top-down Approaches

In a recent article, Masanet et al. propose a modern bottom-up approach to estimating AI electricity use, arguing that despite higher power requirements, AI data centers can be modeled similarly to conventional ones (3). This method enables scenario analysis of future electricity demand based on various factors. The authors identify several “traps” that often lead to overestimations in AI data center electricity consumption, such as multiplying AI server rated power by sales data, using advertised power capacities and assumed Power Usage Effectiveness (PUE) values, and basing projections on utility permits or assumed growth rates. To implement this approach, various data points are crucial, including AI hardware power profiles, data center reporting, cooling technology performance data, and market evolution. Masanet et al. identify shortcuts that fail to account for discrepancies between rated and actual power use, unclear definitions, non-representative cooling approaches, and overprovisioning. While this bottom-up approach is appealing, it requires significant alignment across stakeholders to provide critical data, positioning it as a structural and long-term method for forecasting AI electricity use.

In contrast, De Vries’ study employs a top-down approach, utilizing industry-level assumptions and trends to project future AI energy consumption (2). This methodology relies on broad statistics and simplified parameters, and using market leader NVIDIA data for information on AI server energy consumption. While this approach allows for high-level estimates and scenario-based analysis, it has limitations due to the lack of detailed data from individual AI systems or companies (4). Additionally, the assumption of servers operating continuously at full capacity may overestimate energy consumption. However, despite these shortcomings, the study serves as an important starting point for discussing and addressing the potential environmental impact of AI’s rapid growth (5).

Both approaches highlight the need for more granular data and transparency from AI developers to improve the accuracy of energy consumption projections (6). Hence, as the field of AI continues to evolve rapidly, a combination of bottom-up and top-down approaches, along with increased data sharing and collaboration among stakeholders, may provide the most comprehensive understanding of AI’s energy footprint (7).

This methodological reflection isn't novel. The balanced approach required for AI energy consumption forecasting echoes Olaf Helmer's (1910-2011) (8) pioneering work in futures studies. His Delphi method (9), which blends expert opinions with quantitative analysis, offers valuable insights for addressing complex forecasting challenges. By mitigating forecasting biases, this methodology demonstrates conceptual relevance to AI projections (10) . Helmer's emphasis on systematic analysis, interdisciplinary insights, and long-term perspectives provides a robust theoretical foundation (11), adaptable to rapidly evolving fields like AI, balancing data-driven analysis with expert judgment in novel domains.

Understanding Representations Of The Future Is Essential For Accurate Modeling Of Electricity Use Projections

Most existing projections, grounded in data, often align with a future-specific representation, which can be shaped by researcher’s positionality statements or stakeholder priorities, which significantly influence data interpretation and conclusions (12). These projections typically blend elements of rationalism (bottom-up or top-down data, assumptions, and models to form a logical framework) and subjectivism (cultural biases, litanies, worldviews, and metaphors) that shape interpretations and projections (13). As Ekchajzer et al. (14) caution, we must be wary of overly simplistic, narrow, and static future projections. This is particularly relevant as we witness the emergence of organized thought in the landscape of AI projections, with contributions from a diverse range of entities (15) , including energy analysts, banks, consulting firms, foundations, IT associations, federations, and public institutions.

At the most fundamental level, these projections are driven by individuals whose ideas are disseminated exponentially through journals, articles, books, and social networks. Over time, these intellectual efforts are amplified by organizations, giving rise to distinct practices, movements, and beliefs that reflect collective perspectives. As these practices coalesce, they contribute to deeper intellectual currents that have started to shape the debates on AI trajectory, especially in terms of infrastructural footprint.

Quite naturally, these converging forces are yet forming - explicitly or not - distinct Schools of Thought, still taking shape within an information-saturated debate, each offering unique perspectives on the future of AI. Ultimately, these emerging Schools of Thought will profoundly shape the development of AI. To foster debate and analysis, a plurality of perspectives is essential. While each school of thought has its limitations, their collective insights offer a broader spectrum for thinking AI electricity future (16). These diverse viewpoints can inform more robust scenario planning by discussing assumptions, identifying blind spots in forecasting, and stimulating constructive debates on the multifaceted factors influencing AI electricity use.

System Dynamics For Thinking AI And Electricity Scenarios

The Schneider Electric Sustainability Research Institute approach utilizes System Dynamics modelling (17) to explore four scenarios derived from distinct Schools of Thought. While many existing studies adhere to a school of thought and forecasting approach (either bottom-up or top-down) (18), our research suggests that a combination of these methods can forge a more nuanced and comprehensive vision of potential futures (19). Our approach not only stimulates critical thinking but also equips stakeholders to navigate a broader spectrum of scenarios (20). To ground our work in solid method, we collaborate with Professor Fons Wijnhoven, a recognized specialist in System Thinking, to enhance the robustness of the underlying theory for our system dynamics modeling. To this end, we have begun to qualitatively define four scenarios. These scenarios will then undergo further qualitative development, followed by translation into causal loop diagrams, and finally, into comprehensive system dynamics models.

We begin by examining the Sustainable AI School of Thought, which advocates for AI development that prioritizes sustainable practices within planetary boundaries. This school envisions a future where AI advances harmoniously with environmental stewardship.

The second, the Limits to Growth School, emphasizes the constraints on AI development, including power constraints, supply chain tensions, and material scarcity. This scenario echoes the approach of Dennis Meadows' seminal "Limits to Growth" study from 1972.

The third school, the Abundance Without Boundaries School, believes in technology's ability to overcome challenges but risks leading to unchecked AI growth, potentially concentrating power and exacerbating inequalities. This scenario can be perceived as an extreme case, embedded with risks, but also as a nascent reality in some regions of the world, necessitating careful and data-grounded examination.

Finally, the Energy Crisis School warns of the potential negative impacts of ungoverned AI development, particularly due to electricity supply limitations. This school emphasizes the principle that pushing boundaries too far can trigger disruptions in the ecosystem, such as black swan events with high risks.

While the first two scenarios may be closer to the most likely future evolution, the third and fourth scenarios push the boundaries and highlight potential risks that require careful consideration. In the next series, we will delve deeper into the qualitative description of the four Schools of Thought.

This concludes our second series. I hope you enjoyed it! Let's catch up in a few days to delve deeper into the fundamentals of the Schools of Thought and how we have translated qualitative scenarios into system mechanisms for future modeling.

I look forward to your reactions for further conversations.

References

1. Castro, D. (2024, January 29). Rethinking concerns about AI's energy use. Center for Data Innovation. https://www2.datainnovation.org/2024-ai-energy-use.pdf
2. de Vries, A. (2023). The growing energy footprint of artificial intelligence. Joule, 7(10), 2191–2194. https://doi.org/10.1016/j.joule.2023.09.004
3. Masanet, E., Lei, N., & Koomey, J. (2024). How will the electricity use of AI data centers evolve? To answer this question, energy analysts need better data. https://doi.org/10.13140/RG.2.2.11203.00801
4. Schwartz, R., Dodge, J., Smith, N. A., & Etzioni, O. (2020). Green AI. Communications of the ACM, 63(12), 54-63. https://doi.org/10.1145/3381831
5. Rahman-Jones, I. (2024). Google’s greenhouse gas emissions up 48% in 5 years. BBC News. https://www.bbc.com/news/technology-12345678
6. Google. (2024). 2024 Environmental Report. https://www.gstatic.com/gumdrop/sustainability/google-2024-environmental-report.pdf
7. Hintemann, R., & Hinterholzer, S. (2022). Energy consumption of data centers worldwide. Borderstep Institute for Innovation and Sustainability. https://www.borderstep.de/publikation/hintemann-r-hinterholzer-s-2022-data-centers-2021-cloud-computing-drives-the-growth-of-the-data-center-industry-and-its-energy-consumption-berlin-borderstep-institute/
8. Helmer, O. (n.d.). In Wikipedia. Retrieved December 9, 2024, from https://en.wikipedia.org/wiki/Olaf_Helmer
9. Delphi method. (n.d.). In Wikipedia. Retrieved December 9, 2024, from https://en.wikipedia.org/wiki/Delphi_method
10. Helmer, O., & Rescher, N. (1967). Analysis of the future: The Delphi method. RAND Corporation. https://www.rand.org/pubs/papers/P3558.html
11. Helmer, O. (1977). The utility of long-term forecasting (Professional Paper PP-77-005). International Institute for Applied Systems Analysis. https://pure.iiasa.ac.at/id/eprint/750/7/PP-77-005.pdf
12. Strubell, E., Ganesh, A., & McCallum, A. (2019). Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (pp. 3645-3650). https://arxiv.org/abs/1906.02243
13. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (pp. 610-623). https://doi.org/10.1145/3442188.3445922
14. Ekchajzer, D., Hilty, L. M., Kern, E., & Penzenstadler, B. (2024). Decision-making under environmental complexity: The need for moving from avoided impacts of ICT solutions to systems thinking approaches. In ICT4S 2024: International Conference on ICT for Sustainability. https://hal-lara.archives-ouvertes.fr/LITEM/hal-04637677v1
15. Lacoste, A., Luccioni, A., Schmidt, V., & Dandres, T. (2019). Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700. https://arxiv.org/abs/1910.09700
16. Milmo, D. (2024). Can the climate survive the insatiable energy demands of the AI arms race? The Guardian. https://www.theguardian.com/business/article/2024/jul/04/can-the-climate-survive-the-insatiable-energy-demands-of-the-ai-arms-race
17. Forrester, J. W. (1994). System dynamics, systems thinking, and soft OR. System Dynamics Review, 10(203), 245-256. https://doi.org/10.1002/sdr.4260100211
18. Sterman, J. D. (2000). Business dynamics: Systems thinking and modeling for a complex world. McGraw-Hill Education. https://www.mheducation.com/highered/product/business-dynamics-sterman/M9780072317202.html
19. Petit, V. (2021). Digital economy and climate impact. https://www.se.com/ww/en/insights/electricity-4-0/digital-for-recovery/digital-economy-and-climate-impact.jsp
20. Kerr, D. (2024, July 12). AI brings soaring emissions for Google and Microsoft, a major contributor to climate change. NPR. https://www.npr.org/2024/07/12/g-s1-9545/ai-brings-soaring-emissions-for-google-and-microsoft-a-major-contributor-to-climate-change