Explaining Variation in Energy Forecasts: The Role of AI and Data Centers

Written by: Owen Casey

Edited by: Ashley Yeung

Underlying the dazzling achievements in Artificial Intelligence is a more concrete reality: energy. The close connection between AI and energy stems in large part from the role of AI data centers. Data centers are vital for the successful operation of AI models as they hold the “computing infrastructure used to process, manage, and store large volumes of digital data” (Hooper and Hayden, 2025). The data and computing infrastructure stored in these centers allow for the large scale data ingestion that current Large Language Models (LLMs) rely on, making them vital to current AI operations. Data center development has already spiked in recent years, and firms have committed extraordinary levels of capital to data center projects. Amazon, Meta, Google, and Microsoft are projected to spend US$750 billion on AI-related capital expenditures, such as data centers, in the next two years (Milmo, 2025). Labelling it the “next industrial revolution,” Brookfield estimated that over the next nine years, US$2 trillion of the US$7 trillion spent on AI-related infrastructure will be directly spent on data centers (Brookfield, 2025). The massive amounts of investment into AI has turned data centers into a central component of future energy demand.

For energy forecasters, AI data center development has introduced a new element of volatility and uncertainty. Although there is consensus that AI development is set to dramatically increase energy demand, forecasting the exact amount of growth has become increasingly disputed. On the high end of predictions, the Boston Consulting Group predicts that U.S. energy use will increase to 1050 TWh/year by 2030 (Muir, 2025). McKinsey estimates that power centers in the U.S. will need 606 TWh by 2030 (McKinsey & Company, 2024), meanwhile the International Energy Agency forecasts that data centers will require just over 400 TWh per year by 2030 (Muir, 2025). Hence, there exist serious discrepancies in the projections of future energy demands. Energy projections differ due to a variety of different factors, however, AI data centers have undoubtedly become a central reason for the current disagreements. This article will examine how the development of AI data centers is contributing to the variation in energy demand projections.

Importance of Energy Forecasts

Before examining how AI data centers are complicating energy forecasts, one ought to understand the importance of accurate energy forecasts. Forecasts can overestimate or underestimate actual energy demand, both of which have harmful consequences. For one, overestimations of future energy demand lead to the construction of energy infrastructure that becomes underutilized. This can create stranded assets, which are assets that degenerate into liabilities because of unexpected low use (Grantham Research Institute, 2022). Utilities often shift the financial burden of stranded energy infrastructure to consumers by raising consumer electricity prices. For example, during the late 2010s, before the AI boom, the U.S. utility generating capacity exceeded required energy reserve margins by 30%, thereby costing customers billions per year (Dyson and Engel, 2017). Additionally, inaccurate energy demand forecasts can lead to higher spot prices, which are the market price of energy sold for immediate use. If high spot prices are sustained and volatile, it can lead to higher costs for households and firms (Goodarzi et al., 2019). Renewable energy development is also harmed by imprecise energy forecasts: overestimating energy demand leads to a surplus of energy supply, which will lower energy prices, and therefore, decrease revenues. The financial loss incurred by overestimations thus deters investors from expanding renewable energy capacity, given the increased risk of lower or negative returns. Conversely, underestimations can lead to the excess use of renewable energy infrastructure, which accelerates the infrastructure’s deterioration (Ugbehe et al., 2025). Given the harmful consequences of inaccurate energy demand forecasts, the current variation in energy forecasts is cause for concern and warrants examination.

AI Bubble

One of the driving factors behind the current discordance over future energy demand is due to the preoccupation over the potential consequences of an “AI bubble”. An economic bubble occurs whenever a particular industry or technology attracts significant investment and inflated expectations, causing a disconnect between its economic fundamentals and the high valuations. When a bubble bursts, there can be minor or major market corrections in which people rush to withdraw capital, leading to a drop in asset values. Anxiety over the existence of an AI bubble stems from a variety of factors, however, there are two primary reasons behind such fears. Firstly, there are concerns over the limitations caused by energy supply and the likelihood of energy availability serving as a bottleneck for AI development. While there is disagreement over exactly how much energy AI will demand, there is a broad consensus that the current energy generation trajectory will not be able to match it. It is predicted that 44GW of power will be required by new data centers by 2028 (Rosner-Uddin et al., 2025). However, only 25GW of new power is predicted to come online in the next three years, meaning that the current approach will lead to a 19GW deficit (Rosner-Uddin et al., 2025). If this bottleneck materializes, current and planned AI investments could be put at risk, forcing firms to scale back or abandon projects. In such a scenario, a lack of energy availability could deter the development of the AI industry, causing a burst of the AI Bubble. The resulting decline in AI development could significantly lessen energy demand relative to many current forecasts.

Another major reason behind fears of an AI Bubble is the doubt over AI’s broad adoption across business and society. If AI doesn’t allow companies to achieve sufficient gains in productivity and profits, there will not be enough demand to support the wide-scale infrastructure development that it is currently thought to warrant. This would translate into a lower energy demand than many are currently forecasting. As the Swiss Institute of AI writes, “[Energy] demand will depend on how quickly models are used, business adoption, and profit margins that justify ongoing expansion. If…the AI bubble deflates, [energy] projections that assume steady increases will be off target” (O’Neill, 2025).

Critics and analysts point to several indicators to support their concerns. In a recent survey of 4,454 CEOs, over half admitted that their companies have not seen any financial return from investments into AI (PwC, 2026). Furthermore, transactions related to AI have mainly been circular or within the AI industry rather than from a broader range of users. For example, the company Nvidia, which makes the silicon chips and software packages critical to AI systems, agreed to annually invest US$10 billion into OpenAI, who will mainly use the money to buy Nvidia’s chips. Transactions such as these, in which AI companies buy or lend among themselves, constitute a significant portion of current AI business. These circular transactions obscure actual societal demand for AI services. If AI doesn't diffuse across society and the wider business landscape, there could be significant investment rollback and a corresponding drop in energy demand.

Importantly, the bursting of an AI Bubble by any of the aforementioned reasons could result in a range of outcomes with different levels of severity. If the bubble is major, there could be massive devaluation of AI assets and a huge rollback in data center development. Alternatively, the bubble may only result in a minor correction where there is a slight downturn but an eventual recovery. There is also the chance that there is no AI Bubble, and all of the investment will be justified. The wide range of potential outcomes and the uncertainty over their fruition is a large reason why energy forecasting is currently so conflicted.

Phantom Data Centers

Another AI-related factor contributing to the disparities in energy forecasts is the onset of “Phantom Data Centers”. This phenomenon arises when developers approach multiple different utilities with the same project in order to find the lowest-priced power to connect their data center to (Muir, 2025). The quest for the lowest prices results in multiple utilities counting the same project in their demand projections, despite only one project actually being developed. Hence, a sudden rise in phantom data centers can lead to an overestimation of energy demand. It is for this reason that multiple American utilities have recently reduced their data center-related projects (Muir, 2025). Energy forecasters will have to learn to differentiate between what data center projects are legitimate rather than just speculative and account for this in their projections.

Energy Efficiency

An additional way in which AI complicates energy demand forecasting is the uncertainty concerning potential advancements in energy efficiency. Currently, AI requires extraordinarily high levels of energy compared to other internet search functions. It is estimated that a ChatGPT query uses around 10 times more energy than a Google search (Towler, 2025). Nonetheless, given the historical progress in energy efficiency and the financial and climate incentives, improvements in AI energy efficiency are likely to occur. Josh Parker, a director of sustainability at Nvidia, notes that each new generation of AI GPU chips significantly improves performance and energy efficiency (Parker, 2025). The latest GPUs provide thirty times more computational performance with a twenty-five-fold increase in energy efficiency compared to those from two years ago (Parker, 2025). One study found that if AI could improve its efficiency at one-tenth the rate of its adoption, it would compensate for the additional energy required by data centers (Bernard et al., 2025). Improvements in AI energy efficiency, therefore, complicate energy demand forecasts because future electricity consumption will largely rest on the scale of efficiency gains.

Ironically, one potential path toward improved energy efficiency may come from AI itself.  AI’s ability to consume and analyze vast amounts of data enables more accurate energy demand forecasting and more efficient automation of energy distribution. For example, in commercial buildings, where 30% of energy used is wasted, AI could help reduce energy consumption by 25% (Parker, 2025). However, although AI may improve energy efficiency, it does not necessarily mean that total energy demand will decrease. The Jevons Paradox suggests that advancements in the efficient use of a resource often lead to an increase in the actual use of the resource, given its lower cost. Therefore, it is difficult for forecasters to predict whether energy efficiency driven by AI will increase, decrease, or maintain energy demand.

Conclusion

Energy demand forecasting has been complicated by the onset of AI. Forecasts have estimated vastly different levels of future energy demand due to several reasons. For one, due to the extraordinary investment into AI, there is significant concern over the potential consequences of an AI Bubble. If the bubble bursts, whether due to an energy bottleneck, a shortage in the business adoption of the technology, or another reason, there will be significant consequences for energy demand. Additionally, the rise in Phantom Data Centers has further muddied the job of forecasters. Finally, uncertainty regarding developments in energy efficiency adds another layer of complexity to energy forecasting. These ambiguities are just a few of the many variables that complicate energy demand forecasts. Given the economic and environmental stakes of inaccurate forecasting, it is essential that policy makers, utilities, and investors continue to account for all potential sources of forecasting error.

References  

“29th Global CEO Survey | PWC.” PwC, January 19, 2026. https://www.pwc.com/gx/en/issues/c-suite-insights/ceo-survey.html.

“AI’s Power Binge.” McKinsey & Company, November 6, 2024. https://www.mckinsey.com/featured-insights/week-in-charts/ais-power-binge.

Bernard, Robert N., Renate de Lange , Sammy Lakshmanan, and Scott Likens. “Could Net-Zero AI Become a Reality? | PWC.” PwC, April 29, 2025. https://www.pwc.com/gx/en/issues/value-in-motion/ai-energy-consumption-net-zero.html. 

“Building the Backbone of Ai.” Brookfield , August 2025. https://www.brookfield.com/sites/default/files/documents/Brookfield_Building_the_Backbone_of_AI.pdf.  

David, O’Neill. “Rethinking AI Energy Demand: Planning for Power in a Bubble-Prone Boom.” Swiss Institute of Artificial Intelligence, November 3, 2025. https://siai.org/memo/2025/11/202511282776?utm_source=chatgpt.com. 

Dyson, Mark, and Alex Engel. “The Billion-Dollar Costs of Forecasting Electricity Demand.” RMI, October 23, 2017. https://rmi.org/billion-dollar-costs-forecasting-electricity-demand/?utm_source.  

Goodarzi, Shadi, H. Niles Perera, and Derek Bunn. “The Impact of Renewable Energy Forecast Errors on Imbalance Volumes and Electricity Spot Prices.” Energy Policy 134 (November 2019): 110827. https://doi.org/10.1016/j.enpol.2019.06.035. 

Hooper, Alex, and Hayden Toohey . “Estimating Data Centre ‘Phantom Demand.’” Oxford Economics, November 21, 2025. https://www.oxfordeconomics.com/resource/estimating-data-centre-phantom-demand/.  

Milmo, Dan. “Boom or Bubble? Inside the $3tn AI Datacentre Spending Spree.” The Guardian, November 2, 2025. https://www.theguardian.com/technology/2025/nov/02/global-datacentre-boom-investment-debt.  

Muir, Martha. “‘Phantom’ Data Centres Muddy Forecasts for US Power Needs.” Financial Times, November 13, 2025. https://www.ft.com/content/331f8e5c-a813-48d4-9af6-806c8482eede.   

Parker, Josh. “Busting the Top Myths about AI and Energy Efficiency .” Atlantic Council, February 20, 2025. https://www.atlanticcouncil.org/content-series/global-energy-agenda/busting-the-top-myths-about-ai-and-energy-efficiency/. 

Rosner-Uddin, Rafe, Nassos Stylianou, Caroline Nevitt, and Jamie Smyth. Subscribe to the Financial Times, December 8, 2025. https://ig.ft.com/ai-power/.  

Towler, Louise. “Search Engines vs AI: Energy Consumption Compared.” Kanoppi, February 17, 2025. https://kanoppi.co/search-engines-vs-ai-energy-consumption-compared/.   

Ugbehe, Prosper O., Ogheneruona E. Diemuodeke, and Daniel O. Aikhuele. “Electricity Demand Forecasting Methodologies and Applications: A Review.” Sustainable Energy Research 12, no. 1 (April 24, 2025). https://doi.org/10.1186/s40807-025-00149-z.  

“What Are Stranded Assets?” Grantham Research Institute on Climate Change and the Environment, July 27, 2022. https://www.lse.ac.uk/granthaminstitute/explainers/what-are-stranded-assets/.