02
AI is unlikely to lead to very high energy demand

02 AI is unlikely to lead to very high electricity demand

Introduction

There is a lot of hype about the forecasted signiﬁcant electricity demand for AI
We believe that there will be increases, but most modelling ignores the energy reduction that will come from improvements to technology. Therefore, we believe current forecasts are exaggerated

The world has seen many predictions of high electricity demand, from the rise of the internet to the move to cloud and now Artiﬁcial Intelligence (AI). Yet the reality is that electricity demand in Europe at least has been falling in the last few years. In addition, almost every energy forecast and prediction of the past 20 years has proven to be wrong.

Why has there been so much hype about AI driving massive electricity demand, particularly in the US? Is it real or are the chip makers fuelling the hype? There is certainly a lot of data centre discussion in the US. At the same time the pace of innovation in AI, chip design and processer architectures and systems, is palpable.

This article explores the latest forecasts in AI electricity demand. It then goes on to contemplate Koomey’s law and the latest innovations. Finally, we examine the existing AI use cases driving down demand and a potential tipping point, before reaching conclusions.

Dan Jeavons

VP Computational Science & Digital Innovation, Shell

Dan Jeavons is an industry thought leader and experienced energy executive who previously worked at both Accenture and at Shell. At Shell, he most recently held accountability for computational science, data science and AI globally. His industry leading work in digital transformation has been recognised by his listing in the BT150 and his work has been referenced by the Harvard Business Review, Forbes and the Financial Times.

James Forrest

EVP: Global Industry Leader for Intelligent Industry, Capgemini

James led Capgemini’s global business for energy and utilities for five years. He has three decades of experience in energy across retail, generation, operations, and the energy transition. He has lived and worked in Europe, Asia Paciﬁc and North America.

02 AI is unlikely to lead to very high electricity demand

There is a lot of variation in current forecasts

35% global increase between now and 2030 (IEA) to 250% growth in the US alone

Forecasts for the impact of AI on global-power demand vary widely, reﬂecting deep uncertainty about the pace and scale of AI adoption. Conservative estimates, such as those from the International Energy Agency (IEA), suggest a modest global increase in power demand of 35% to 100% by 2030. Similarly, the U.S. Federal Energy Regulatory Commission (FERC) projects a relatively restrained growth in US data center power consumption, estimating an increase of 21GW to 35GW by 2030.

In contrast, several private sector analysts predict much steeper rises.

Boston Consulting Group anticipates US data center power demand will triple by 2030, while Goldman Sachs expects it to reach 8% of total US power use, necessitating 47GW of new supply. The Electric Power Research Institute estimates that data centre energy demand could be up to 9% of the US energy capacity by 2030. McKinsey estimates 11-12% by 2030, up from 3-4% currently.

Semi Analysis bases its projections on manufacturer shipment forecasts, particularly from Nvidia. Their model suggests that the three million AI accelerators expected in 2024 alone could require 4.2GW of power. Their base case sees global data center demand tripling to 1,500 terawatt-hours by 2030, rising from 1.5% to 4.5% of global electricity use and necessitating 100GW of new power infrastructure.

02 AI is unlikely to lead to very high electricity demand

Koomey’s law is likely to have

an impact

Much of the current analysis does not account for Koomey’s law

Koomey’s Law describes a long-standing trend in computing hardware: for roughly ﬁfty years, the number of computations achievable per joule of energy has doubled approximately every 18 months. This principle, identiﬁed by Professor Jonathan Koomey in a 2010 paper, implies that for a ﬁxed computing workload, the energy—or battery—required would halve every year and a half.

However, this increase in energy eﬃciency may be oﬀset by the Jevons Paradox. This phenomenon suggests that as technological improvements make a resource more eﬃcient and cheaper to use, overall consumption of that resource can actually increase due to rising demand. In the context of computing, even as devices become more energy-eﬃcient, the growing appetite for computational power—especially driven by AI and data-intensive applications—could lead to greater total energy use.

It’s also important to note that Koomey’s Law was originally based on Central Processing Units (CPUs), which perform one calculation at a time. Today’s cutting-edge processors are Graphics Processing Units (GPUs), which are not only larger but also function as complex systems capable of handling many tasks in parallel. This shift raises questions about whether Koomey’s Law still applies in the same way, and whether the Jevons eﬀect might dominate, driving energy demand upward despite gains in eﬃciency.

02 AI is unlikely to lead to very high electricity demand

Innovation is moving at a great pace

A GPU is not just a single chip—it’s a sophisticated system comprising multiple processors, short-term caches, long-term memory, communication switches, specialized functional modules, and task controllers that coordinate operations across all components. Performance and eﬃciency gains can be achieved not only within each of these elements but also through their integration and the optimization of the algorithms they run.

Nvidia CEO Jensen Huang highlighted this synergy with a striking example: training a 1.8 trillion-parameter AI model using the new Blackwell GPUs required just 4 megawatts of power, compared to 15 megawatts using the previous Hopper architecture—a 75% reduction in energy use, despite each Blackwell GPU consuming 70% more power individually.

These massive improvements in chip performance have been paired with improved architectures and communication improvements such as NVLink.

This step change has been recently illustrated by CoreWeave’s MLPerf Training v5.0 results that showcase the real-world impact of these architectural advancements, particularly with the Blackwell chipset. In collaboration with NVIDIA and IBM, CoreWeave deployed the largest-ever NVIDIA GB200 NVL72 cluster, utilizing 2,496 NVIDIA Blackwell GPUs across 39 racks. This massive deployment achieved breakthrough performance on the Llama 3.1 405B model benchmark, completing training in just 27.3 minutes and delivering more than 2x faster training performance compared to NVIDIA Hopper-based systems of similar cluster sizes.

Similarly, AMD has improved energy efficiency by 38 times since 2020 with its Instinct GPU based MI350 Series and is now suggesting that this can be improved a further 20 times by 2030 with Rack scale improvements. Deepseek’s big announcements in January were focused on techniques like PTX programming, which allow for more fine-grained control over GPU instruction execution, leading to more efficient GPU usage.

02 AI is unlikely to lead to very high electricity demand

These energy efficiencies are being supported by massive improvements in the cooling technology. Energy giants such as Shell, Exxon and BP (Castrol) are investing heavily in coolants for data centres and two-phase direct-to-chip cooling has emerged as the mainstream solution for handling the increased thermal demands of AI workloads and high-density server racks. Further into the future, immersion cooling systems, which completely submerge IT equipment in non-conductive cooling fluid, offer approximately 75% greater efficiency than traditional air-cooling methods and provide more precise temperature management, reduced space requirements, and offer better protection against contaminants.

Furthermore, GPUs themselves may be challenged as the most efficient architecture for AI workloads which could result in further energy efficiencies. Google’s Tensor Processing Units (TPUs), purpose-built for AI rather than graphics, are expected by some experts to outperform GPUs in both speed and energy efficiency. In addition, Cerebras is pushing the boundaries with Wafer-Scale purpose-built AI architectures; this approach reduces memory bottlenecks, allowing AI models to operate with dramatically lower power consumption and higher throughput compared to traditional GPU-based clusters. Other more exotic architectures like IBM’s neuromorphic computing or even quantum computing also demonstrate potential significant efficiency gains. With the pace of innovation accelerating, it's likely that new architectures and breakthroughs will continue to push the boundaries of energy efficient AI computing.

Energy giants such as Shell, Exxon and BP (Castrol) are investing heavily in coolants for data centres and two-phase direct-to-chip cooling has emerged as the mainstream solution for handling the increased thermal demands of AI workloads and high-density server racks.

02 AI is unlikely to lead to very high electricity demand

AI may make traditional high performance computing models far more eﬃcient

Beyond the architectures and the chips, there is evidence that GPU-based AI computing may well replace a broader range of conventional high performance computing tasks than previously thought. AI has already been gradually introduced into many high performance computing workloads, such as seismic processing, simulation and optimization. NVIDIA has estimated that by transitioning from CPU-only operations to GPU-accelerated systems, HPC and AI workloads can save over 40 terawatt-hours of energy annually, equivalent to the electricity needs of nearly 5 million U.S. homes. NVIDIA sees this as a market opportunity and recent announcements of extensions to the CUDA framework to support solvers like CPlex or Ansys’ computational ﬂuid dynamics simulators point in this direction.

AI is already reducing demand and holds signiﬁcant further potential

Whilst the focus of the discussion around AI has been around the impact of AI on energy demand, AI is driving significant energy efficiency through its application to the energy sector. A report by the world economic forum and Accenture has estimated that digital technologies and AI have the potential to reduce emissions by as much as 20%. In fact, AI is already driving a wave of innovation in the energy sector. In the two case studies, we demonstrate some large energy efficiency improvements using AI on buildings and wind turbines. AI is also helping generators design, schedule maintenance, and optimize the output of wind and solar farms using sophisticated techniques such as wake prediction for wind farms. It is being used to help grid operators squeeze more capacity out of existing transmission lines, to optimize the utilization of grid-scale batteries, to identify foliage encroachment on power lines, and to spot fires before they spread. AI is improving weather forecasting, which helps reduce peak heating and cooling loads. Leading utilities are using it to generate time-of-use price signals to send to customers, helping to nudge demand to match periods of high renewable supply. AI-driven power trading is improving the utilization of grid-connected batteries and other resources. It is also worth noting that AI is being used to measure emissions – most notably fugitive methane emissions through novel sensors paired with AI, enabling these to be eliminated at source. The list of potential use cases is endless.

02 AI is unlikely to lead to very high electricity demand

AI reducing energy usage with Capgemini’s Energy Command Center

Based at our Bengaluru site, Capgemini’s Energy Command Center (ECC) is built on IoT based architecture for intuitive resource management. It harnesses a data-driven approach and digitalization to monitor and manage performance of our energy assets. Since its launch, the ECC has reduced energy usage by 29% for the campuses that support Capgemini’s 170,000 employees in India. The EEC uses AI to measure and predict metrics like indoor air quality, energy intensity, water intensity, health of critical assets, critical operations, renewable energy generation, and the overall performance across all our energy assets. It also monitors and improves building energy and asset performance across Capgemini’s campuses as well as predicting condition-based maintenance.

Equinor are using AI to increase throughput on wind turbines. They have used AI to predict maintenance failures in wind turbine blades. The idea is that if they can fix multiple issues on a turbine at once, while it is already shut down for maintenance, they can reduce the total downtime when the turbine is not generating electricity. With the new AI solution on the Sheringham Shoal field, Equinor can already provide power to an additional 1000 households per annum due to reduced downtime.

IMAGE 1
Energy Command Center

IMAGE 2
Predictive Maintenance using AI at Sheringham Shoal field, UK

02 AI is unlikely to lead to very high electricity demand

We are at the engineering / AI tipping point for huge beneﬁts in the energy sector

The next phase of the AI journey further increases the potential. The impact of conventional forms of AI on energy has been constrained by the challenges of explainability. In a high-risk, safety-critical industry, many processes have remained untouched by AI because the algorithms were seen as a “black box”. However, the latest advances in AI are showing a path towards explainability. New reasoning models with chain of thought can explain their working. Scientific models like Alpha Fold are producing scientifically coherent predictions and emergent techniques and new AI-driven physics simulation based on methods such as Fourier neural operators (FNOs) or physics-inspired neural networks (PINNs) are demonstrating significant speed up over conventional simulation methods.

Latest AI research has demonstrated that new methods can radically accelerate traditional physics-based simulation approaches by 100,000x+. The integration of these new scientifically inspired AI methods opens up new horizons for the energy sector, where AI models can become more reliable and explainable and hence more integrated into core operations, resulting in further energy efficiency gains. All of this brings us back to the logical perspective that the improvements in hardware and software around the AI revolution will reduce the additional energy demand, as demand growth from new data centres is significantly offset by improvements in hardware and algorithms. This allows exponential improvement in AI with modest energy growth.

02 AI is unlikely to lead to very high electricity demand

Putting it all together

We disagree with the prevailing narrative that artificial intelligence (AI) will drive massive increases in global electricity demand. Whilst acknowledging that AI will contribute to higher energy use, we argue that many forecasts—some projecting up to 250% growth in U.S. electricity demand by 2030—are overly pessimistic. These projections often overlook the counterbalancing effects of technological advancements.

Historical trends like Koomey’s Law, which shows that computational efficiency doubles roughly every 18 months, suggest that energy use per unit of computation is falling rapidly, even as demand for AI grows. We believe the evidence supports this, with massive investment being made in improving the energy efficiency of chips and the associated architectures, alongside rapid software improvement and the opportunity to simplify existing workloads. In the medium term, we are positive about the impact that AI can have in reducing energy consumption, as tech is not just a source of energy demand, it is also a powerful tool for reducing it. AI is already being used to optimize energy systems, from improving the efficiency of wind and solar farms to enhancing grid-management and weather forecasting. Real-world examples, such as Capgemini’s 30% energy savings in building operations, demonstrate AI’s potential to drive down consumption.

So, whilst electricity demand will rise, the most extreme forecasts are likely exaggerated, and we may well be approaching a tipping point where AI becomes a net reducer of energy demand.

Is the forecasted signiﬁcant electricity demand for AI exaggerated?

Yes - by a lot
Yes - Somewhat
No

01 The electric grid investment paradox

Our Convictions

1

There is a big range in the forecasts for AI related electricity demand.

2

Innovation is happening in compute eﬃciency at a pace not experienced before.

3

AI is already being applied to reduce demand in the energy systems today.

4

We believe the industry has reached a tipping for further application of AI to energy infrastructure.

5

Therefore, we believe that current forecasts for increases in electricity consumption are over hyped - there will be rises but not 230% by 2030 as some analysts believe.