AI’s energy equation: How clean energy and efficiency could balance demand growth

The rise in energy demand driven by the adoption of generative artificial intelligence (AI) technologies has been widely written about; many consider it a potential bottleneck to growth. However, in our view, its relative scale and the resultant impact on the environment is not well understood.

Many forecasters conclude that while energy demand from AI data centres will grow rapidly, it will remain a small contributor to overall energy demand. Indeed, the International Energy Agency (IEA) asserts that AI data centres will account for less energy demand than electric vehicles (EVs) and industry growth.

Furthermore, the operators of large-scale data centres, the so-called hyperscalers, have set net zero targets and are rapidly increasing their use of power purchase agreements (long-term contracts for electricity supply) and low carbon energy; they are also improving the energy efficiency of their data centres.

AI data-centre energy forecasts may also underappreciate AI’s role in reducing energy demand in end-use cases; this could cancel out increased data-centre energy demand. Thus, the impact on the environment may be much less material than initially anticipated.

“AI data-centre energy forecasts may also underappreciate AI’s role in reducing energy demand in end-use cases.”

AI’s energy consumption

Observation: AI’s power needs are growing rapidly but will likely only account for a minor share of global energy demand.

AI computing is inherently more energy intensive than traditional computing due to the large amounts of data processed in both training and usage. According to the IEA, global energy consumption from data centres (all workloads, including AI) is expected to grow at a 15% compound annual growth rate (CAGR) from 269 terawatt-hours (TWh) in 2024 to 946 TWh in 2030¹.

Accelerated servers, used mainly in AI computing, are expected to grow at a 30% CAGR to 2030, and are therefore set to contribute 50% of the growth in total data-centre energy demand during this time.

Despite this rapid growth, the contribution to overall energy demand is expected to remain small. The IEA estimates that total data-centre share of overall global energy demand will increase from 1.5% to only 3% by 2030 in its base case.

Even in its scenario of rapid adoption of AI, data-centre demand reaches only 4% of total energy demand by 2030. It forecasts that as a sector data centres will be the fifth largest contributor to growth in energy demand, with industry (excluding heavy industry) contributing 3x as much and electric transport 50% more.

The impact on emissions is also not as material as one might think. The IEA estimates in its base case that AI will account for 1% of all energy emissions by 2030 and decline thereafter as AI efficiencies are brought to the energy system.

AI and clean energy

Observation: Growth in energy demand will be powered by renewables as tech companies target emissions reductions.

Most cloud-computing companies have voluntarily set 2030 net zero targets, except Amazon which targets 2040. These companies plan to achieve this through the use of clean energy, renewable energy purchases, data-centre efficiencies and 24/7 clean energy matching.

Given the pace of investment in AI, and thus the need for power, hyperscalers prioritise fast-to-market solutions. This perhaps explains their preference for power purchase agreements (PPAs), which involve long-term contracts to purchase renewable energy from an existing renewable generator.

In 2024, Amazon, Google, Microsoft and Meta signed deals for 25 gigawatts (GW) of clean PPAs, equivalent to 43% of clean PPAs globally.

Nuclear power is also becoming an attractive option for tech companies as it is low carbon emitting and provides the consistent baseload power needed for data centres. However, available capacity is limited, and capital expenditure (capex) requirements and lead times are significant.

Despite this, since 2023 cloud providers have signed a total pipeline of 32GW worth of nuclear power contracts in the US, according to BloombergNEF (BNEF)². Some companies are using the restart of existing nuclear power plants to circumvent the long lead times involved.

For example, Microsoft is funding the restart of Three Mile Island, an 837MW nuclear plant in Pennsylvania, which will commence operations in 2027. Meanwhile, Amazon is investing in new technologies – specifically, 5GW of small modular reactors (SMRs), which will come online over the next 15 years.

Cloud computing providers are also turning to other clean technologies, such as hydrogen fuel cells as contracted by Oracle. Other businesses, such as xAI, are using utility scale battery storage (BESS) to smooth electricity demand thus enabling the use of variable renewables supply.

Some hyperscalers are starting to turn to natural gas given its quick time to power. For example, Meta’s data centres in Louisiana will be powered by 2GW of gas-fired power. Others are hesitant about the resultant emissions footprint from this approach, with Microsoft only considering natural gas with carbon capture technology.

“Nuclear power is also becoming an attractive option for tech companies as it is low carbon emitting and provides the consistent baseload power needed for data centres.”

Data centre efficiency

Observation: The energy efficiency of data centres will continue to improve.

Energy efficiency in data centres has played a critical role in curbing growth in energy demand. During the last decade of cloud adoption, energy demand has remained flat³.

This can partly be explained by a shift of workloads from enterprise data centres to hyperscalers. The latter have an efficiency (measured in power usage effectiveness (PUE)⁴) 60% higher than enterprises. It can also be explained by hyperscalers improving their PUE during this time. Google, for example, has reduced its PUE from 1.20 to 1.10 over 14 years⁵. 

The rate of improvements in energy efficiency of the total data centre fleet will likely slow as the shift to hyperscalers slows down. However, hyperscalers will likely continue to improve the energy efficiency of their data centres. This will include:

• Hardware: photonic integrated circuits, chiplets used for specific tasks, energy efficient memory, storage and innovative cooling techniques, such as direct-to-chip cooling. Google and Nvidia have already reported 80-fold and 25-fold improvements in the energy performance of their new AI chips⁶.
• Software: smaller AI models for specific tasks; for example, OpenAI’s o3-mini, and energy efficient models such as DeepSeek.
• Others: edge and quantum computing and virtualisation.

Overall, the IEA expects energy saving initiatives will result in a 17% reduction in data-centre energy demand between 2024 and 2030, largely in conventional servers and power infrastructure⁷.

In the best case, it expects a 15% improvement in data-centre efficiency, which would result in data centres, including AI and non-AI, only accounting for 2.6% of energy demand by 2030.

Using AI to reduce energy consumption

Observation: AI will also enable energy savings in end-use cases.

The energy savings that AI could enable through optimising energy generation, distribution and end consumption are often overlooked, in our view. Applications in real time electricity dispatching could have material impacts. Other end-use applications include building management, robots in manufacturing, predictive maintenance, fleet management and EV charging.

Some of the companies in our portfolios are actively deploying AI technologies to reduce their own energy consumption and those of their customers. For example:

• Johnson Controls – OpenBlue, Johnson Controls’ AI-optimised smart building system of connected heating, ventilation and air conditioning (HVAC) and other appliances, can deliver a 30% reduction in energy spend for customers.
• Schneider Electric – The company’s site in Hyderabad, India, uses AI predictive monitoring, and real-time data generation and internet of things (IoT) capabilities. This has resulted in a 59% reduction in electricity and 61% reduction in emissions, as well as a 57% reduction in water.
• Siemens – The company’s factory in Chengdu, China, is using digital energy management, predictive maintenance and AI-based automation to achieve a 24% reduction in electricity production per unit.

The IEA estimates that, in the case of rapid adoption of AI, 8% energy savings could be achieved in light industry and as much as 20% in EVs by 2035. Overall, in this scenario, AI would lead to a reduction of 1.4 gigatonnes of carbon dioxide. Further, research by consultants PwC finds that from 2024 to 2035 total energy savings outside data centres would be 0.1% to 1.0% and would roughly offset an increase in energy usage due to AI⁸.

“The energy savings that AI could enable through optimising energy generation, distribution and end consumption are often overlooked, in our view.”

Other studies have drawn similar conclusions showing that AI can have meaningful impact on the energy transition⁹. For example, Google’s DeepMind model has identified 45 times the number of theoretical crystals than discovered so far, helping to further the development of renewable energy and storage technologies. However, the rate of adoption of these technologies is uncertain and the impact of rebound effects, that is an increase in demand due to cost savings, is also very unclear.

There are of course other environmental risks such as the increased use of water associated with data-centre growth. Though water is predominantly used in energy generation, it is also increasingly employed in cooling. New technologies in liquid cooling, such as direct-to-chip cooling, could address this issue.

Conclusion: Clean energy and use of AI may be silver lining to growth in energy demand.

Growth in energy demand as a result of AI adoption is expected to be material. However, this must be put into the context of growth in electrification more broadly. Motivated by their net zero targets, hyperscalers are increasingly contracting clean energy to meet energy demand; they are also investing heavily in energy and water efficiency, thus reducing their impact on the environment.

AI’s role in reducing energy demand should also be considered. If rapidly adopted, it could offset growth in energy demand, as well as emissions from AI computing. However, the rate of adoption is of course still uncertain. It appears therefore that every cloud, even those run by hyperscalers, may have a silver lining.

¹ IEA, ‘Energy and AI – World Energy Outlook Special Report’, (iea.org), April 2025.
² BNEF, ‘Power Hungry Data Centres Are Driving Green Energy Demand’, (about.bnef.com), August 2025.
³ Oxford Energy Forum, ‘AI’s indirect impacts on climate outweigh concerns over its direct energy footprint’, (oxfordenergy.org), Issue 145, May 2025.
⁴ The energy efficiency of a data centre is measured by the Power Usage Effectiveness (PUE). This measures the total energy used by the facility compared to the energy used by its IT equipment.
⁵ Google, ‘2024 Environmental Report’, (sustainability.google), July 2024.
⁶ Oxford Energy Forum, ‘AI’s indirect impacts on climate outweigh concerns over its direct energy footprint’, (oxfordenergy.org), Issue 145, May 2025.
⁷ IEA, ‘Energy and AI – World Energy Outlook Special Report’, (iea.org), April 2025.
⁸ PWC, ‘Could net-zero AI become a reality?’, (pwc.com), April 2025.
⁹ Scientific American, ‘AI’s Climate Impact Goes beyond Its Emissions’, (scientificamerican.com), December 2023 and Nature, ‘Green and intelligent: the role of AI in the climate transition’, (nature.com), June 2025.