AI's role in powering renewable energy growth

What happened

The surge in data center power consumption has become a significant factor in global electricity demand. A recent report from the International Energy Agency (IEA) highlights that the electricity consumption of data centers, which includes activities such as AI model training and bitcoin mining, could potentially double by 2026.

In 2022, data centers consumed approximately 460 TWh of electricity, representing almost 2% of the total global electricity demand. Computing and cooling account each for 40% of data centers’ power consumption, with the remaining 20% consumed by other associated IT equipment.

Latest data suggests that data centers contributed to about 14% of 2023’s global growth in electricity demand (+90 TWh). This increase is comparable to the electricity demand growth from air conditioning and exceeds that of electric vehicles.

Looking forward, the IEA projects that by 2026, the electricity demand for data centers will range between 620 TWh and 1,050 TWh (base case at ~800 TWh). This projection, while dependent on various factors such as technological advancements and the evolving trends in AI and cryptocurrency, indicates a potential increase of up to 590 TWh from 2022 levels. Such an increase would be roughly equivalent to adding the entire electricity consumption of a country like Sweden or even Germany to the global grid.

Impact on our Investment Case

Data Center Power Demand: Geographical Focus

There are currently about 11'000 data centers globally, with approximately 50% located in the U.S., 20% in the European Union, and about 5% in China. 

In the U.S., the IEA projects that data centers' power demand will reach nearly 260 TWh by 2026 (IEA's base case), representing 6% of U.S. power demand (from 4% today). Over the past decade, the U.S. has seen virtually zero growth in power demand on average, meaning that the growth in renewables has primarily replaced fossil-based power sources like coal. Despite economic and population growth, improvements in energy efficiency have offset increased demand. Innovations such as the widespread adoption of LED lighting and the shift from traditional on-premise to cloud and hyperscale data centers have significantly reduced power usage. However, as the demand for digital services continues to surge, these efficiencies are no longer sufficient to curb the rising power consumption of data centers. 

In China, electricity demand in the data center sector is expected to reach approximately 300 TWh by 2026. Unlike the U.S., China's electricity demand has grown significantly in recent years, with an average annual growth rate of about 5.9% from 2012 to 2022. This sustained increase in electricity demand could further stimulate the existing boom in renewable energy deployment, as China continues to expand its wind, solar, and hydroelectric power generation to meet rising energy needs.

In the European Union, data center electricity consumption is projected to reach almost 150 TWh by 2026, driven by the increasing digitalization and AI computations. With data centers primarily concentrated in financial hubs like Frankfurt, London, Amsterdam, Paris, and Dublin, the demand for electricity in the EU data center sector will significantly contribute to the power consumption in these specific cities.

Power Consumption Dynamics

The rapid growth in power consumption by data centers is significantly driven by the increasing incorporation of AI and cryptocurrency activities. Understanding the distinct power demands of these technologies is key for projecting their future impact on electricity consumption.

AI vs. Traditional Computing

AI applications, particularly those involving large-scale machine learning models, consume significantly more power than traditional computing tasks. A typical Google search uses about 0.3 Wh, whereas an AI-powered request to OpenAI's ChatGPT uses approximately 2.9 Wh. With Google handling around 9bn searches daily, fully implementing AI in google search could potentially add 10 TWh of annual electricity demand.

The power requirements for AI come from both training and inference processes:

• Training: This initial phase involves AI models learning from vast datasets, requiring extensive computational power typically provided by GPUs. Training state-of-the-art models can consume thousands of kilowatt-hours of electricity. For instance, the training of GPT-3 model consumed approximately 1'300 MWh of electricity, equivalent to the annual power consumption of 130 U.S. homes.
• Inference: This phase applies trained models to new data. While less energy-intensive than training, inference still requires significant power, especially when scaled across billions of requests (e.g., ChatGPT queries or self-driving car operations).

Currently, the AI sector is estimated to consume less than 8 TWh of electricity annually. This consumption is expected to increase tenfold over the next four years, with some studies projecting annual usage to reach between 85 and 184 TWh by 2027. This would represent approximately 0.5% of global electricity consumption.

Cryptocurrency Mining

Cryptocurrency mining, particularly Bitcoin, is a significant contributor to data center power consumption. In 2022, cryptocurrencies consumed around 110 TWh of electricity (out of the total 460 TWh for data centres), equivalent to 0.4% of global electricity demand. This demand is projected to rise to approximately 160 TWh by 2026, a 40% increase.

Mining involves validating transactions and adding them to the blockchain ledger, requiring substantial computational power provided by specialized hardware like ASICs and GPUs. This process, known as proof of work, is energy-intensive due to the need to solve complex cryptographic puzzles. As more miners join the network, the difficulty of these puzzles adjusts to ensure a consistent block production rate, often leading to higher energy consumption.

Despite improvements in mining efficiency, which allow for more operations with less energy, these gains are not necessarily sufficient to offset the overall increase in mining activity and growing network difficulty. Currently, the network's weighted average efficiency is 34 watts per terahash (W/T), with an 8% improvement in the past year and a 28% improvement over the past three years. However, Bitcoin mining hash rate (representing the total computational power being used by miners) has increased by over 100% and over 250% over the same periods, respectively. Note that by mid-2026, mining efficiency levels could reach as low as 10 W/T due to advancements in chip design and more efficient hardware.

Technical Innovations and Power Efficiency

While the shift towards more efficient data centers (e.g., cloud and hyperscale centers) has historically driven power efficiency gains, the rapid growth in demand for AI and cryptocurrency operations has begun to outpace these improvements. Innovations in AI hardware, such as NVIDIA’s DGX systems, demonstrate both the challenges and potential efficiencies. The NVIDIA DGX A100 system delivers 5 petaFLOPS (a measure of processing speed) while consuming 6.5 kW, whereas the newer DGX H100 system offers 32 petaFLOPS with a consumption of 10.2 kW. This trend shows a significant reduction in power intensity (kW per petaFLOPS) but an overall increase in power consumption per server due to higher computational capacities.

Driving Renewable Energy Growth

The rising power demand from data centers presents a significant opportunity for renewables. Hyperscale data centers, operated by major tech companies like Amazon, Microsoft, Meta, and Google, are very often committed to 100% renewable energy. This commitment is expected to significantly boost the demand for renewables, especially in the U.S.

Large corporations fulfill their renewable energy commitments through Corporate Power Purchase Agreements (C-PPAs), which are long-term contracts between companies and renewable energy providers. C-PPAs have been major drivers of renewables growth, with 80% focused on solar and 20% on wind from 2021 to 2023. Over the past five years, C-PPAs accounted for nearly 64 GW of solar capacity in the U.S., about 80% of total utility-scale solar installations.

As previously mentioned, data centers in the U.S. are expected to increase their share of power demand from 4% to about 6% by 2026. Some analysts even predict they could account for 8% by 2030. While the exact additional power demand from AI and cryptocurrencies is difficult to predict, various scenarios have been proposed.

One study from Goldman Sachs suggests that data centers will require approximately 47 GW of additional power generation capacity in the US by 2030, with a mix of 60% natural gas and 40% renewable sources. This translates to around 19 GW of new renewable capacity (13 GW solar and 6 GW wind).

A recent UBS analysis projects 175 TWh of incremental load growth for AI data centers by 2030, on top of the previously forecast 175 TWh from regular data centers, totaling 350 TWh. UBS expects this incremental demand to be fully met by renewables (80% solar, 20% wind), translating into about 45 GW of new utility solar and 12 GW of new wind capacity.

In 2023, U.S. utility-scale solar installations totaled 22.5 GW, and new wind installations reached ~8 GW. Adding 2-7 GW of utility-scale solar and 1-2 GW of wind per year due to new data center demand, which was not previously expected or accounted for, makes a non-negligible difference.

It is challenging to determine if these numbers are accurate, and data center demand might exceed all current scenarios. Historically, the IEA and other analysts have underestimated the real deployment of renewables. Our belief is that most scenarios and forecasts underestimate the real impact of AI demand (and, to a lesser extent, cryptocurrency) on renewable deployment, suggesting more upside potential.

Key investment opportunities

The surge in power demand from data centers, driven by AI and other digital services, presents significant opportunities for various players in the renewable energy sector. U.S.-based manufacturers of renewable technologies like First Solar, Array, and Nextracker Inc stand to gain significantly. First Solar's dominant market position in thin-film solar panel manufacturing, combined with benefits from domestic manufacturing tax incentives and import tariffs, gives it a strong edge in the market. Similarly, Array Technologies and Nextracker, specializing in solar tracking systems, will see increased demand as new solar projects aim to maximize efficiency and output.

Schneider Electric SE is also poised to benefit greatly from this trend. The data center segment, (~20% of their sales), is seen as the largest near-term opportunity. Schneider offers a broad range of products and solutions, including power management systems, cooling systems, racks, and IT management systems, all designed to meet the evolving needs of data centers

Modernizing the grid is essential to handle the increased supply and demand from data centers. Companies like Itron Inc, which specializes in smart grid technologies and advanced metering, will benefit from the need to upgrade the grid to support the integration of renewable energy sources and meet dynamic data center demands.

Our Takeaway

Demand for data centers is driving renewable growth, especially in the U.S., making utility-scale renewable energy more essential than ever. While different scenarios exist, it is difficult to predict precisely tomorrow’s power needs. However, we believe that demand will exceed expectations, driven not only by AI training and inference, and crypto mining but also by the adoption of electric vehicles, air conditioning, and heat pumps.

Renewables are not the only generation technology, and future demand will undoubtedly be met by other power sources such as natural gas. However, the share of renewables will continue to increase. Additionally, utilities' capital expenditure will go into transmission and distribution technologies, as well as digitizing the grid, offering ample investment opportunities across the value chain.