Artificial intelligence is rewiring the modern world. In classrooms and clinics, boardrooms and research labs, AI tools have evolved from novelty to necessity in less than two years. But AI’s seemingly effortless digital magic masks a crucial vulnerability: its enormous appetite for electricity.
Behind every use of AI is a vast physical infrastructure of power plants, transmission lines and data centers fighting to keep pace with ever-growing demand. AI’s need for power is intensifying the strain on electrical infrastructure and potentially driving up energy costs for consumers worldwide. At the same time, governments and corporations are racing to meet ambitious carbon reduction targets.
The numbers are staggering. Training a single large language model (LLM) can take as much power as 100 American households consume in a year. 1 A query on ChatGPT uses 10 times more energy than the equivalent Google search, and a single AI-generated image takes enough electricity to charge a smartphone. 2 As AI applications multiply across industries, some experts project they could account for 3–4% of global power demand by 2030 3 — a share that rivals the power consumption of entire nations.
The AI revolution has come with assumptions that computing and energy needs will grow exponentially, resulting in massive tech investments in both data centres and the means to power them, bolstering energy stocks. Deepseek’s launch of an open-source R1 model using around 2,000 Nvidia chips, at just a fraction of the computing power generally thought necessary to train similar programs, has raised questions about these power consumption assumptions. However, more efficient AI training will enable new models to be made with less investment, increasing the pace and breadth of adoption. This is the “Jevons Paradox,” the idea that increased efficiency can lead to increased consumption of a resource, rather than conservation. Deepseek and more efficient AI training approaches will likely reduce models’ power requirements; but by making AI modeling more accessible, it could increase AI’s overall power demand.
”Deepseek’s launch of an open-source R1 model using around 2,000 Nvidia chips, at just a fraction of the computing power generally thought necessary to train similar programs, has raised questions about these power consumption assumptions. However, more efficient AI training will enable new models to be made with less investment, increasing the pace and breadth of adoption.
Addressing these energy demands remains critical to AI’s sustainable development. The required infrastructure overhaul will drive unprecedented demand for industrial metals and critical minerals, transforming supply chains and creating new bottlenecks in AI’s expansion. Success in the next phase of the AI revolution may hinge less on computational breakthroughs than on securing the physical materials needed to power them.
What makes AI power-intensive?
From training complex models to maintaining the facilities that house them, AI operations create multiple layers of power demand. Broadly, they can be broken into computing requirements and infrastructure needs.
Computing requirements
Training Large Language Models
Training LLMs involves working with huge datasets and complex algorithms—and that calls for a lot of computing power. These models may have millions or even billions of parameters, requiring long training sessions on powerful graphics processing units (GPUs) or specialized hardware like tensor processing units (TPUs). Although Deepseek’s technology has shown greater efficiency, it can take up to half a million kilowatt hours to train existing models, and models must be continually retrained and tweaked.
Neural network calculations
Once a model has learned patterns from a dataset, the next stage is inference, the process by which a trained model makes predictions or classifications. While there is debate over whether training or inference uses more power, inference may overall have the greater impact; while the bulk of training is done once, inference is replicated over and over. Over time, inference could account for 70% to 80% more energy than training.4
Data center power capacity demands
AI applications typically run in data centers designed to handle high-performance computing tasks. These often involve many servers working together simultaneously. Additionally, data centers need advanced networking systems and equipment that allow servers to communicate with each other and share data efficiently, further increasing the need for power. Globally, the current data center power capacity is 59 GW. Driven by the computational needs of AI, Appian forecasts that global power capacity will grow by 23% through 2030, resulting in 217 GW of capacity in 2030.6
Figure 1—Data center power capacity and copper demand until 2040
Infrastructure needs
24/7 operation requirements
Many AI applications need to run continuously in order to provide real-time services or support ongoing research. The need for constant operation contributes significantly to the overall power consumption of AI technologies and is further compounded by the need for cooling and backup systems.
Cooling systems
High-performance computing systems generate substantial heat; effective cooling systems are necessary to prevent overheating and ensure hardware reliability. Cooling systems, however, are highly energy-intensive and can account for as much as 40% of a data center’s total energy consumption.6
Backup power
AI operations often require high availability and reliability, and downtime can be costly. Backup power systems are essential, but they must be maintained and powered continuously. They can contribute to as much as 10–12% of electrical distribution system losses.7
Generating power in the age of AI
As data centers consume ever-increasing amounts of electricity, power providers face a dual challenge: to dramatically expand capacity while reducing carbon emissions. These goals are reshaping the energy landscape, driving renewed interest in nuclear power and accelerating the adoption of renewable forms of energy.
”Updating the energy grid calls for significant investments in infrastructure, power transmission and energy storage solutions. Key materials such as copper, uranium, graphite, lithium and nickel play critical roles in this transition.
The nuclear renaissance
Long controversial, nuclear power is experiencing a global resurgence. Asia is expected to drive much of its growth, with the region’s share of global nuclear generation projected to reach 30% by 2026.8 The US announced plans at COP29 in Baku to triple its nuclear power capacity by 2050.9 President Trump’s “Unleashing American Energy” Executive Order has asked for a review of all policies that hinder the development and use of domestic nuclear power and has asked for Uranium to be potentially included in the US’ list of critical minerals.10 Meanwhile, Canada, France, Japan and the UK are collaborating with the US in developing an alliance designed to undermine Russia’s hegemony over nuclear supply chains. 11
Recent innovations in nuclear technology, especially the development of small modular reactors (SMRs), are also contributing to the renewed interest. SMRs are designed to be safer, cheaper and quicker to build compared to traditional large-scale reactors. As they can be prefabricated and installed more easily than standard plants, they allow for incremental capacity increases as demand grows. This flexibility makes them appealing for both new projects and upgrades to existing facilities.12
Driven largely by the growth of AI, the tech sector is leading the nuclear charge. Microsoft has signed an agreement to have its data centers powered by the Three Mile Island power plant, once notorious as the site of the worst nuclear accident in the US. Not to be outdone, Amazon and Google are investing billions into SMRs to power their expansion into AI technologies.13
Nuclear expansion and the rise of uranium
The uranium market has seen limited new mine development over the past decade due to historically low prices following the Fukushima disaster in 2011. However, uranium prices have been steadily increasing over the past two years, prompting companies to reopen old mines or develop new ones in regions such as Canada, India and the US.
Thanks to the growing reliance on nuclear power, the International Atomic Energy Agency (IAEA) projects that global nuclear power capacity could more than double by 2050, with demand for uranium potentially reaching up to 100,000 tonnes per year by 2040.14
Renewable energy integration
While nuclear energy provides reliable baseline power, renewable energy sources are becoming increasingly important to energy grid stability. Solar and wind power installations have grown exponentially, with costs dropping by over 70% in the past decade. However, this transition requires massive infrastructure updates.
Copper is set to play a pivotal role in the transition. Offshore wind has a copper intensity of up to 10 tonnes per megawatt of capacity, driven primarily by cabling required to connect the wind farm to the onshore grid. Onshore wind farms have up to 4.5 tonnes of copper per megawatt of capacity, while solar plants have up to 5 tonnes of copper per megawatt of capacity.15 The grid infrastructure and transmission systems (see below) needed to use and connect these renewable sources will require even more.
Stabilizing the energy grid
Already under stress from growing populations and trends toward electrification, aging power grids around the world are struggling under the additional demands of AI and data centers. Without significant upgrades, electricity costs are likely to surge while power outages increase.
Updating the energy grid calls for significant investments in infrastructure, power transmission and energy storage solutions. Key materials such as copper, uranium, graphite, lithium and nickel play critical roles in this transition.
Infrastructure
The backbone of any energy grid is its infrastructure, which includes transmission lines, substations and generation facilities. Copper plays an indispensable role here thanks to its exceptional electrical conductivity. It is used extensively in wiring for power lines, transformers and generators.
Copper is even more important when it comes to renewable energy systems. These require more copper per megawatt than traditional fossil fuel or nuclear plants because of the extensive cabling needed to connect widely dispersed solar panels and wind turbines to the grid. Driven primarily by growth in renewable energy and power infrastructure requirements, the demand for copper is projected to increase from 25.7 million metric tonnes in 2024 to 33.3 million metric tonnes in 2030, requiring a significant global supply response to meet demand.16 The fundamentals for copper remain favourable given structural supply challenges caused by declining ore grades, ageing mines, higher costs and fewer new projects being developed, calling into question the ability of future mined copper supply to meet demand.
Power transmission
Increasingly, modern grid stability depends on the ability to move power efficiently across vast distances, particularly as renewable energy sources are often located far from population centers. However, many of the high-voltage lines currently in use are outdated and operating at or beyond capacity.
Modernizing and expanding transmission capacity will require a large amount of minerals and metals, specifically copper and aluminum, the two main materials in wires and cables. The International Energy Agency (IEA) projects that over 80 million kilometers—the equivalent of the entire existing global grid—of grid capacity will need to be built or refurbished by 2040.17 The majority of this will be used for distribution. The IEA projects that copper used for power transmission could grow by as much as 50% by 2030.18
Energy storage
The final leg of the power grid tripod, energy storage systems (ESSs) balance electricity supply and demand by storing excess energy during low-demand periods and releasing it when demand peaks. ESSs are particularly important to efficient renewable energy use as they can store excess renewable energy when production exceeds demand and release it when needed.
The two most prominent forms of energy storage are batteries and pumped storage hydropower.
Pumped hydro storage
Pumped hydro storage (PHS) is the oldest and most widely used form of grid-scale energy storage. The system moves water between two reservoirs at different heights, using excess electricity during low-demand periods to pump water upward, then releasing it through turbines to generate power when demand increases. 19
This tried-and-tested technology provides massive storage capacity, with global installations reaching 160GW in 2021. However, it is limited by the need for specific geographical conditions.
Grid-scale battery storage
Grid-scale battery storage has gained significant traction in recent years, particularly with advancements in lithium-ion technology. These systems store electricity directly in batteries, which can be charged during periods of excess generation (e.g., from renewable sources) and discharged when demand peaks.
Power infrastructure and grid scale storage will represent a significant component of critical minerals demand growth over the coming decades, although it must be noted that EVs will have a much more significant impact.
Electric vehicle (EV) and battery energy storage system (BESS) demand currently represent 88% of global battery demand, with this increasing to 95% by 2040.
BESS demand is forecast to grow from 201 GWh in 2024 to 1,038 GWh in 2040, growing at a 10.8% CAGR over the period.
This growth in battery use for EVs and BESS will impact demand for several key metals and minerals 20:
Powering AI’s growth with critical minerals
As data centers proliferate and power grids expand to meet the demands of AI, the pressure on critical mineral supplies is reaching unprecedented levels. This is creating new challenges for both technology companies and resource providers.
Supply chain challenges
As easily accessible deposits become scarcer, mining companies require increasingly sophisticated and costly extraction methods. Additionally, operations must navigate stricter environmental, social and governance (ESG) standards that ensure sustainable and responsible resource extraction. Geopolitical tensions further complicate the supply chain: Key mineral resources are often concentrated in politically sensitive regions, potentially affecting stable access to these
critical materials.
The path forward
The changing and surging power needs of the global economy are helping drive the mining sector to invest in innovative exploration technologies and more efficient extraction methods. The race between AI’s growing power demands and mineral supply capabilities remains tight, however. The success of the AI revolution will likely depend on how effectively nations and companies can secure and manage these crucial resources while maintaining sustainable practices and stable supply chains.
Sources:
1. https://www.contrary.com/foundations-and-frontiers/ai-inference
2. https://www.theverge.com/24066646/ai-electricity-energy-watts-generative-consumption
3. https://www.goldmansachs.com/insights/articles/AI-poised-to-drive-160-increase-in-power-demand
4. https://semiengineering.com/ai-power-consumption-exploding/
5. Appian Capital Advisory LLP, J.P. Morgan, McKinsey, Semi Analysis, Wood Mackenzie
6. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/investing-in-the-rising-data-center-economy
7. https://www.csemag.com/articles/evaluating-ups-system-efficiency/
8. https://www.iea.org/reports/electricity-2024/executive-summary
9. https://www.bloomberg.com/news/articles/2024-11-12/cop29-us-has-plan-to-triple-nuclear-power-as-energy-demand-soars
10.https://www.newsweek.com/donald-trump-nuclear-energy-popular-democrats-1944649
11. https://www.power-technology.com/news/canada-france-japan-uk-us-nuclear-alliance/
12. https://www.iaea.org/newscenter/news/what-are-small-modular-reactors-smrs
13. https://www.ecowatch.com/amazon-google-nuclear-energy-ai-data-centers.html
14. https://www.iaea.org/newscenter/news/iaea-symposium-examines-uranium-production-cycle-for-sustainable-nuclear-power
15. https://www.visualcapitalist.com/sp/copper-intensity-of-renewable-energy/)
16. Appian Capital Advisory LLP
17. https://www.iea.org/reports/electricity-grids-and-secure-energy-transitions/executive-summary
18. https://www.iea.org/reports/the-role-of-critical-minerals-in-clean-energy-transitions/mineral-requirements-for-clean-energy-transitions
19. https://www.energy.gov/eere/water/how-pumped-storage-hydropower-works
20. https://www.visualcapitalist.com/sp/visualizing-the-future-demand-for-battery-minerals/
