Key Questions on Energy and AI (IEA)
Author: International Energy Agency
IEA (2026), Key Questions on Energy and AI, IEA, Paris https://www.iea.org/reports/key-questions-on-energy-and-ai, Licence: CC BY 4.0
0 General Document comments
0 Sentence and Paragraph comments
0 Image and Video comments
Key Questions on Energy and AI
Executive summary
The AI and energy nexus continues to evolve rapidly
The largest technology companies are contributing to a surge in data centre investment, as their capital expenditure exceeded USD 400 billion in 2025 – and is expected to jump by another 75% in 2026. Capital expenditure of just five technology companies is now larger than global investment in oil and natural gas production. Many jurisdictions are seeing project pipelines accelerate dramatically, although not all projects will come to fruition. Those that are moving forward are doing so at pace: the IEA’s unique satellite-based tracking shows that “artificial intelligence (AI) factories” – cutting-edge data centres specifically designed for AI – have more than tripled in capacity in the past 18 months. Meanwhile, the capabilities of AI are improving quickly, increasing the likelihood that it will reshape economic growth, innovation and competitiveness and disrupt established industries and jobs.
- This paragraph is mostly making a claim, then stacking facts under it to make that claim feel urgent.
- The main argument seems to be that AI is driving a huge wave of data-centre spending, and that this wave matters because it could reshape the economy, business competition, and jobs.
- The facts are the proof: the USD 400 billion number, the comparison to oil and gas, the faster project pipelines, and the tripling of AI-factory capacity all support that bigger point.
- It also has a subtle push in its wording: phrases like “at pace,” “cutting-edge,” and “increasing the likelihood” nudge the reader to see this as a major shift, not just neutral information.
Try this in Your Thoughts: name the sentence that feels most like the core claim, then ask yourself whether the other sentences act more like evidence, comparison, or prediction—or would you rather click Continue with AI or go to Create Document to start something new?
In April 2025, the IEA published its landmark Energy and AI report, which provided first-of-its-kind global analysis on the links between AI and energy. Since then, the field has evolved rapidly: new questions have emerged and new data has come to light. This report builds on the foundation of previous work, providing fresh analysis on key issues. It covers the outlook for data centre electricity demand considering recent market developments; innovations in powering data centres; and the implications of these trends for energy security, affordability, competitiveness and overall energy demand.
When I first read the title of the article, my mind went to what my students often say about AI – that is environmentally harmful (water use, for example). So I’m interested in what this report says about those kinds of issues. Paragraph 97 speaks to some social concerns, but doesn’t say a lot about environmental impacts beyond, “The emissions associated with data centres double in IEA projections, reaching around 350 million tonnes in 2035, but still make up about 2% of global electricity sector emissions by that date.”
- Yes—people have written about ways AI can do less environmental harm, and a lot of that work focuses on smaller models, better chips, cleaner electricity, and more careful use of data centers (UNESCO, “AI Large Language Models: new report shows small changes can reduce energy use 90,” 2025; IEA, “AI and climate change,” 2025).
- One useful thread in the writing is that “less harmful” usually means design choices: compressing models, using smaller task-fit models, and improving hardware and cooling systems rather than assuming bigger is always better (UNESCO, “AI Large Language Models: new report shows small changes can reduce energy use 90,” 2025).
- The IEA also frames this as a mixed picture: AI can raise data-center energy demand and emissions, but it may also help cut emissions in sectors like transport, power systems, and industry depending on how it is deployed (IEA, “Energy and AI,” 2025; IEA, “AI and climate change,” 2025).
- If you want to respond to paragraph 97, one strong move might be to note that the report gives a broad emissions number, while other writers push for a fuller picture that also includes water, hardware manufacturing, and the cumulative effect of everyday inference requests (MIT News, “Explained: Generative AI’s environmental impact,” 2025; “Holistically Evaluating the Environmental Impact of Creating Language Models,” 2025).
What if you wrote one quick note in Your Thoughts that starts with this contrast: the report measures one kind of harm clearly, but other sources ask us to count harms that are easier to miss—do you want to click Continue with AI to sort those ideas, or go to Create Document to upload a source and keep reading or writing?
Energy consumption per AI query has declined massively, but much more energy-intensive use cases are becoming increasingly popular
Measured per individual task, the energy efficiency of AI is improving at a rate unprecedented in energy history. Software and hardware advances have resulted in the energy use per AI task dropping by at least an order of magnitude annually in recent years. Simple text queries now typically consume less electricity than running a television over the same period of time. If all conventional internet searches were performed with simple AI text queries, it would consume less than 4 terawatt-hours (TWh) of electricity annually, equivalent to less than 1% of total data centre consumption today.
However, new energy-intensive AI applications are increasingly being launched and used, such as those for video generation, reasoning and agentic tasks. These kinds of tasks can consume hundreds or thousands of times more energy per query than simple text generation. The energy demand of AI is therefore the result of three rapidly evolving and uncertain trends: improvements in efficiency, surging uptake, and changing model capabilities, which can unlock new and, in many instances, far more energy-intensive use cases. To improve the robustness of the outlook for AI’s energy demand, close monitoring, frequent updates and cooperation with the tech sector, including more systematic energy consumption disclosures, will remain important.
The global electricity demand of data centres – the critical infrastructure for training and running AI models – grew by 17% in 2025, in line with IEA projections. Electricity consumption from AI-focused data centres grew even faster, surging 50% in 2025. While there are no comprehensive statistics on the frequency and depth of AI usage around the world, major model providers reported a threefold increase in active users and a fivefold increase in revenue over the past year, highlighting the rapid growt/h of demand.
Capital expenditures by hyperscalers
AI use on OpenRouter/
Across the AI value chain, a scramble for electricity, grid connections, manufacturing capacity, chips and capital has set in
The speed of the AI revolution is increasingly contrasting with the speed of the physical, social and economic systems that underpin it. Bottlenecks across energy supply chains and advanced chip manufacturing have tightened since our last report. Planning and regulatory systems are being stretched by the wave of project applications for data centres, amid a broader trend of rapid load growth and electrification. Social acceptability is also a growing issue, as communities push back against data centre projects, and concerns about affordability and environmental impacts rise. Essential elements within the IT industry are currently facing limitations; notably, a shortage of high-bandwidth memory – integral to AI chip production – has developed over the past six months and is anticipated to persist through at least the end of 2027.
what does this mean?
- This sentence says a key computer part is in short supply, and that part is needed to make AI chips.
- High-bandwidth memory is a very fast kind of memory, so if there is not enough of it, companies cannot make AI chips as easily or as quickly.
- The sentence also says this shortage started about six months ago and may last until late 2027, so the problem is not expected to end soon.
- In simpler words: demand for AI chips is very high, but one important ingredient is hard to get right now.
What would help more right now: breaking the sentence into smaller parts, or explaining what “high-bandwidth memory” does inside a computer while you either click Continue with AI or go to Create Document to start writing or upload a source?
- High-bandwidth memory is a stack of memory chips built mostly from silicon, with tiny metal wiring inside and between the layers.
- It also uses transistors and capacitors to store data, plus very small solder bumps and a base layer that helps the stack connect to a processor.
- The key idea is not a special “mystery material,” but the way the memory is packed in layers and linked with very short vertical connections.
- So when people say there is a shortage, they usually mean making enough of these advanced stacked memory packages is hard, not that the world ran out of one simple raw ingredient.
Want to focus next on the physical materials, or on why this kind of memory is harder to manufacture than regular memory while you either click Continue with AI or go to Create Document to start writing or upload a source?
Data centre investments have grown too large to be funded from company balance sheets alone, and large amounts of funding from capital markets will be critical for their buildout. As a result, the pace of data centre growth, and the resulting increase in energy consumption, will be sensitive to market sentiment, including expectations for returns on investment in data centres and AI deployment, as well as to broader macroeconomic and financing conditions. Understanding the energy implications of AI therefore also means following closely the technology’s economic trajectory.
Our updated data centre electricity demand outlook sees near-term bottlenecks but longer-term upside
The central projection for electricity demand from data centres remains close to the trajectory set out in the IEA’s 2025 report. Our updated projections see electricity consumption from data centres roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, accounting for around 3% of global electricity demand by that date. Electricity consumption from AI-focused data centres grows much faster than overall data centre electricity consumption, tripling in this period. Bottlenecks across the value chain, however, are reducing the likelihood of more aggressive near-term scenarios, despite booming investment and surging project pipelines.
The mid- to longer-term outlook for data centre electricity demand sees a possible upside. Investments in relieving bottlenecks across energy equipment and chip manufacturing, and rapid uptake of energy-intensive use cases of AI, raise the prospect that there could be an even higher upside case after 2030. The IEA will continue to update its projections regularly.
Electricity demand growth in 2025
Electricity consumption by data centres, 2020-2035
- Case range
- Historical
- Base case
AI is pushing data centre power density to the limits of today’s technologies
An individual server rack within an advanced data centre is only the size of a large refrigerator, but by 2027 it could have peak power demand equivalent to that of 65 households. The speed of this shift is remarkable: between 2020 and 2025, the power density of AI servers increased by 11 times; by 2027, it is set to see a further fourfold increase. This will test the capacity to ramp up supply chains for key electricity technologies such as power electronics and transformers. Some of these technologies also depend on critical inputs from a small number of producers, notably China. Care is therefore needed to ensure that supply chains for the critical emerging technologies going into data centres are diverse and resilient. Power density Power distribution Load variability Heat density Energy storage Weight In the next few years, AI data centres are set to concentrate a peak power demand equivalent to 65 households in a space equal to a household refrigerator. AI data centres are shifting towards high voltage, direct current power distribution, building on technologies pioneered in EVs and renewables. AI data centres see repeated swings of server load of more than 50% of rated capacity within a second. AI data centres need to evacuate the heat equivalent to 30 natural gas boilers per server rack. To manage load swings, storage technologies, including lithium- ion batteries, are becoming critical for AI data centres, potentially making them more flexible, grid-friendly assets. In an AI data centre, a single server rack the size of a refrigerator can weigh more than a pick-up truck. AI data centres versus Conventional
Unlike traditional data centre operations, AI training and model use induce large and rapid power swings, making energy storage critical to ensure that electricity is always supplied reliably. By 2030, around 20-25 GW of battery storage could be installed in data centres globally, potentially making them a grid asset if the incentives are right. Recent months have seen a data centre operator sign an agreement for the largest battery project ever by energy capacity (four times larger than the previous record-holder), helping to commercialise long-duration energy storage.
Total data centre electricity demand covered by renewables PPAs
- United States (light blue)
- Europe (dark blue)
- Rest of world (green)
PPAs and offtake agreements in firm power technologies for data centres
- Existing nuclear
- SMR
- Battery and demand response
- SMR
Constrained by slow grid connections, data centre developers in the United States are pushing forward projects with onsite natural gas-based power generation. Satellite tracking of these projects indicates that around one-fifth have started land clearing or construction. This highlights that onsite gas power is an emerging solution, with key design, supply chain, regulatory and financial questions remaining. New IEA analysis shows that providing reliable onsite gas-fired electricity to meet critical and variable data centre load requires overbuilding onsite generation infrastructure by 30% to 70% relative to demand. Yet in the context of a supply crunch for gas turbines, it is not clear that onsite generation necessarily promises a faster route to market for data centres at scale. Though uncertainties are high, around 15‑27 GW of onsite natural gas may power data centres by 2030, mostly in the United States. However, this does not remove the urgency of addressing grid bottlenecks, as most data centres prefer connecting to the grid.
The AI boom could therefore accelerate deployment and innovation in the electricity sector, if spending continues and if government support is also aligned. In this report, we analysed share price movements of AI and energy companies to understand how financial markets expect AI’s energy demand to impact energy companies. On the one hand, financial valuations since the launch of ChatGPT do not suggest that AI demand will provide a generalised uplift to the energy sector; it is simply too small in the context of the energy sector as a whole. In contrast, manufacturers of gas turbines and electrical equipment, some nuclear companies, and some energy startups have seen their valuations become more strongly linked to AI, highlighting the pull that data centre demand is providing for innovation and growth in their businesses.
Global natural gas turbine orders, 2024-2025
- Associated with data centres
- Other
Pipeline of onsite natural gas power projects for data centres in the United States, 2026
Even if AI boosts economic growth, the impact on energy demand will be lower
Initial signs indicate AI is boosting productivity in some sectors, which may push up overall economic growth. There are a wide range of projections of the impact of AI on GDP. Based on cooperation with the OECD and economic modelling of the possible task-by-task productivity boost coming from AI, this report provides a first-of-its-kind analysis of the implications of an AI-driven GDP boost on the energy sector.
Stronger economic growth from AI does not translate one-to-one into higher energy demand. It is concentrated in knowledge-intensive services and high-income countries, where the elasticity between energy demand and economic activity is lower. Estimates show that, depending on the scale of uptake, an AI-driven growth boost could raise the level of global energy demand by between 1-4% in 2035 compared with trends without this AI boost. Effects are concentrated in advanced economies, particularly the United States, although emerging and developing economies also benefit from increased economic activity.
Ultimately, what matters most for energy demand are energy policies and energy technologies. In our analysis, the impact of energy policies and energy technology developments on energy demand is much larger than the potential impact of an AI-driven economic growth boost.
Estimated energy demand impact of AI-induced GDP growth, 2035
Energy savings potential from AI, 2035
In an unstable world, the links between AI and energy security have tightened
Over the past year, energy and technology supply chains have become further strained. Trade restrictions have targeted key components going into data centres, such as the critical minerals needed for advanced power electronics, or the batteries and battery components needed to smooth AI loads. A 70% surge in gas turbine orders in 2025 has highlighted chokepoints in energy technology supply chains. And data centres themselves have been targeted in conflict zones, underscoring their role as critical infrastructure. The broader implications of the energy crisis tied to the conflict in the Middle East are still unknown, but they could extend to the choices that countries and companies make about the fuels and technologies used to power data centres, and where sites are located.
AI has the potential to be an important tool to enhance energy security and sustainability. For example, AI technologies monitor grids, transformers and other energy equipment to reduce unexpected failures and outages, and AI and digital grid-enhancing technologies are key to optimising the use of existing grid capacity, helping to offset lengthy and costly grid expansions. At the same time, AI risks making cyberattacks easier and more powerful, and an increasingly digitalised energy sector presents new vulnerabilities.
With increasing deployment of AI-enabled robotics, automation and efficiency solutions, AI will be critical to industrial competitiveness. Advances in data, models and hardware now allow for broader automation in industrial design and production, speeding up development, boosting innovation and lowering costs. The AI-enabled optimisation of production processes could reduce energy costs by 3-10 percentage points in energy-intensive industries, where energy is a critical production input and margins tend to be low. Well-documented AI use cases have the potential to save over 13 exajoules (EJ) of energy by 2035, equivalent to 3% of global final energy consumption, if barriers to wider uptake are overcome. Longer-term, AI could unlock another wave of industrial innovation and productivity: the almost two-fold jump in venture capital going into this field in 2025 is a marker of the potential. The race to develop more powerful AI models is being run alongside another race to apply AI across the economy for innovation and productivity, with China particularly focused on the latter.
But the energy sector is not yet taking full advantage of AI opportunities
An IEA survey of energy companies reveals that the lack of digital skills is the single largest barrier to greater AI adoption in the energy sector. The adoption of AI is also limited by fragmented data and concerns related to data protection, privacy, and cybersecurity. For example, only 10% of global electricity consumption is covered by open electricity data policies. The lack of adequate digitalisation of equipment can also be a barrier. Globally, less than half of energy demand is covered by policy frameworks aiming to promote the uptake of AI in the energy sector.
Share of policy coverage in energy demand, 2026
- Covered by policy
- Not covered
Share of global stock that is digitally connected, 2026
- Digitally connected
- Non-connected
Social concerns around AI have grown, focusing particularly on the environment and electricity prices…
Data centres are a highly visible flashpoint for concerns around energy prices and the environment. Energy prices have risen up the agenda since the inflationary shock of Covid-19 and the 2022 global energy crisis, with the energy crisis sparked by the conflict in the Middle East compounding concerns. In addition to the potential impact of data centres on electricity prices, surveys increasingly show that citizens are concerned about AI’s effect on jobs, the economy, the environment, and society more broadly. The emissions associated with data centres double in IEA projections, reaching around 350 million tonnes in 2035, but still make up about 2% of global electricity sector emissions by that date.
…but with the right conditions, data centres do not necessarily raise electricity prices
The impact of electricity demand growth on electricity prices depends on a combination of fundamental factors and policy choices. In systems with tighter demand-supply balances, load growth can indeed trigger a need for new investments, potentially raising prices. On the other hand, in systems with excess electricity supply, load growth can result in more efficient capital utilisation and lower prices. The shape of the load growth matters too. In the same way that high and reliable passenger numbers can allow airlines to lower average ticket prices on established routes, predictable baseload electricity demand can increase the efficiency of the use of capital-intensive power plants and grids, potentially lowering electricity prices.
Data centres can create special challenges for electricity affordability. Data centres are large, concentrated, rapidly developed infrastructure that are likely to trigger a need for new generation and grid investments in systems that host them. Actual peak load from data centres is often uncertain, as data centres are filled progressively with operating servers and often oversize their grid connections initially. The mismatch between fast-moving data centres and slow-moving energy investments, and the uncertainty over their load, creates risks of misalignment between electricity system investment and data centre demand that could raise electricity prices in some places if unmitigated.
Policy remains key to ensuring AI and data centres play a constructive role in energy systems
The IEA proposes three principles to help ensure AI deployment is leveraged for the benefit of the energy sector, and that data centres minimise any adverse impacts on electricity systems.
- Proactive management of data centre project pipelines and electricity sector investment can support adequate and reliable electricity for the sector without adversely affecting prices. This includes reforming the management of data centre connection queues and streamlining permitting timelines. Similarly, more robust demand projections, building on greater disclosure from technology companies and strong cooperation with system operators, can help ensure investments are well aligned. Finally, how costs are allocated in the electricity system is ultimately up to policymakers. Tools such as tariff schemes can support the fair allocation of the costs of grid upgrades and new generation capacity.
- Approaches that promote electricity system flexibility can help accelerate grid connections and ensure electricity affordability. This can help defer the need for large investments and improve the efficiency of expensive grid and generation investments. System operators can explore non-firm grid connections and incentivise data centre developers to provide demand response in return for faster connection processes. Increasingly sophisticated, grid-interactive onsite power assets, such as battery storage and gas-fired generators, can help data centres support grid operations, moving data centres from grid loads to grid resources.
- Removing barriers to AI adoption in the energy sector can ensure AI is leveraged to enhance energy security and sustainability. Comprehensive policy frameworks that address data availability, cybersecurity, skills, and interoperability are crucial for boosting AI uptake. As public concerns grow around the local impacts of data centres, and the broader implications of AI on jobs and fairness, social acceptability will hinge on demonstrating AI’s benefits for affordable, secure and sustainable energy systems, including at the local level.
General Document Comments 0
- This looks fairly trustworthy as a source for current energy-policy analysis because it comes from the International Energy Agency, a major intergovernmental body that publishes data-rich energy reports and updates them often ((IEA, “Key Questions on Energy and AI,” 2026)).
- It is not fully neutral, though, because the IEA has an institutional point of view: it exists to guide policy, frame risks, and influence decision-makers, so its choices about what to count, model, and emphasize can shape the story even when the data work is serious ((IEA, “Key Questions on Energy and AI,” 2026)).
- Its reliability is strongest for big-picture trends like data-centre demand, investment pressure, bottlenecks, and policy tradeoffs, but weaker for precise long-range forecasts because the report itself says AI uptake, efficiency, and energy use are changing fast and remain uncertain ((IEA, “Executive Summary – Key Questions on Energy and AI,” 2026)).
- A good way to judge it is to separate its observable evidence from its projections and policy framing: facts about what happened in 2025 are usually more solid than claims about what may happen by 2030 or 2035 ((IEA, “Energy Demand from AI,” 2025)).
A strong note you could write with Your Thoughts is this: which part of the report feels most evidence-based to you, and which part feels most shaped by forecasting or policy goals—or would you rather click Continue with AI to test that idea, or go to Create Document to upload another source and compare them?
- Governments usually do not “control” this kind of power all at once; they break it into parts like energy access, grid connections, competition, disclosure, safety, and procurement rules ((IEA, “Key Questions on Energy and AI,” 2026)).
- The EU is already doing this with the AI Act by assigning enforcement to the European AI Office and national authorities, with different rules phasing in through August 2, 2026, and after that for some systems ((European Commission, “AI Act,” 2026)).
- In the U.S., one likely path is less direct control and more standards, audits, procurement pressure, and sector-by-sector oversight, since NIST’s AI Risk Management Framework is voluntary guidance rather than a single national AI law ((NIST, “AI Risk Management Framework,” 2026)).
- For the power problem you noticed, governments can also slow or shape company growth through utility regulation, interconnection queues, permitting, tariff design, and rules about who pays for grid upgrades when giant data centers arrive ((IEA, “Key Questions on Energy and AI,” 2026)).
One sharp note you could write with Your Thoughts is this: is the bigger issue here AI itself, or the fact that five firms can shape electricity, capital, and infrastructure at a global scale—or would you rather click Continue with AI to test that claim, or go to Create Document to compare this report with another source?
on the uploaded document.
0 archived comments