
Artificial intelligence feels weightless. You type a question, an answer appears, and nothing seems to have been consumed in between. But every response runs on hardware in a building somewhere, and that building runs on electricity.
The relationship between AI and energy is now one of the defining infrastructure stories of the decade. AI is becoming a major new source of electricity demand. It is also becoming a powerful tool for running energy systems more cleanly and efficiently. Both things are true at once, and the balance between them will shape who benefits from this technology and who is left out.
At SELF, we watch this story from a particular vantage point. Close to a billion people still live without reliable electricity. As the world races to power ever larger data centers, that gap deserves to be part of the conversation.
Why AI uses so much energy
Training a large AI model means adjusting billions of parameters through repeated calculations. That work runs on thousands of specialized chips, such as GPUs and TPUs, operating in parallel for weeks or months at a time. Models also need frequent retraining to stay current, which adds to the total.
The scale is hard to picture. A typical AI-focused data center consumes as much electricity as 100,000 households. The largest facilities now under construction will consume about 20 times that. Only a handful of companies can afford to build and run infrastructure at this level, because the costs of hardware, power, cooling, and maintenance are so high.
The numbers add up quickly. According to the International Energy Agency (IEA), data centers used around 1.5 percent of the world’s electricity in 2024, about 415 terawatt-hours. The United States accounted for 45 percent of that, followed by China at 25 percent and Europe at 15 percent. Since 2017, data center electricity use has grown about 12 percent per year, more than four times faster than electricity consumption overall.
The share is still small globally, but local effects can be large. Data centers cluster geographically, and nearly half of U.S. capacity sits in just five regional clusters. In those places, a single industry can reshape the local grid.
How much more energy will AI need?
The IEA projects that global data center electricity consumption will more than double to around 945 terawatt-hours by 2030. By 2035, its base case reaches roughly 1,200 terawatt-hours, though the range across scenarios is wide, from 700 to 1,700.
In the United States, researchers at Penn State note that data centers consumed 4.4 percent of national electricity in 2023, a figure that could triple by 2028. The IEA expects U.S. data centers to account for nearly half of national demand growth through 2030.
Those projections carry real uncertainty. Much depends on how quickly AI is adopted, how fast models and chips become more efficient, and whether energy infrastructure can keep pace. The IEA estimates that around 20 percent of planned data center projects could face delays if grid constraints go unaddressed. Building new transmission lines can take four to eight years in advanced economies, and wait times for transformers and cables have doubled in three years.
Where the power will come from
Meeting this demand will require a mix of sources. The IEA expects renewables to supply about half of the growth, supported by storage and the broader grid, with renewable generation growing by over 450 terawatt-hours to 2035. Natural gas and nuclear each contribute roughly 175 terawatt-hours of additional generation over the same period, and the first small modular reactors are expected around 2030.
Renewables have an edge for good reasons. They deploy quickly, they are cost-competitive, and technology companies are increasingly buying clean power directly. Solar in particular can be built at almost any scale, from a rooftop to a utility-scale plant. That flexibility matters well beyond the data center boom, and it is the principle behind our work on affordable and clean energy for communities that the grid has never reached.
The environmental footprint beyond electricity
Electricity is only part of the picture. Researchers point to several other pressures.
Water. Cooling AI hardware takes large volumes of water, which can strain supplies in regions already facing scarcity. The link between power and water runs in both directions, which we explore in the connection between water security and energy infrastructure.
Emissions. The IEA estimates that emissions from data center electricity will rise from about 180 million tonnes today to 300 million tonnes by 2035 in its base case, and up to 500 million tonnes if AI adoption accelerates. That is still under 1.5 percent of energy sector emissions, but data centers are among the fastest growing sources.
E-waste and materials. AI chips have short useful lives, and discarded hardware adds to the electronic waste stream. Manufacturing also depends on rare earth and critical minerals, with supply chains concentrated in a few countries.
How AI is helping transform the energy system
The same technology that strains the grid can also help run it better. The IEA identifies several areas where AI is already delivering value.
Smarter grids. AI-based fault detection can pinpoint problems quickly and cut outage durations by 30 to 50 percent. Sensors combined with AI management can increase the capacity of existing transmission lines, potentially unlocking up to 175 gigawatts without building anything new. That is more than the projected rise in data center load to 2030.
Better renewable integration. Solar and wind output varies with the weather. AI can forecast generation more accurately, which reduces curtailment and helps grid operators balance supply and demand. This matters most as grids become more decentralized and complex.
Efficiency in buildings, industry, and transport. Applying existing AI tools to heating and cooling could save around 300 terawatt-hours of electricity globally. Industrial optimization and smarter transport could save even more, though the IEA cautions that rebound effects, such as cheaper autonomous vehicles drawing riders away from public transit, could erode some gains.
Faster discovery. The IEA notes that only 0.01 percent of next-generation solar materials have been produced and tested experimentally. AI can help researchers screen candidates for better solar cells, batteries, and carbon capture molecules far faster than traditional methods. Yet only 2 percent of equity raised by energy start-ups has gone to AI-focused companies, so the opportunity remains largely untapped.
For a broader look at how these tools apply to the climate crisis, see our piece on AI and climate change.
Making AI more sustainable
The research points to practical steps that can reduce AI’s footprint without halting progress.
- Efficient models. Domain-specific models built for fields like healthcare or chemistry need far less computing power than general-purpose models trained from scratch.
- Better hardware. Specialized accelerators, including neuromorphic and optical chips, could deliver large efficiency gains over today’s GPUs.
- Clean power at the source. Moving data centers onto solar and wind reduces fossil fuel dependence, though storage and grid adaptation remain challenges.
- Smarter scheduling. Shifting computing workloads across time zones can align demand with periods of peak renewable output.
- Flexible operation. Locating data centers where grids have spare capacity, and running servers or backup storage more flexibly, can ease pressure on constrained networks.
None of these is a silver bullet. The IEA estimates that widespread use of existing AI applications could reduce energy-related emissions by around 5 percent in 2035. That is meaningful, but far short of what climate goals require. AI can support the transition. It cannot replace the policy and investment needed to deliver it.
The question of who gets left out
Here is the part of the story that gets the least attention. The IEA reports that emerging and developing economies outside China account for half of the world’s internet users but less than 10 percent of global data center capacity.
One reason is power reliability. In regions with frequent outages or poor power quality, running a data center is risky and expensive, so businesses host their computing overseas. That leaves local innovators dependent on infrastructure built elsewhere, with all the cost, latency, and control issues that implies.
This is energy poverty showing up in a new form. We have long documented how unreliable electricity limits health, education, and income. The AI era adds another layer: Communities without dependable power cannot participate in the digital economy being built around this technology. When solar electrification of schools brings computers and connectivity to classrooms, it is closing part of that gap. The same applies to clinics that can run diagnostic tools, farmers who can access forecasts, and entrepreneurs who can reach online markets.
The IEA does see an opening. Developing economies can leapfrog straight to AI-enabled solutions, skipping legacy approaches, if the basics are in place. Reliable, affordable electricity is the first of those basics.
Why decentralized solar belongs in this conversation
Large data centers will draw on large grids. But the lesson for energy access is different. Grids that are centralized, aging, and slow to expand struggle to serve remote communities. They are also vulnerable. SELF has written about how decentralized energy can rebuild resilience after disasters destroy central infrastructure.
Community-scale solar offers a practical alternative to large grids. Community-based solar gardens put generation close to where power is used. Community-owned solar gives residents a stake in the system and a role in maintaining it. Together with storage, these systems can deliver the kind of dependable power that digital tools require, and they are an important part of fighting energy poverty through solar solutions.
AI can help here too. Better forecasting, smarter load management, and predictive maintenance can make small systems cheaper to run and easier to keep working. For off-grid communities, that could mean longer equipment life and more prosperity.
What needs to happen next
The IEA’s conclusion is that the energy and technology sectors are now deeply intertwined, and that deeper collaboration between them is essential. We would add a few priorities from our own vantage point.
Treat access as part of the AI agenda. Plans for powering AI should not be separate from plans for the people who still lack electricity. Investment that serves both is possible and should be sought out.
Prioritize clean, flexible supply. Renewables paired with storage can meet new demand without locking in decades of additional emissions.
Support decentralized solutions. Community-scale systems reach people that large grids will not serve for years, if ever.
Build skills. The IEA notes that AI skills are less common in the energy sector than elsewhere. Training local technicians and engineers helps communities operate and benefit from the systems they depend on.
The bottom line
AI is changing the global energy system in two directions at once. It is creating major new demand for electricity, and it is offering new tools to produce, distribute, and use energy more intelligently. How that balance plays out depends on choices being made now by governments, companies, and communities.
The most important choice may be who is included. A world that powers its data centers while billions still lack basic electricity will have missed something essential. Energy access is not a side issue to the AI era. It is the foundation that determines who can take part in it.




0 Comments