Every chatbot reply, generated image and automated workflow begins in a warehouse-sized building filled with processors, cooling systems and electrical equipment. The artificial intelligence boom is often described in terms of software and algorithms, but it is also an enormous construction and energy project. In 2026, the physical footprint of AI has become one of the most consequential economic and environmental trends in the world.
The scale is already large. BloombergNEF estimates that global data centre capacity reached 84 gigawatts in 2025, consuming around 500 terawatt-hours of electricity, about 1.9 per cent of all electricity demand worldwide. That share may sound modest, but it is concentrated in specific regions, grows quickly and arrives on top of other sources of rising demand, including electric vehicles, heat pumps and industrial electrification. Where data centres cluster, they can dominate local grid planning. Utilities in some regions report requests for connections that dwarf the capacity they have built in decades.
The investment behind this is a macroeconomic force in its own right. The International Monetary Fund credits AI-related spending as a significant support for US growth, and it notes that several Southeast Asian economies are benefiting from data centre activity and from the upswing in the global technology cycle. Construction, electrical equipment, cooling technology, semiconductors, land and skilled labour all feel the effects. This is one of the reasons the IMF describes global growth as being reallocated according to AI-related demand strength.
But power is the bottleneck. Chips can be manufactured, and buildings can be erected relatively quickly, while new power plants, transmission lines and grid connections often take years to permit and construct. This mismatch is pushing technology companies into the energy business. They are signing long-term contracts for wind, solar and nuclear power, investing in battery storage, experimenting with on-site generation, and in some cases funding the restart or extension of existing plants. Some developers are choosing locations based on available electricity and cool climates rather than proximity to customers, since training large models does not require being near users.
The energy crisis stemming from the Strait of Hormuz closure complicates the picture. High fuel and gas prices raise electricity costs in import-dependent regions and sharpen debates about who should pay for new infrastructure. If households see their bills rise while data centres secure priority access, political backlash is likely. Regulators in several jurisdictions are already examining special tariffs for very large power users and requiring them to bring their own new supply.
Supporters of the AI build-out argue that the technology will pay back its energy cost. Efficiency gains in chips and software are steady, and many AI applications may reduce emissions elsewhere by optimising logistics, buildings, grids and industrial processes. They also point out that data centre demand can encourage investment in clean generation: large, predictable buyers are helpful in financing solar, wind and storage projects. BloombergNEF’s outlook sees solar becoming the world’s largest source of electricity by 2032, helped by very low prices, which could make clean supply easier to find for new loads.
Sceptics raise several concerns. First, the speed of demand growth may outrun clean supply, leading to more gas or coal use in the short term. Second, water use for cooling can strain local resources in dry regions. Third, the economic returns from AI may be uncertain; if demand disappoints, some of the capacity being built could stand underused, and communities and utilities could be left with costly commitments. The IMF itself expects the AI technology cycle to stabilise rather than keep accelerating, and it flags a sharp reassessment of AI valuations as a possible risk to global markets.
Efficiency is therefore a central battleground. Newer chips deliver more computation per unit of electricity, and techniques such as smaller specialised models, better scheduling and improved cooling reduce waste. The rise of AI agents, which plan, call tools and retry tasks, increases the amount of computation per task, so efficiency improvements are partly offset by heavier use. Economists call this a rebound effect: when something gets cheaper, people use more of it.
What should readers watch? Look for whether announced data centre projects actually connect to the grid on schedule, since many proposals never get built. Track the mix of power sources feeding them and whether new clean capacity is truly additional or simply redirected from other customers. Notice how regulators handle cost allocation, and whether communities receive benefits such as grid upgrades or jobs in return for hosting facilities.
The AI boom is teaching an old lesson in a new setting: digital technologies rest on physical foundations. The countries and companies best prepared to supply abundant, reliable, affordable power are likely to have an edge in the next phase of the technology. Those that cannot may find that their ambitions are limited not by ideas or talent, but by electricity.
