Artificial intelligence is becoming an energy-security challenge as data centres built to train and run advanced models place mounting pressure on electricity grids, power equipment supply chains and climate targets across the world’s largest technology markets.
Global data-centre electricity consumption is on course to more than double to about 945 terawatt-hours by 2030, a level close to the current annual electricity use of Japan. AI-focused facilities are expanding faster than the wider data-centre sector because frontier model training, inference workloads and cloud-based enterprise tools require dense clusters of high-performance chips running around the clock.
The strain is most visible in the United States, where Microsoft, Amazon, Alphabet and Meta are spending hundreds of billions of dollars on AI infrastructure while utilities struggle to connect large new loads quickly enough. Some proposed campuses require power at the scale of heavy industrial plants, with individual facilities approaching or exceeding one gigawatt, enough to supply hundreds of thousands of homes.
Grid bottlenecks have become a commercial risk for the AI industry. Developers face delays in securing interconnection approvals, transformers, turbines, skilled workers and transmission upgrades. Power demand is rising after two decades of relatively flat consumption in many mature economies, forcing regulators and utilities to decide who pays for network expansion and how costs should be shared between technology companies, households and other businesses.
The result is a shift in strategy among the largest AI players. Rather than relying only on conventional utility connections and renewable energy certificates, technology groups are now seeking dedicated power arrangements, direct generation partnerships and long-term contracts for firm electricity. Microsoft has backed a plan to restart a unit at the Three Mile Island nuclear site in Pennsylvania through a long-term power purchase agreement. Google has signed an advanced nuclear agreement with Kairos Power, with the first deployment targeted for 2030 and a broader goal of 500 megawatts by 2035. Meta has sought proposals for 1 to 4 gigawatts of new nuclear generation capacity in the United States.
Natural gas is also gaining ground as a near-term option because it can be deployed faster than nuclear power and can provide continuous supply when renewable generation fluctuates. Microsoft has entered an exclusivity arrangement with Chevron and Engine No. 1 for potential gas-fired power supply to support data-centre demand, while energy groups are examining liquefied natural gas investments linked to AI infrastructure growth. These moves underline the tension between corporate net-zero pledges and the practical need for reliable electricity.
Europe faces similar constraints, although the geography is different. A 200-megawatt grid connection agreement in northern Sweden for an AI data-centre project shows how developers are moving towards regions with cooler climates, cleaner power mixes and available grid capacity. Ireland, the Netherlands and parts of Germany have already confronted public debate over whether data centres are crowding out housing, industry and local climate objectives.
The problem is not only the total amount of electricity consumed. AI workloads can create sharp shifts in demand, particularly when model training runs are started, paused or relocated. This makes the integration of batteries, backup systems, workload scheduling and grid-interactive data centres more important. Operators are exploring ways to shift non-urgent computing tasks to periods of lower demand or higher renewable output, but real-time services such as search, coding assistants, translation and business automation require constant availability.
Efficiency gains remain the main counterweight. Chipmakers are improving performance per watt, cloud companies are designing custom processors, and cooling systems are becoming more sophisticated. Liquid cooling, waste-heat reuse and advanced power management can reduce the energy intensity of computing. Yet efficiency alone may not solve the problem if demand for AI services grows faster than improvements in hardware and software.
Public policy is now moving into the centre of the debate. Governments want AI investment because it promises productivity gains, national-security advantages and high-value jobs, but grid planning cycles often run slower than data-centre construction timelines. Transmission lines can take years to approve, large transformers face long lead times, and local opposition can slow both power and data infrastructure.
Global data-centre electricity consumption is on course to more than double to about 945 terawatt-hours by 2030, a level close to the current annual electricity use of Japan. AI-focused facilities are expanding faster than the wider data-centre sector because frontier model training, inference workloads and cloud-based enterprise tools require dense clusters of high-performance chips running around the clock.
The strain is most visible in the United States, where Microsoft, Amazon, Alphabet and Meta are spending hundreds of billions of dollars on AI infrastructure while utilities struggle to connect large new loads quickly enough. Some proposed campuses require power at the scale of heavy industrial plants, with individual facilities approaching or exceeding one gigawatt, enough to supply hundreds of thousands of homes.
Grid bottlenecks have become a commercial risk for the AI industry. Developers face delays in securing interconnection approvals, transformers, turbines, skilled workers and transmission upgrades. Power demand is rising after two decades of relatively flat consumption in many mature economies, forcing regulators and utilities to decide who pays for network expansion and how costs should be shared between technology companies, households and other businesses.
The result is a shift in strategy among the largest AI players. Rather than relying only on conventional utility connections and renewable energy certificates, technology groups are now seeking dedicated power arrangements, direct generation partnerships and long-term contracts for firm electricity. Microsoft has backed a plan to restart a unit at the Three Mile Island nuclear site in Pennsylvania through a long-term power purchase agreement. Google has signed an advanced nuclear agreement with Kairos Power, with the first deployment targeted for 2030 and a broader goal of 500 megawatts by 2035. Meta has sought proposals for 1 to 4 gigawatts of new nuclear generation capacity in the United States.
Natural gas is also gaining ground as a near-term option because it can be deployed faster than nuclear power and can provide continuous supply when renewable generation fluctuates. Microsoft has entered an exclusivity arrangement with Chevron and Engine No. 1 for potential gas-fired power supply to support data-centre demand, while energy groups are examining liquefied natural gas investments linked to AI infrastructure growth. These moves underline the tension between corporate net-zero pledges and the practical need for reliable electricity.
Europe faces similar constraints, although the geography is different. A 200-megawatt grid connection agreement in northern Sweden for an AI data-centre project shows how developers are moving towards regions with cooler climates, cleaner power mixes and available grid capacity. Ireland, the Netherlands and parts of Germany have already confronted public debate over whether data centres are crowding out housing, industry and local climate objectives.
The problem is not only the total amount of electricity consumed. AI workloads can create sharp shifts in demand, particularly when model training runs are started, paused or relocated. This makes the integration of batteries, backup systems, workload scheduling and grid-interactive data centres more important. Operators are exploring ways to shift non-urgent computing tasks to periods of lower demand or higher renewable output, but real-time services such as search, coding assistants, translation and business automation require constant availability.
Efficiency gains remain the main counterweight. Chipmakers are improving performance per watt, cloud companies are designing custom processors, and cooling systems are becoming more sophisticated. Liquid cooling, waste-heat reuse and advanced power management can reduce the energy intensity of computing. Yet efficiency alone may not solve the problem if demand for AI services grows faster than improvements in hardware and software.
Public policy is now moving into the centre of the debate. Governments want AI investment because it promises productivity gains, national-security advantages and high-value jobs, but grid planning cycles often run slower than data-centre construction timelines. Transmission lines can take years to approve, large transformers face long lead times, and local opposition can slow both power and data infrastructure.
Topics
Technology