The artificial intelligence revolution has an enormous and rapidly growing energy footprint. Training large language models, running inference on AI queries, and powering the cloud infrastructure that underpins AI applications requires vast quantities of electricity — and the demand is accelerating faster than most grid planners and energy analysts anticipated. In 2026, the intersection of AI and energy has become one of the most discussed topics in both the technology and energy sectors, with implications for electricity prices, grid infrastructure, renewable energy procurement, and carbon emissions worldwide.

The Scale of Data Centre Power Demand

Data centres — the facilities that house the servers, networking equipment, and cooling systems required for digital services — consumed approximately 400–500 terawatt-hours (TWh) of electricity globally in 2025, representing around 1.5–2% of global electricity consumption. This figure is projected to roughly double by 2030, driven primarily by AI workloads. The IEA has described data centre electricity demand as one of the fastest-growing components of global power consumption, with AI-specific workloads growing at annual rates of 30–50% or more.

The energy intensity of AI is considerably higher than conventional computing. Training a large foundation model — such as those underlying major AI assistants and image generators — requires the equivalent of running thousands of high-performance chips simultaneously for weeks or months. The electricity consumption of a single training run for a frontier model can rival that of a small town over the same period. Inference — serving queries to users of deployed AI applications — runs continuously at massive scale and represents an even larger share of total AI electricity demand than training.

Geographic Hotspots

Data centre construction is concentrated in a relatively small number of geographic clusters, driven by land availability, power infrastructure, fibre connectivity, and regulatory environment. In the United States, Northern Virginia (the largest data centre market in the world), Phoenix, Dallas, Chicago, and the Pacific Northwest are primary hubs. Amazon Web Services, Microsoft, Google, and Meta are building hyperscale data centres at extraordinary pace, with construction pipelines that represent hundreds of billions of dollars in capital investment over the next five years.

Europe’s major data centre markets — Dublin (which hosts a disproportionate share of US tech companies’ European infrastructure due to its tax environment and English-speaking workforce), Frankfurt, Amsterdam, London, and Stockholm — are experiencing surging demand and, in some cases, electricity supply constraints that are limiting further construction. Amsterdam’s municipality effectively declared a moratorium on new data centre approvals in 2022 due to grid capacity concerns; Ireland’s grid operator has warned that data centre load growth could threaten system stability without major transmission investment.

In Asia, Singapore — long the data centre hub for Southeast Asia — has similarly imposed restrictions on new construction due to electricity and water constraints. Malaysia has emerged as an alternative, with Johor Bahru attracting massive investment from Chinese and US technology companies. Japan, South Korea, and India are all experiencing rapid data centre growth as domestic cloud and AI demand increases.

The Power Procurement Challenge

Technology companies have made ambitious renewable energy commitments, and the hyperscalers — Amazon, Microsoft, Google, Meta — are among the world’s largest corporate buyers of renewable energy. Power purchase agreements (PPAs) signed by tech companies for renewable electricity have been a major driver of clean energy project development in multiple markets. Microsoft’s recent agreement to purchase power from the restarted Three Mile Island nuclear plant in Pennsylvania — the largest corporate nuclear power deal in history — illustrated the lengths to which major AI operators will go to secure large quantities of carbon-free electricity.

The challenge is that AI data centres require not just renewable energy on an annual average basis, but power that is available 24 hours a day, 7 days a week. Solar and wind are intermittent, generating only when the sun shines or wind blows. Matching AI’s constant power appetite with intermittent renewable generation requires either storage (expensive), transmission from complementary renewable regions (infrastructure-intensive), or firm low-carbon power sources such as nuclear, geothermal, or hydro. This is driving renewed corporate interest in nuclear energy — both existing plant restarts and new advanced nuclear technologies including small modular reactors (SMRs).

Grid Stress and Infrastructure Investment

The clustering of large data centres in specific locations is creating acute grid stress in those areas. In Northern Virginia, electricity utilities have been struggling to keep pace with load growth, with transformer and substation lead times extending to years due to global supply chain constraints. New transmission lines, substations, and generation capacity are all required to meet data centre demand — investment that takes years to plan, permit, and build.

Electricity prices in data centre hotspots are rising, and utilities are introducing special rate structures for large load customers to ensure cost recovery. Some data centre operators are exploring on-site generation — including natural gas backup generators (a significant source of local air pollution concern), rooftop solar, and even on-site nuclear reactors — as a way of reducing grid dependence and securing power supply certainty.

The Efficiency Response

Technology companies are investing heavily in improving the energy efficiency of their AI infrastructure, recognising that power costs represent a significant and growing operational expense. Advances in chip design — including purpose-built AI accelerators from NVIDIA, AMD, Google, and others — have dramatically improved the computational work performed per unit of energy. Cooling efficiency improvements — including liquid cooling systems that remove heat more efficiently than traditional air cooling — are also reducing the overhead of keeping servers at operating temperature.

The concept of Power Usage Effectiveness (PUE) — the ratio of total data centre energy consumption to IT equipment energy consumption — has improved significantly over the past decade, with leading hyperscale operators achieving PUEs approaching 1.1 (meaning only 10% overhead for cooling and other systems), compared to industry averages of 1.5–1.6 in older facilities. These efficiency gains are meaningful but are being more than offset by the explosive growth in the volume of AI computation being performed. Explore more on the energy transition’s technology dimension in our AI & Energy section and the latest energy news.

Outlook

AI’s energy footprint will continue to grow rapidly through the rest of the decade. The combination of expanding model complexity, growing user bases, and proliferating AI applications across industries means electricity demand from AI workloads will be a major structural driver of power sector investment globally. For energy markets, this represents both a challenge — managing grid infrastructure to cope with concentrated, large-scale loads — and an opportunity, as large, creditworthy corporate buyers drive investment in clean energy generation and storage. The companies building AI are becoming some of the most important players in the clean energy transition — not by design, but by necessity.