AI’s infrastructure race is becoming a race for electricity

AI arrives as a line of text on a screen, but the service is built from heavy things. Behind the prompt sit advanced chips, high-density servers, cooling equipment, substations, transmission capacity and power generation. A product that feels weightless to its user can impose a multi-year delivery problem on an electricity system.

That is a reversal of the cloud era’s central promise. When Amazon Web Services launched storage and compute services in 2006, customers could build without first buying and sizing their own machines. Infrastructure became an on-demand service and, from the customer’s point of view, almost disappeared. AI retains that convenient interface while making the machinery behind it strategically important again.S08

The scale has moved quickly. The International Energy Agency estimates that global data-centre electricity use reached about 485 TWh in 2025. It estimates year-on-year growth of 17% for data centres overall and 50% for AI-focused facilities. Its central projection reaches around 950 TWh by 2030, while electricity consumption by AI-focused centres triples. Separately, it reports that the total capital spending of five large technology companies exceeded USD 400 billion in 2025; the projected 75% increase in 2026 is an estimate, not an AI-only spending total.S01

Investment data shows the same change from another angle. UN Trade and Development’s preliminary monitor put announced foreign investment in data centres above USD 270 billion in 2025, over one fifth of all greenfield project value. The value of newly announced semiconductor projects rose by 35%. The build-out is therefore not a single property boom. It is a linked expansion of compute, chipmaking and energy infrastructure.S02S03

The tension lies in the clocks. A software service can add users quickly, and a customer can request more cloud capacity without seeing the equipment behind it. A new power project cannot move that way. Google’s account of its energy procurement describes a purchase agreement as the start of design, engineering, permitting, construction and interconnection—not the finish. The IEA now identifies delays at those same physical interfaces. Demand and announced spending can therefore become visible well before the system is ready to serve them. That gap between commercial speed and infrastructure delivery is the organising problem for the rest of the report.S01S08S09

The first internet build-out was visibly physical: fibre routes, mobile towers and rooms full of servers. Cloud computing changed both the geography and the customer experience. Workloads concentrated in much larger campuses while software teams bought computing as a service. The machines did not vanish; responsibility for finding, powering and operating them moved behind the interface.S08S16

Power was becoming strategic well before generative AI. Google signed its first corporate power-purchase agreement in 2010 for 114 MW of wind generation in Iowa, on a grid serving its data centres. The company’s own account also makes the delivery sequence plain: a signed agreement can precede design, permitting, construction and grid connection by years. Electricity procurement was already part of cloud infrastructure; AI has increased the size and urgency of the obligation.S09

Cloud’s achievement was to separate the customer’s decision from the operator’s infrastructure work. A developer could rent storage or servers in minutes because a provider had already assembled the land, machines, networks and power. That model rewarded large, standardised campuses and a small group of operators able to plan across regions. It also created an illusion: because capacity felt elastic at the application layer, the power system beneath it could be mistaken for elastic too. In reality, a workload can move between cloud regions far more easily than a transmission line, substation or generator can be built. AI exposes that difference because its new loads are both concentrated and large.S08S09

The electricity record makes the progression visible. A later national-laboratory model estimates that United States data centres used 58 TWh in 2014 and 176 TWh in 2023, then places 2028 use in a 325–580 TWh range. What was once a specialist load inside a mature power market is now large enough to influence generation plans, transmission investment and the timing of other industrial connections.S04

The source vintages show why a long arc must preserve its methods. A 2016 Lawrence Berkeley National Laboratory report estimated US data-centre use at about 70 TWh in 2014 and described growth slowing sharply through the cloud-efficiency era. The later national model used 58 TWh for the same year before reaching 176 TWh in 2023. That is a revision to the historical estimate, not a fall in electricity use. A June 2026 update then gave a 2030 reference case of 649 TWh and a wider compounded range of 521–843 TWh. Those points belong in the same history, but they cannot be joined into one continuous series.S17S04S18

The global series also has a model vintage. The IEA’s 2025 Energy and AI report estimated 415 TWh of data-centre electricity use in 2024 and a 945 TWh Base Case for 2030. Its April 2026 update instead estimates 485 TWh for 2025 and gives a central projection of about 950 TWh in 2030. The current series is the better orientation, while the earlier values remain useful only when their vintage is explicit. The global share can still look modest, but that average hides the problem that matters: AI demand arrives in very large blocks, in particular places, and often on a much shorter timetable than new grid assets.S05S01

That is the longer arc. The digital economy spent three decades becoming lighter for the user and heavier behind the interface. AI has brought the weight back into view. Location now matters because compute, cooling, construction, grid access, equipment and approvals must arrive on one compatible timetable.

Three changes have landed together. AI workloads call for denser clusters; their campuses ask for very large blocks of reliable power; and governments increasingly treat advanced compute as a strategic capability. Data-centre development, semiconductor policy and electricity planning are therefore becoming one delivery question.

The equipment chain is already signalling strain. The IEA identifies tighter supply for gas turbines, transformers, advanced chips and other IT components, alongside delays in grid connections and approvals. None of these constraints is novel on its own. What is new is that the same projects need all of them at the same time. A campus can secure land and chips yet remain stranded by a substation, or secure power while waiting for cooling equipment and planning consent.S01

People, contracts and finance run on their own clocks too. In the United States, the Department of Energy now places skilled-workforce expansion alongside energy infrastructure and computing capacity, and its data-centre framework expects technology companies to pay for required power-delivery upgrades and invest in workforce development. A project therefore needs specialist contractors, electrical trades, materials, funded work packages and clear cost allocation—not only a capital announcement. If those owners and dates are missing, construction or connection can slip even when land, chips and demand are present.S02S23

Europe’s public programmes show how far the definition of AI infrastructure has widened. The European Commission says 19 AI Factories and 13 linked antennas are operational, with at least nine more AI-optimised supercomputers due to more than triple current EuroHPC AI capacity. Its InvestAI initiative is intended to mobilise EUR 200 billion, including a EUR 20 billion fund for as many as five AI gigafactories. The Commission describes each gigafactory in terms of more than 100,000 advanced processors, power capacity, reliable supply chains, networking and efficiency—not chips alone.S10S11

The campus cycle also meets a second industrial build-out upstream. UN Trade and Development reported that the value of newly announced semiconductor projects rose by 35% in 2025. Those factories expand the potential supply of compute, but they do not remove the campus constraint; they add another class of power-intensive, equipment-heavy project to national industrial plans. The relationship runs both ways. More processors make larger AI clusters possible, while confidence in future powered campuses supports the case for additional chip capacity. Governments are therefore trying to coordinate a chain that crosses semiconductor manufacturing, digital infrastructure and electricity rather than treating each as a separate sector.S03S10

Geopolitics sharpens the same dependency from the other direction. In December 2024 the US Bureau of Industry and Security announced controls covering 24 types of semiconductor-manufacturing equipment, three types of software tools and high-bandwidth memory, alongside 140 Entity List additions. The action explicitly linked those inputs to leading-edge semiconductor production and AI at scale. It is a regulatory action at a date, not a measure of China’s achieved capability. For the infrastructure race, its significance is narrower and concrete: access to chips and the tools that make them can alter which planned campuses can install frontier equipment, even after a site and power route exist.S20

Connection queues show how crowded the interface has become. More than 2,500 GW of generation, storage and large-load projects are waiting in queues worldwide. That figure is not a list of assets certain to be built; queues contain duplication and speculation. It does show that the administrative and engineering path onto the grid has become a scarce resource in its own right.S06

This changes how the build-out should be read. An announced campus value is evidence of intent, not delivered compute. A chip order is not a powered rack. A power-purchase agreement is not necessarily a new generator or a firm grid connection. The useful question is no longer how much has been announced, but which projects have closed the gaps between every layer of delivery.

The build-out is global, but there is no universal route to a working AI campus. Mature cloud markets are trying to add power to established clusters. China is moving computing demand across a national network of hubs. Europe is combining shared supercomputing with national infrastructure programmes. Gulf states are designing new campuses around sovereign partnerships and new generation. Singapore is linking additional capacity to efficiency and green-energy conditions.S07S10S13S14

In the United States, the scale of the load is the story. The later national model estimates a rise from 58 TWh in 2014 to 176 TWh in 2023, then places 2028 use in a 325–580 TWh range. That spread is large enough to change generation and network plans. The decisive evidence will be local: completed grid studies, funded network work, firm equipment dates and a route to energisation.S04

China is using geography as an organising tool. By June 2024, more than CNY 43.5 billion had been directly invested in eight computing hubs under the ‘east data, west computing’ project, and the government said those hubs had driven more than CNY 200 billion of wider investment. A state official separately reported that the programme had more than 1.95 million racks. The system is intended to send data from economically advanced eastern regions to computing capacity inland. It is a national attempt to match demand, networks, land and energy rather than leave every city to solve the problem alone.S13

Europe is combining a shared compute network with national competition for sites. France said more than EUR 109 billion of AI infrastructure projects were announced at the Paris summit in February 2025 and presented low-carbon electricity, high-voltage grid access, suitable sites and faster procedures as the offer. At EU level, the AI Factory network broadens access to compute, while InvestAI is intended to support much larger gigafactories. The European route is therefore as much about assembling power and permission as raising the processor count.S10S11S12

Europe is also building a measurement layer around the physical system. Commission Delegated Regulation (EU) 2024/1364 requires operators of data centres with at least 500 kW of installed IT power demand to report defined information and indicators, including installed IT power, total energy, water and reused heat, to a European database. Reporting does not create a grid connection or shorten an equipment lead time. It can make the starting point clearer by forcing operators to distinguish the IT boundary from wider facility consumption and by giving public policy a more consistent record of resource use.S19

The Gulf route begins with the ability to plan at campus scale. Abu Dhabi’s announced UAE–US AI Campus is designed for 5 GW, including a 1 GW Stargate cluster whose first 200 MW was expected to enter service in 2026. Saudi Arabia offers a useful operating counterpoint to these large plans: the Saudi Press Agency reported that operational data-centre capacity rose from 68 MW in 2021 to more than 440 MW in 2025, across more than 60 centres built by over 20 companies. Two Saudi announcements then show the planned ceiling: NEOM and DataVolt described a phased 1.5 GW project with a 2028 target for its first phase, while centre3 and HUMAIN described infrastructure capable of supporting up to 1 GW of AI workloads. Announced gigawatts and state-reported operating megawatts remain separate columns.S14S15S21S22

Singapore represents the constrained-city model. Its Green Data Centre Roadmap aims to make at least 300 MW of additional capacity available in the near term, with further growth tied to green-energy deployment. The policy treats power and resource efficiency as boundaries on capacity, not matters to settle after a campus has been promised.S07

The investment map remains more concentrated than the global announcement wave can suggest. UN Trade and Development identified France, the United States and the Republic of Korea as the leading data-centre investment destinations in 2025. They are different markets, but each can point to an established combination of digital demand, power-system capability, public coordination or industrial depth. That concentration is not permanent, and the Gulf programmes are designed to change it. It does show why the first question for a new geography is not whether leaders want AI capacity. It is whether the country can move a specific site through power, network and approval gates quickly enough to compete with places that already have a delivery base.S03S12S14S15

These models are not directly comparable: some figures describe national electricity use, some operating capacity, some project announcements and some public funding intentions. The comparison still reveals the common strategic problem. Each region is trying to create a place where compute equipment, electricity, networks and permission can meet. Cheap land on its own is not a strategy; a power-ready site is.

It is tempting to name one scarce input: chips, power, transformers, turbines or water. The evidence points to a more difficult answer. The bottleneck is the ability to sequence them. Semiconductor production can expand on one timetable, transmission on another and permitting on a third. A declared campus moves from plan to operation only when those clocks meet at the same site.S01S06

Grid access is especially awkward because the physical asset and the commercial right to use it are different things. A region may have abundant generation but no transmission path to the chosen site. A developer may hold a queue position without a finished network study. A utility may be willing to connect the load but unable to source transformers on the proposed schedule. Each condition can turn an apparently advanced project back into an option.S01S06

The queue number needs careful reading. The IEA’s total of more than 2,500 GW covers generation, storage and large-load projects waiting for connection, not a list of data centres certain to operate. Requests can overlap, change or be withdrawn. The figure is useful because it measures congestion at a shared gateway: engineering studies, network design and commercial allocation. It cannot say which campus will be served, in what order or by what date. For that, a project needs its own completed studies, agreement and funded works.S06

The delivery model also spans different planning horizons. Chips and servers follow a fast technology cycle, while substations, transmission and generation are planned as long-lived parts of a public system. A customer may want flexibility about when and how much capacity it uses; a utility needs enough certainty to build assets that remain after a server generation has changed. The resulting questions are practical and public: who funds dedicated works, what happens if the load arrives late, and whether the campus can reduce or shift demand when the grid is tight.S06S09

Even the unit of measurement can hide the mismatch. One announcement may quote IT load, another total facility load and another contracted power. A gigawatt described at campus level may include later phases that have no near-term opening date. Electricity-demand forecasts use energy over a year, while connection decisions also depend on peak demand at a particular time and place. Unless those boundaries and dates remain visible, unlike projects can appear equivalent and planned capacity can be mistaken for delivered capacity. Clear definitions are therefore part of the delivery evidence, not a footnote to it.S01S04S05S14

That is why this cycle is best read as a delivery system. Technology firms, utilities, network operators, governments and equipment suppliers depend on one another’s schedules. Regions that solve the coordination problem can turn AI demand into operating capacity. Those that do not may collect announcements, queue positions and incomplete infrastructure instead.

The next stage will not be confirmed by a larger total of press releases. It will be confirmed by projects crossing hard gates: contracted chips, approved grid studies, funded substations, ordered long-lead equipment, secured water and cooling, and a credible route to energisation. Those records show when strategic intent is becoming operating capacity.S01S06

The most useful indicators sit at the junctions between industries. Watch whether equipment lead times ease, whether connection queues move rather than merely grow, and whether new power is genuinely additional. Also watch the geography of completed projects. If investment keeps concentrating in a few countries despite a global announcement wave, power-system readiness—not demand for AI—will have decided the map.S02S03S06

It helps to separate leading, intermediate and lagging evidence. Announcements, funding intentions and queue applications are leading signals of demand, but they can still disappear. Planning consent, completed grid studies, connection agreements and firm equipment orders are intermediate evidence that a project is surviving the delivery process. Commissioned buildings, energised megawatts and operating electricity use are lagging confirmation. No single layer is enough on its own: a market with only lagging data is slow to read, while a market judged only by announcements can look larger and faster than it is. The useful record follows the same project across all three stages.S01S02S06S09S14S15

Efficiency is the most important counterweight to the physical-constraint thesis. If chips, cooling and utilisation deliver much more useful compute per unit of electricity, the upper end of today’s power ranges may not arrive. If total demand still doubles while efficiency improves, however, grid access remains central. A second test is diffusion: if regions with limited power-system readiness begin completing large campuses as quickly as established markets, the strategic value of the power-ready site has been overstated. Those are genuine weakening conditions, not caveats added after the fact, and the report’s interpretation should move if the evidence does.S01S04S05