The AI Race Is Becoming an Energy Race

The competition surrounding artificial intelligence is usually described through the technologies closest to the user: models, semiconductors, software, algorithms and the companies capable of assembling those components into increasingly powerful systems. Those are unquestionably important parts of the story, but the extraordinary rate at which computational infrastructure is expanding is exposing another layer of competition beneath them. The ability to develop advanced artificial intelligence increasingly depends upon the ability to construct and operate enormous amounts of physical infrastructure, which means that a contest commonly understood as a software and semiconductor race is becoming, in part, an energy race as well.

This shift follows directly from the production chain examined throughout this series. Machine intelligence requires computation, computation requires processors operating inside physical facilities, and those facilities require electrical infrastructure capable of supplying extraordinarily large and increasingly concentrated loads. Once the quantity of intelligence an economy can deploy begins to depend partly upon the quantity of computation it can economically operate, the power system beneath that computation becomes strategically relevant to technological capacity itself.

The International Energy Agency's updated 2026 outlook estimates that global data-center electricity consumption will rise from roughly 485 terawatt-hours in 2025 to about 950 TWh by 2030, while electricity consumption at AI-focused facilities grows substantially faster and roughly triples over the same period. The United States and China account for nearly 80 percent of the projected increase in global data-center electricity use through the end of the decade, placing much of the physical expansion of machine intelligence inside the world's two largest economies.

That concentration makes the issue larger than whether the technology industry can purchase enough electricity. It raises a broader question about national capacity: which countries can expand generation, transmission, compute infrastructure, semiconductor supply and industrial support systems quickly enough to sustain the next phase of machine intelligence without allowing one layer of the system to become the bottleneck for all the others?

Intelligence Is Becoming an Infrastructure Business

One of the most consequential characteristics of artificial intelligence is that the quantity of usable machine cognition is not fixed. Additional processors can perform additional computation, and additional computational infrastructure can serve more models, agents, simulations and automated systems. Improvements in hardware and software continue to reduce the resources required for many individual tasks, but aggregate demand has nevertheless expanded rapidly because the number and sophistication of economically useful applications have grown even faster.

The IEA reported that electricity consumption from data centers increased approximately 17 percent during 2025 while consumption associated with AI-focused facilities grew about 50 percent, far exceeding the roughly 3 percent increase in global electricity demand overall. At the same time, five major technology companies collectively spent more than $400 billion on capital investment in 2025, with the IEA estimating that their expenditures could rise another 75 percent in 2026.

Those capital flows reveal how dramatically AI is moving from a software story into an infrastructure story. The investment does not stop at processors and buildings because every new accelerator ultimately requires electricity, every facility must connect to a grid or secure another source of power, and every cluster of facilities increases demand for transmission systems, substations, transformers, cooling equipment and generation assets that were often planned long before contemporary AI loads existed. The strategic significance of energy therefore arises not because electricity replaces chips, models or talent as the determinant of technological leadership, but because it has become one of the physical constraints through which all of those other advantages must ultimately pass.

The Competition Is Moving Down the Stack

The earliest public discussion surrounding the AI race concentrated heavily on model capability because improvements were visible and comparatively easy to compare. One system could write better software, solve more difficult mathematics or process larger contexts than another, while benchmark scores provided a convenient shorthand for technological progress.

As the industry has grown, however, the underlying competition has moved steadily downward through the technology stack. Advanced models require accelerators, which created extraordinary demand for GPUs and specialized memory; semiconductor demand increased the strategic importance of fabrication capacity and manufacturing equipment; enormous clusters of accelerators created demand for data centers and networking systems; and those facilities are now placing unprecedented pressure upon electricity systems. The IEA's 2026 assessment describes a simultaneous scramble for electricity, grid connections, manufacturing capacity, chips and capital, while identifying tightening supplies of high-bandwidth memory, transformers, gas turbines and other infrastructure as constraints on continued expansion.

This progression matters because the location of the bottleneck determines what the economy must build next. Scarcity of advanced chips makes computational capacity expensive, while shortages of transformers can leave completed data centers waiting for grid connections. When gas turbines have multi-year delivery schedules, developers cannot respond to projected power shortages simply by ordering new dispatchable generation and expecting it the following year. When transmission projects require considerably longer than the data centers seeking connection, a region may possess proposed generation and investment capital yet still fail to convert either into usable compute.

The AI stack has consequently expanded well beyond what is conventionally considered the technology industry. Electrical engineering, utility planning, turbine manufacturing, nuclear engineering, natural-gas infrastructure, power electronics, copper supply, transformers, construction and transmission development are all becoming part of the same capacity-building problem.

The Grid Has Become a Compute Constraint

Perhaps nowhere is the mismatch between digital ambition and physical infrastructure clearer than in electrical grids. The IEA estimates that more than 2,500 gigawatts of proposed generation, storage and large-load projects are currently stalled in grid-connection queues around the world. It also estimates that annual global grid investment will need to increase by approximately 50 percent from today's roughly $400 billion level by 2030 to accommodate anticipated demand and generation growth.

The most consequential mismatch concerns timing, since a data center may be built within roughly one to three years, while major grid infrastructure can require five to fifteen years to plan, permit and complete. Prices for important components have also risen while manufacturing lead times for transformers, cables and related equipment have lengthened.

This creates a technological constraint with an unusual character: private capital can now assemble computational facilities faster than electrical systems can necessarily accommodate them. A technology company can secure land, order servers, finance construction and develop software on timelines substantially shorter than those required for a utility to complete a new high-voltage transmission corridor. The resulting scarcity tends to appear regionally rather than as a national shortage of electricity, particularly where transmission capacity has already been absorbed, substations cannot accept another large connection or the incremental load requires system upgrades whose completion dates extend beyond the investment timetable of the proposed facility.

That distinction reinforces a principle established earlier in this series: energy abundance has to be measured where power can actually be delivered. A country can possess enormous generating capacity while still constraining AI infrastructure if the network cannot move sufficient electricity toward the computational clusters consuming it. The technology race therefore includes a grid-building race whether policymakers describe it that way or not.

China Is Building at the Scale of an Electrifying Industrial Economy

China provides a useful example of why the energy dimension of technological competition cannot be separated from broader industrial development. According to China's National Energy Administration, the country ended 2025 with approximately 3.89 billion kilowatts, or 3.89 terawatts, of installed generating capacity, representing growth of 16.1 percent from the previous year. Solar capacity reached roughly 1.2 TW while wind capacity reached about 640 GW.

Installed capacity and actual electricity production are not interchangeable, particularly when comparing resources with different utilization rates, but the magnitude of the buildout demonstrates the scale at which China is expanding the physical system beneath an increasingly electrified economy. The IEA reports that Chinese electricity consumption surpassed 9,500 TWh in 2025 and expects the country to add approximately 2,600 TWh of annual electricity demand by 2030, an increase roughly comparable to the European Union's current total electricity consumption.

Only part of that growth comes from data centers, while industrial production, transportation, buildings, air conditioning and wider electrification account for enormous portions of Chinese demand. That broader context is precisely why the comparison is relevant: AI infrastructure does not operate on a separate computational grid but competes inside an economy where factories, vehicles, households and other industries are simultaneously increasing their consumption.

China's electrical expansion includes large additions of solar, wind, nuclear and other generation, while coal remains a substantial component of the power system. The IEA expects nearly 30 GW of additional Chinese nuclear capacity to enter service between 2026 and 2030, while renewable generation continues growing rapidly.

The significance for AI competition lies less in the performance of any single generating technology than in China's willingness and ability to expand the system as a whole. Large-scale intelligence infrastructure rests upon large-scale electrical infrastructure, and the latter must be built before the former can be sustained at scale.

The United States Is Re-entering an Era of Electricity Growth

The United States approaches the same challenge from a different historical position. American electricity consumption remained comparatively stagnant for roughly fifteen years, which encouraged utilities, regulators and investors to plan within an environment where rapid load growth no longer appeared normal. AI, advanced manufacturing and other new electrical loads have disrupted that assumption.

The IEA expects U.S. electricity demand to grow by nearly 2 percent annually through 2030, more than twice the rate of the previous decade, with data centers accounting for roughly half of the total increase. The latest September 2026 forecast from the U.S. Energy Information Administration similarly projects record electricity sales, reaching approximately 4,135 billion kilowatt-hours in 2026 and 4,211 billion kWh in 2027, with data-center development and increased manufacturing activity among the principal drivers.

The generation system is responding through a substantial construction program, with EIA reporting that developers planned a record 86 GW of new utility-scale generating capacity for 2026, if all scheduled projects are completed, following 53 GW of additions in 2025. Solar represents slightly more than half of the planned 2026 additions, while battery storage and wind account for much of the remainder. Utility-scale battery capacity reached nearly 52 GW by June 2026 after averaging approximately 70 percent annual growth during the previous three years.

At the same time, data-center and manufacturing demand has increased the importance of dispatchable generation and transmission infrastructure. The Department of Energy's 2026 draft National Transmission Needs Study describes a pressing need for additional transmission due partly to rapidly growing loads from hyperscale data centers, manufacturing and other industrial development.

The American challenge is therefore not best understood as a simple shortage of generating resources. The country possesses enormous energy resources, deep capital markets, advanced technology companies and a large existing grid, but it is attempting to convert an electrical system developed during an era of modest demand growth into one capable of supporting rapidly expanding computational and industrial loads. The strategic question is whether that transition can occur quickly enough to prevent infrastructure from becoming the limiting factor on technological expansion.

Energy Diversity Becomes a Strategic Advantage

The scale and speed of AI demand also make arguments centered around a single generating technology increasingly inadequate. The electricity system serving the intelligence economy will require resources with different development timelines, cost structures and operating characteristics because no individual technology can satisfy every part of a rapidly expanding load profile.

The IEA expects renewable generation to supply almost half of the additional electricity consumed by data centers globally through 2030, reflecting the rapid construction timelines and competitive economics of solar and wind in many markets. Natural gas supplies much of the remaining near-term dispatchable growth, particularly in the United States, while nuclear generation becomes more important later in the decade and during the 2030s as restarts, life extensions and new reactors begin contributing additional firm capacity.

This mix illustrates why an abundant energy system is better understood as an engineering portfolio than as an ideological commitment to one resource. Solar can be deployed quickly but does not produce continuously; storage can shift electricity across hours but remains a reservoir rather than a primary energy source; natural gas can provide dispatchable generation but depends upon fuel infrastructure and turbine availability; nuclear can supply large quantities of continuous power but generally requires longer development timelines and significant upfront capital.

The strategic advantage belongs to electrical systems capable of combining these resources effectively enough that rapid growth in one sector does not force reliability or cost problems onto the rest of the economy. That challenge becomes more difficult because AI is not the only source of additional demand. Advanced manufacturing, semiconductor fabrication, transportation electrification, building electrification, industrial investment and population growth can all increase electrical loads simultaneously.

The IEA now describes this broader transformation as an emerging “Age of Electricity,” projecting global power demand to grow more than 3.5 percent annually through 2030 and electricity consumption to expand substantially faster than overall energy demand. Artificial intelligence sits inside that transition, intensifying it while also increasing the economic value of getting the underlying system right.

The Industrial Base Behind Power Is Becoming Part of the Competition

Building additional generation begins with the supply chains capable of manufacturing the equipment from which power systems are constructed, making those industries essential to every electrical project. Transformers illustrate why manufacturing capacity for even fundamental electrical components cannot be expanded instantly, because large power transformers require specialized facilities, materials and skilled labor. Similar constraints exist across switchgear, cables, turbines and other components.

The IEA's 2026 assessment identifies tightening supplies of gas turbines, transformers, advanced chips and high-bandwidth memory among the physical bottlenecks already affecting AI infrastructure expansion. This creates an industrial recursion within the intelligence economy: AI companies need data centers; data centers need electricity; electricity systems need equipment; and equipment manufacturers need factories, materials, skilled workers and energy of their own.

A country that wants to increase electricity generation but cannot obtain transformers, turbines or other critical equipment quickly enough may find that nominal energy resources do not translate into usable power on the required schedule. Similarly, possessing sufficient generating capacity does not automatically produce computational capacity if advanced semiconductors remain unavailable.

The competitive unit therefore encompasses the complete industrial stack, from energy resources, generation and grid infrastructure through data centers, semiconductors and compute to the models and applications that create productive capability. The strategic value lies in how effectively those layers reinforce one another rather than in the isolated strength of one component.

Capital Is Competing for the Same Physical Bottlenecks

The scale of investment adds another dimension to the problem. Data-center construction has become so capital intensive that the IEA expects financing requirements to extend beyond the balance sheets of even the world's largest technology companies, increasing reliance upon capital markets and making infrastructure expansion more sensitive to borrowing costs, expected returns and broader financial conditions.

Energy infrastructure is simultaneously experiencing enormous capital requirements of its own. The IEA expects total global energy investment to reach roughly $3.4 trillion in 2026, with approximately $2.2 trillion directed toward renewables, nuclear power, grids, storage, efficiency, electrification and other lower-emissions technologies.

The significance of these figures is not simply that infrastructure is expensive, but that AI expansion increasingly competes within global capital markets for many of the same resources required by the wider electricity system: engineering capacity, construction labor, manufacturing equipment, financing and industrial materials. A proposed data center may require a transmission expansion that competes for engineering resources with other grid projects, while a new natural-gas plant may enter the same turbine queue as utilities responding to industrial demand elsewhere. Semiconductor fabrication facilities require their own electrical infrastructure even while supplying chips to the data centers creating additional demand.

The intelligence economy therefore amplifies pressure across multiple layers at once, which is why some of the most consequential economic effects of AI may occur several steps away from the companies developing the models. Electrical equipment manufacturers, engineering firms, energy producers, utilities, construction companies, semiconductor suppliers and infrastructure financiers all participate in the system through which machine intelligence becomes physically possible.

Energy Security and Compute Security Are Beginning to Converge

As machine intelligence spreads into more economically and strategically important activities, the reliability of the infrastructure supplying it acquires a different significance. Compute already supports communications, finance, logistics, scientific research, military systems, cybersecurity and government operations, while AI is increasingly being integrated into each of those areas.

An economy that depends heavily upon computational infrastructure consequently depends upon the energy system operating underneath it. A prolonged electricity shortage, fuel disruption or grid failure capable of removing large amounts of computational capacity would not remain confined to the technology sector; its effects could propagate into any activity relying upon those systems. Energy security and compute security therefore begin to overlap, particularly as larger portions of economic and strategic activity depend upon continuous access to machine intelligence.

This does not imply that every data center must become electrically self-sufficient or that countries should attempt technological autarky. Interconnected grids, cloud computing, international trade and diversified supply chains can strengthen resilience precisely because they allow resources to be shared across locations.

The concern arises when concentration becomes sufficiently extreme that failure within one physical layer can disable disproportionately large amounts of capability above it. A resilient intelligence economy therefore rests not simply upon redundant servers and duplicated software, but upon diverse energy supplies, robust electrical networks, sufficient equipment inventories and industrial capacity capable of repairing or expanding critical systems when conditions change.

The AI Race Is Also a Construction Race

One of the more underappreciated distinctions between software innovation and infrastructure expansion is that infrastructure cannot be copied. A successful algorithm can be distributed globally in seconds, whereas a new transmission corridor requires physical construction. A trained model can be reproduced across multiple computing clusters relatively quickly once hardware is available, while a nuclear plant, gas pipeline, substation or turbine factory must be physically constructed at a particular location through a sequence of engineering, financing, permitting and construction decisions that may take years.

This creates a temporal mismatch between the speed of intelligence innovation and the speed of the systems required to support it, which increases the economic value of competent execution. Regions capable of moving electrical projects from planning into operation within predictable timelines can translate investment intentions into actual compute more quickly than regions where infrastructure remains caught between application, interconnection study, permitting and construction. At national scale, installed generation, completed transmission, operating semiconductor factories and functioning data centers create capability through the completion of work that announcements and project pipelines can only anticipate.

Artificial intelligence may be among the most sophisticated technologies ever developed, but scaling it requires extraordinarily conventional activities: pouring concrete, erecting steel, manufacturing equipment, installing transformers, laying transmission lines and operating power stations. The competition to expand intelligence is therefore partly a competition over who can still build.

Energy Abundance Expands the Intelligence Frontier

The deeper economic importance of energy abundance appears when considering what sufficient infrastructure makes possible after it has been built. A region with electrical headroom does not have to know precisely which future computational applications will prove most valuable before constructing the capacity capable of supporting them. Additional generation and grid capacity create options for entrepreneurs and investors to discover those uses later.

That optionality resembles the role other forms of infrastructure have played throughout economic history. An interstate highway network supports industries whose precise future shipping patterns could not have been predicted when the roads were constructed, just as telecommunications capacity enables businesses that did not exist when fiber was installed. Industrial land and water infrastructure allow communities to accommodate investments whose technologies may change from one decade to another.

Electrical abundance performs a similar function because it raises the ceiling on what can be built. The importance of that ceiling increases when energy can be translated into cognition, since a megawatt can now support not only motors, furnaces, pumps and lighting but computational systems capable of performing analysis, designing products, operating robots, coordinating transportation, interpreting medical information, generating software and accelerating scientific research. Greater electricity capacity therefore expands the range of machine intelligence that can economically exist inside an economy.

That relationship does not make every additional kilowatt-hour equally valuable, nor does it justify building infrastructure without regard to cost. Capital still has alternative uses, electricity systems still require disciplined planning, and large data centers themselves must eventually generate sufficient returns to justify the extraordinary capital being committed to them. The argument is instead that energy abundance raises the frontier of potential intelligence production, while persistent scarcity lowers it. Over time, that difference can become economically significant.