Energy Scarcity Becomes Intelligence Scarcity

Artificial intelligence is often discussed as though intelligence itself were becoming almost infinitely abundant. A model can answer another question without educating another human being for twenty years; software can be duplicated across thousands of users; computational systems can operate continuously; and improvements in algorithms, semiconductors and model architecture continue to reduce the resources required to perform many individual tasks. Seen only from the application layer, machine intelligence can appear to be following the familiar trajectory of digital technology toward lower marginal costs and increasingly ubiquitous availability.

Beneath that apparent abundance, however, sits an industrial system governed by constraints that are far less abstract. Models require computation, computation requires machines, machines require capital and physical infrastructure, and every useful operation eventually requires electricity. The expansion of machine intelligence therefore creates an unusual economic relationship in which something experienced by the end user as software depends upon an increasingly large industrial base of generation, transmission, semiconductors, cooling systems, networking equipment and data centers.

The implication is broader than the observation that AI consumes electricity. If machine cognition becomes an important productive input throughout the economy, then limitations on the systems that produce it can influence how cheaply, extensively and rapidly organizations are able to deploy that cognition. Energy scarcity can constrain compute; constrained or expensive compute can raise the cost of machine intelligence; and differences in the availability and price of that intelligence can eventually become differences in productivity, investment and economic capability. This relationship forms the next stage of the Power → Compute → Intelligence argument.

Intelligence Has a Cost Structure

Human intelligence has always had an economic cost, although we rarely describe it in those terms. A physician, engineer, financial analyst, attorney or scientist represents years of education, training and accumulated experience, supported by institutions that themselves required enormous investments of capital and human effort. Organizations purchase access to that intelligence through wages, salaries, professional fees and the broader cost of maintaining skilled workforces.

Machine intelligence has a different cost structure, but it does not escape economics. Training advanced models requires enormous computational resources, while operating them requires continuing expenditures on processors, memory, networking, facilities, maintenance, cooling and electricity. Even when the marginal cost of an individual inference becomes very small, producing millions or billions of such inferences across businesses, public institutions, scientific systems, industrial processes and consumer applications creates a substantial aggregate requirement for physical resources.

The International Energy Agency estimates that global data-center electricity consumption could reach roughly 945 terawatt-hours by 2030 in its base case, more than twice the 2024 level. Electricity consumed by accelerated servers, which are closely associated with AI workloads, is projected to grow much faster than conventional server demand, at around 30 percent annually through the end of the decade.

Those figures should not be interpreted as evidence that every interaction with an AI system is inherently energy intensive. The opposite is increasingly true at the level of individual tasks. The IEA's 2026 analysis found that energy consumption per AI task has been falling extraordinarily quickly as hardware and software improve, with simple text queries now requiring relatively modest amounts of electricity. At the same time, increasingly computationally demanding applications are becoming practical and popular, while the total volume of AI usage continues to expand. That distinction between unit efficiency and aggregate demand is essential. A civilization can become dramatically better at producing a unit of machine intelligence while simultaneously consuming far more machine intelligence overall.

Efficiency Does Not End the Abundance Question

Computing history provides a useful precedent. Processors became vastly more efficient, storage became cheaper, networks became faster and software became more capable, yet society did not respond by deciding that it had finally accumulated enough computation. Instead, lower costs made entirely new uses economical, which expanded demand into areas that previously could not justify the expense.

Artificial intelligence appears likely to follow a similar pattern: a tenfold reduction in the cost of interpreting medical images gives organizations an incentive to analyze more images, just as cheaper automated engineering analysis allows engineers to run more simulations. Lower inference costs can also make continuously operating autonomous agents economical across previously unaffordable business processes, while inexpensive machine vision can turn cameras throughout factories, transportation systems, warehouses and infrastructure into continuous analytical instruments rather than passive recording devices.

Efficiency therefore expands the frontier of what intelligence can economically be used for. It reduces the amount of energy required for a particular computational task, but it can simultaneously increase the number of tasks worth performing.

The IEA's broader outlook reflects precisely this tension. Even while assuming continuing improvements in hardware, software and infrastructure efficiency, its base case has global data-center electricity consumption roughly doubling by 2030. A more aggressive efficiency scenario reduces the eventual electricity footprint substantially, while a faster AI-deployment scenario drives it considerably higher, demonstrating how heavily future demand depends upon the interaction between technological efficiency and the rate at which new uses emerge.

For economic analysis, continuing gains in AI energy efficiency raise a further question: whether improvements in efficiency can outrun the expansion of useful applications indefinitely, particularly as machine intelligence moves from occasional interactions with software into continuous operation across the physical and institutional economy.

Scarcity Appears at the Margin

Energy scarcity should not be understood here as a country literally running out of electricity. Modern economies rarely experience scarcity in such a simple form. Scarcity appears through prices, waiting periods, congestion, interconnection queues, rationed capacity, delayed projects and investment decisions made elsewhere.

A company may be technically capable of purchasing electricity, for example, while discovering that a proposed data-center site cannot receive another 300 megawatts without a transmission upgrade that will take several years. Another business may have access to power but face electricity prices that undermine the economics of operating energy-intensive computational infrastructure locally. A metropolitan area may have enough generation across the broader region while lacking substations or transmission capacity at the location where investment is proposed.

This distinction is particularly important because data centers create large, geographically concentrated loads. The IEA estimates that global data-center consumption remains a relatively modest share of total electricity demand, reaching just under 3 percent in its 2030 base case, yet the local effects can be much more substantial because infrastructure clusters in particular regions. Nearly half of U.S. data-center capacity was already concentrated within five major regional clusters in the IEA's 2025 assessment.

The grid itself is increasingly where this scarcity becomes visible. The IEA reported in 2026 that more than 2,500 gigawatts of proposed generation, storage and large-load projects, including data centers, were sitting in grid-connection queues around the world. Meeting forecast electricity demand through 2030 would require annual global grid investment to rise by roughly 50 percent from current levels, according to the agency's analysis.

The economic constraint is therefore not simply whether a nation can generate another kilowatt-hour somewhere within its borders. It is whether the complete system can deliver sufficient power, at an acceptable price and level of reliability, to the places where productive activity is attempting to expand.

The Price of Electricity Can Become Part of the Price of Intelligence

This leads to a second dimension of scarcity that deserves greater attention: cost. Electrical power is already a major determinant of competitiveness for traditional energy-intensive industries. The IEA reported that final electricity prices for large energy-intensive industrial customers in the European Union remained more than twice U.S. levels in 2025 and nearly 50 percent above Chinese levels. Those differences do not explain every variation in industrial competitiveness, but the IEA explicitly identifies energy prices alongside labor, capital, raw materials, productivity, infrastructure and innovation as structural determinants of where production can compete effectively.

Machine intelligence will not have the same cost structure as aluminum smelting or chemical production, and electricity represents only one part of the enormous expense involved in AI infrastructure. Semiconductors, networking equipment, land, cooling, financing, engineering and software all contribute to the economics. Nevertheless, as computational infrastructure becomes larger and more continuously utilized, differences in electricity prices increasingly become part of the cost of producing compute.

The relationship becomes especially relevant when AI services operate at very large scale. A modest electricity-cost difference may mean little when applied to a single query, but differences compounded across thousands of accelerators operating continuously for years become economically material. Companies building such infrastructure therefore have incentives to locate capacity where they can combine suitable power prices with reliability, grid access, telecommunications infrastructure, favorable regulation and access to capital.

This is where energy economics begins to bleed directly into the economics of intelligence. Regions with structurally expensive electricity may still participate successfully in the AI economy through research, software, specialized applications and the purchase of remote computing services, just as countries with expensive energy remain capable of producing enormous economic value in other sectors. What changes is the relative attractiveness of locating the most electricity-intensive portions of the computational stack there. Over time, this can influence where the physical production of machine intelligence concentrates.

Compute Can Be Imported, but Dependence Has a Price

Cloud computing complicates the argument because machine intelligence does not necessarily have to be produced where it is consumed. A company in Ohio can call a model running in Virginia; a business in Germany can purchase computational services provided from another European country or another continent; and smaller economies can gain access to extraordinarily capable AI systems without constructing frontier-scale data centers themselves.

That is one of the great economic advantages of digital networks, and it prevents the energy-to-intelligence relationship from becoming a simplistic claim that every country must build its own complete AI infrastructure. Trade, specialization and distributed systems allow economies to share the benefits of computational infrastructure, while attempting to duplicate every level of the supply chain within every political jurisdiction would squander capital rather than create resilience.

Dependence becomes strategically significant, however, when an economy loses meaningful control over resources upon which increasingly important portions of its productive system depend. A business that cannot obtain sufficient domestic compute can rent it elsewhere, but the price, availability and contractual terms of that compute are then determined partly by infrastructure outside its immediate control. At national scale, the same dependency becomes relevant to scientific research, defense, cybersecurity, critical infrastructure, financial systems and advanced industrial activity.

The distinction resembles other strategically important commodities. A country does not need to extract all of its own petroleum, fabricate all of its own semiconductors or manufacture every machine tool it consumes, yet governments and firms still care deeply about whether supply chains are sufficiently diversified and resilient to prevent disruption from becoming economic paralysis. Compute increasingly belongs in that discussion.

When Intelligence Becomes a General-Purpose Input

The importance of this issue grows if AI develops into something resembling a general-purpose technology, meaning that its value does not remain confined to one sector but spreads throughout the economy. Machine intelligence is already being incorporated into software development, medical research, financial analysis, logistics, manufacturing, customer service, cybersecurity, scientific modeling and administrative processes. Future systems are likely to become more deeply integrated into robotics, transportation, industrial control, defense, engineering design and infrastructure management.

The economic significance of AI then becomes less analogous to one successful software industry and more analogous to previous technologies whose influence spread through nearly every productive sector. Electricity itself provides the obvious historical comparison. The competitive advantage of electrification did not come primarily from creating an electricity industry that employed enormous numbers of people; it came from allowing factories, offices, households, transportation systems and communications networks to reorganize around electrical power.

AI may produce a similar second-order effect. Its greatest economic value could emerge not from companies selling AI models, but from millions of organizations reorganizing production around the availability of inexpensive machine cognition. If that occurs, the cost of intelligence becomes part of the cost structure of the wider economy.

A manufacturing company with inexpensive access to AI-assisted engineering, predictive maintenance and automated quality control gains capabilities that a competitor may have to obtain at greater expense. A hospital able to deploy machine intelligence throughout scheduling, documentation, imaging and clinical support can reorganize workflows differently from one constrained by high computational costs. A logistics network using continuous optimization across vehicles, inventory and routing can make decisions at a frequency that traditional planning systems cannot economically reproduce. Individual differences may initially appear incremental, yet productivity differences compounded across thousands of organizations become macroeconomic differences.

The Geography of Intelligence Will Follow More Than Talent

The geography of the technology economy has historically been explained primarily through concentrations of human capital, venture financing, universities and entrepreneurial networks. Those agglomeration effects remain powerful, particularly where frontier research and company formation are concerned, but the industrialization of AI introduces physical inputs that do not necessarily concentrate in the same locations.

Compute can be built where electricity, land, fiber, cooling and regulatory conditions are favorable, while the people using that compute may reside hundreds or thousands of miles away. This separation creates the possibility that the geography of intelligence production will become distinct from the geography of intelligence application.

The divergence would not be unprecedented. Financial centers need not sit beside every physical asset being financed; corporate headquarters need not occupy the same cities as factories; and cloud infrastructure already operates according to a geography quite different from that of its users. AI expands the scale at which this separation can matter.

Regions with abundant energy resources, existing transmission capacity, favorable industrial land and a demonstrated ability to build infrastructure may therefore acquire a role in the intelligence economy that would have seemed unlikely during an era when technology development was imagined primarily as a software and talent phenomenon.

This relationship connects directly to Analysis 02, where the megawatt becomes an economic-development metric precisely because electrical capacity is increasingly capable of supporting something more valuable than electricity consumption alone. It can support computational infrastructure that produces a productive resource usable far beyond the physical boundaries of the region generating it.

Abundance Can Become a Competitive Strategy

If scarcity raises the cost and constrains the availability of machine intelligence, energy abundance can create room for new productive uses and investment. A power system with sufficient reserve capacity, robust transmission, competitive generating costs and the ability to expand can respond to new industries without immediately forcing every major investment into a contest over existing electrical capacity. The benefit is not confined to data centers; advanced manufacturing, semiconductor production, electrified industrial processes, transportation infrastructure and future technologies whose demand profiles are not yet fully known can draw upon the same underlying abundance.

The United States is already entering a period in which those demands overlap. EIA's September 2026 outlook projects record U.S. electricity consumption in both 2026 and 2027, with data-center development and increased manufacturing activity among the principal drivers. The IEA expects U.S. electricity demand to rise by more than 420 TWh through 2030, with data centers accounting for roughly half of that increase.

Those figures illustrate why a strategy of abundance must be broader than simply building enough generation to satisfy today's AI projects. Power systems are being asked simultaneously to support computational expansion, industrial investment, population growth, transportation electrification and other new loads. Infrastructure planned narrowly around one sector risks creating bottlenecks somewhere else.

An abundance strategy therefore requires generation, transmission, fuel supply, equipment manufacturing and regulatory systems capable of expanding together. Nuclear generation, natural gas, renewables, hydroelectric resources, geothermal energy, storage and other technologies will contribute differently according to geography, cost, reliability characteristics and development timelines; the economically important question is whether the resulting system can deliver enough dependable electricity for productive investment without making scarcity itself the organizing principle.

Intelligence Rationing Would Be an Economic Failure

There is a temptation whenever demand grows quickly to focus predominantly on controlling consumption. Efficiency, flexible loads and smarter scheduling can certainly reduce strain on electrical systems and should be pursued wherever their economics make sense, particularly because some computational workloads can be shifted across time or geography more easily than conventional industrial processes.

Yet demand management is not a substitute for supply expansion when the underlying demand represents productive economic activity. If businesses are attempting to purchase more computation because that computation generates more value than it costs, preventing the required infrastructure from being built does not eliminate the economic demand; it prevents some of that value from being created domestically. The workload may move to another region, the investment may be delayed, or companies may simply deploy less intelligence than their competitors.

This is the deeper problem with treating rising electricity demand as though it were inherently undesirable. A kilowatt-hour consumed by an economically productive system is not a policy failure simply because consumption increased. The relevant questions concern what value the energy produces, what alternatives exist, whether costs are being properly borne and whether infrastructure can expand without compromising reliability.

The same principle has applied throughout industrial history. Economies did not become wealthier by establishing a fixed ceiling on the amount of mechanical work factories were permitted to perform; they became wealthier by learning to produce mechanical work more efficiently while simultaneously expanding the quantity of useful work available to society. Machine cognition should be examined through a similar lens.

The Constraint Can Move Up the Stack

The Power → Compute → Intelligence framework is useful because constraints can propagate upward. When electrical infrastructure is abundant and inexpensive, more computational infrastructure can be economically deployed. As compute becomes abundant and its cost falls, developers can build applications that would have been uneconomic when inference was expensive. As machine intelligence becomes cheaper and more capable, organizations can apply it to a wider range of productive activities, which can raise productivity and create new forms of economic value.

Scarcity propagates through the same chain in the opposite direction. If electricity is difficult to obtain, the marginal data center becomes harder or more expensive to build. If computational capacity remains scarce relative to demand, developers face higher costs or constrained availability. When intelligence becomes more expensive, fewer applications clear the economic threshold required for deployment, particularly in businesses operating on narrow margins or in sectors where the value of individual AI interactions is modest.

The constraint therefore does not necessarily appear to the final user as an electricity problem. It may appear as an expensive API, a capacity limit, a delayed product, an unavailable service, a cloud-computing premium or simply a business process that never becomes economical enough to automate. That is why infrastructure constraints can be easy to underestimate. By the time scarcity reaches the application layer, the physical origin of the problem may be almost invisible.

Productivity Will Reveal the Difference

No single electricity price, data-center count or gigawatt figure can determine which country will prosper in an AI-intensive economy. Economic performance emerges from the interaction of institutions, entrepreneurship, human capital, markets, technological capability, infrastructure, capital formation and countless decisions made by firms and individuals. Energy belongs inside that system rather than above it.

The thesis of this series concerns the production relationships through which physical resources and technical capabilities become economically useful intelligence. If economically useful machine cognition is created by combining electricity, compute infrastructure, models, software, data and human direction, then the cost and availability of those inputs affect the quantity of intelligence that can economically be produced.

The differences may ultimately become visible in productivity statistics rather than energy statistics. Regions where organizations can cheaply deploy AI across factories, hospitals, transportation systems, professional services, laboratories and infrastructure may discover more opportunities to reorganize work around machine cognition. Regions where computational capacity remains more expensive or difficult to obtain may still adopt the same technologies, but later, less intensively or through greater dependence on infrastructure located elsewhere. Over time, those differences can accumulate.