The artificial-intelligence boom is usually narrated as a software story. Models get larger, chips get faster, and firms race to turn new capabilities into products. But the next stage of the AI economy is increasingly physical. It requires land, transformers, transmission lines, cooling systems, semiconductors, and above all electricity.
That changes the economic question. The constraint on AI is no longer only whether engineers can build better models. It is whether the physical economy can build enough infrastructure, quickly enough, at prices that make the resulting compute worthwhile.
AI investment is becoming infrastructure investment
PwC projects global AI-infrastructure capital expenditure of $31.6 trillion through 2050, with annual data-center capex rising from roughly $800 billion in 2026 to $1.8 trillion by 2050. The exact forecast should not be treated as destiny, but its scale captures the shift: AI is becoming one of the largest capital-allocation stories in the global economy.
This matters because infrastructure behaves differently from software. A software product can be copied at near-zero marginal cost. A data center cannot. It needs financing, construction labor, grid interconnection, hardware replacement, and a reliable stream of power. Those inputs introduce bottlenecks that can persist even while model performance improves.
Electricity is becoming an economic variable in AI
A September 2026 study in Communications Sustainability estimates that AI data centers alone could account for about 1% of global electricity demand by 2030. That number sounds small until one remembers that electricity systems are local and capacity additions take time. A concentrated cluster of new data centers can matter enormously to a particular grid even if its share of global demand remains modest.
Economically, this creates scarcity rents. Locations with abundant generation, available transmission, predictable permitting, and cooling resources become more valuable. Power prices and interconnection queues begin influencing where compute is built. The competitive advantage of an AI company can therefore depend partly on assets far outside the model itself.
The hidden incidence question
Who ultimately pays for the infrastructure? The answer need not be “the AI companies.” If utilities must make major generation or grid investments, regulators face questions about how costs are allocated among large industrial users and ordinary ratepayers. Local governments face similar questions over water, land, tax incentives, and infrastructure upgrades.
This is a classic incidence problem. The party writing the check is not necessarily the party bearing the economic cost. Depending on market structure and regulation, some costs can appear in electricity rates, taxes, rents, or forgone alternative uses of scarce infrastructure.
Capital intensity changes the AI-bubble debate
The debate over whether AI investment is excessive often focuses on valuations. A more useful question is whether the enormous stock of physical capital being created will earn adequate returns. Data centers are not merely financial claims whose prices can fall without changing the productive stock. They are real assets that continue to exist after expectations change.
That creates two possible futures. In the optimistic case, demand for compute expands fast enough that today’s infrastructure becomes the foundation for a large productivity boom. In the pessimistic case, capacity is built ahead of profitable demand, utilization disappoints, and returns on the marginal project collapse. The infrastructure may still be socially useful even if some investors lose money—railroads and fiber-optic networks offer historical analogies—but that distinction matters enormously to capital markets.
AI may raise productivity and resource demand at the same time
There is a tendency to treat productivity improvements as purely disinflationary. In practice, expectations of future productivity can stimulate investment today. Federal Reserve research has explored precisely this mechanism: optimism about future AI-driven productivity can increase current demand before the productivity gains fully arrive.
That means the AI boom can contain two forces at once. Better technology can eventually expand supply, while the race to build the infrastructure required for that technology can create near-term demand for capital, labor, chips, land, and electricity. The macroeconomic effect depends on timing.
The real moat may be boring
AI’s most consequential competitive advantages may end up looking surprisingly old-fashioned: cheap power, financing capacity, procurement relationships, grid access, and the ability to build at scale. Models matter. But once many firms can access capable models, infrastructure economics can determine who can deploy them cheaply enough to win.
The AI economy is not dematerializing the world. It is doing the opposite. It is converting intelligence into a new source of industrial demand. The companies, regions, and policymakers that understand that physical layer will understand much more of the next phase of the boom than those watching model benchmarks alone.