The AI Boom May Make America Richer and Electricity More Expensive

Texas built much of its modern identity around abundance: land, oil, gas, electricity and room to build almost anything.

The artificial intelligence boom has managed to make even Texas wonder whether there is enough.

On September 21, Governor Greg Abbott ordered state regulators to stop issuing permits sought by data centres until their demands on electricity and water could be properly examined. Developers would have to show what they intended to consume, how they intended to obtain the power and whether the infrastructure built for them might leave somebody else paying for it.

“Data centers must pay their own way,” Abbott said.

It was a striking intervention from the governor of the American state most closely associated with cheap energy and rapid development. Texas was not turning against artificial intelligence. It was discovering that the computers could arrive considerably faster than the electricity needed to run them.

That collision is becoming one of the most important economic stories of the AI boom.

Hundreds of billions of dollars are being poured into data centres, computers, software and the power systems required to keep them running. That expenditure is already helping to increase American investment and gross domestic product. Yet a data centre does not really exist in the cloud. It occupies land, sits on vast quantities of concrete, consumes water and, above all, requires an immense and continuous supply of electricity.

If generating capacity, transmission lines and substations cannot expand as quickly as the new demand, the consequences eventually appear in prices. The argument then ceases to be merely about artificial intelligence and becomes a much older economic question: who should pay for scarcity?

The great AI boom may make America richer. It may also make electricity more expensive for everyone else.

The boom has already reached GDP

There has been endless speculation about what artificial intelligence might eventually do to productivity. It may transform medicine, engineering and finance, reduce the number of people required for certain jobs or create industries that barely exist today.

Those questions remain partly in the future. The spending does not.

Federal Reserve Vice Chair Philip Jefferson showed in July just how large it has become. Capital expenditure likely associated with AI, including software, data centres, high technology equipment and power infrastructure, contributed an estimated 1.36 percentage points to US GDP growth in the first quarter of 2026. The same measure contributed about 1.4 percentage points in the first quarter of 2025.

Those are remarkable figures for an industry whose eventual economic return remains uncertain.

The explanation is straightforward. Nobody has to prove that artificial intelligence will change civilisation before building it begins to appear in GDP. Pouring concrete for a data centre creates economic activity immediately. So does buying thousands of servers, installing cooling systems, laying fibre, ordering transformers and expanding the substations and transmission lines required to supply them.

The AI economy therefore begins registering in the national accounts long before anybody can know whether the software inside those buildings will make lawyers twice as productive, discover new medicines or simply allow businesses to produce emails more quickly.

There is another peculiarity to this expansion. Traditional industrial booms often brought armies of workers with them. A car plant, steelworks or shipyard announced its arrival not merely through investment but through thousands of jobs.

Data centres have a different economic shape.

Brookings researchers examining about 1,500 American facilities found that the arrival of a first large data centre typically added only around 100 to 200 local jobs, depending on the kind of facility. Construction can employ large numbers of people for a period, but once a hyperscale centre is operating, its permanent workforce is remarkably small compared with the capital invested in it.

The World Economic Forum’s latest survey of chief economists captures this unusual combination. Seventy-eight per cent expect data-centre investment to make a significant contribution to global economic growth, while 61 per cent do not expect it to make a significant contribution to employment. The same survey found that 78 per cent expect the expansion to increase electricity prices.

The implications are hard to miss. Enormous amounts of capital can produce economic growth without creating anything resembling the employment generated by older industrial expansions, while placing substantial new demands on the electricity system underneath them.

America had almost forgotten how to build for rising electricity demand

For years, the American electricity industry enjoyed an unusual luxury: demand barely increased.

Between 2005 and 2019, electricity consumption grew by only about 0.1 per cent a year. Improvements in efficiency offset much of the effect of economic growth and the proliferation of electrical appliances, allowing utilities to plan for a world in which national demand was essentially flat.

That world is disappearing.

Between 2020 and 2025, electricity demand grew by about 1.7 per cent a year, with the US Energy Information Administration identifying data centres as an important force behind the change.

The numbers now being contemplated would have appeared improbable only a few years ago. Berkeley Lab estimates that data centres could consume 11.8 per cent of all US electricity by 2030, with a possible range between 9.5 and 15.3 per cent depending on the pace of development.

The significance lies not simply in how much electricity is being consumed, but in how quickly that demand can appear.

A technology company can buy land, raise finance and erect a data centre relatively quickly. Electricity infrastructure cannot necessarily follow at the same speed. A new power station may require years of planning and construction. Turbines and transformers must be ordered. Transmission corridors have to be approved and substations expanded, often through regulatory processes that move according to a very different timetable from the technology industry.

PJM, which runs the electricity system serving more than 67 million Americans across 13 states and the District of Columbia, now identifies data centres as the primary driver of its demand growth. It has also warned that these centres can be developed two or three times faster than many of the generating technologies required to serve them.

This is where the digital economy meets the physical one. Computing capacity can be ordered in months. Electricity infrastructure is still built in years.

The shortage has begun to acquire a price

Electricity markets are unusually good at disguising simple problems behind complicated terminology. Capacity auctions, reserve margins and transmission charges can make what is happening sound more obscure than it is.

PJM must ensure that enough generating capacity will be available several years from now to keep the system functioning when demand reaches its highest point. Power producers are paid not merely for electricity they generate today but for maintaining capacity that can be called upon in the future.

In July, the price in PJM’s capacity auction reached its regulatory ceiling of $325 per megawatt-day. Even at that price, the system procured 6,831 megawatts less capacity than its reliability requirement for 2028-29.

Nobody is suggesting that millions of Americans are about to lose their electricity. What the result does reveal is a power system in which demand is rising faster than reliable new supply.

At that point the AI investment boom begins producing two economic effects at once. The building of data centres, servers and power infrastructure adds to investment and GDP, but the scarcity created by rapidly increasing electricity demand can push power costs upwards.

The distinction matters because GDP is not a household ledger.

A technology company may spend several billion dollars constructing a new computing complex. The investment appears in national output, construction companies receive orders and manufacturers sell equipment. Yet the electricity company serving that project may then have to build another transmission line, reinforce a substation or secure more generating capacity.

Those costs exist regardless of how impressive the investment looks in the national accounts. The important question is where they eventually land.

Virginia discovered that a few dollars matter

Virginia has already confronted the problem.

Northern Virginia contains one of the world’s largest concentrations of data centres, and the pace of their expansion has forced regulators to consider what happens when electricity companies invest heavily for customers whose future demand remains uncertain.

A developer may propose a vast computing complex and ask for hundreds of megawatts of capacity. The utility begins reinforcing the network, securing transmission and preparing the connection in expectation of the load. If the project is later delayed, reduced or abandoned, the infrastructure constructed in anticipation of it does not disappear, nor do the financing costs attached to it.

Under an ordinary electricity tariff, some of that expense can filter through the system and eventually be recovered from other customers who played no part in the original investment decision.

Virginia decided to alter the arrangement.

From January 2027, hyperscale data centres and other very large electricity users will have to pay at least 85 per cent of the transmission and distribution capacity reserved for them, even if their eventual electricity consumption falls below that level. Regulators have also created a separate rate class for large users and imposed long-term obligations intended to prevent infrastructure costs being shifted to everybody else.

The State Corporation Commission produced a figure that makes the argument unusually easy to understand. Under the old method of allocating transmission costs, a typical household using 1,000 kilowatt-hours a month faced a projected monthly increase of $2.90. Once a greater share of the cost was allocated to the rapidly growing large users responsible for much of the new infrastructure, the projected household increase fell to 94 cents.

A saving of $1.96 a month sounds modest until it is multiplied across millions of households and repeated year after year.

Nothing about the physical infrastructure had changed. The same transmission assets still had to be financed. What changed was the judgment about which customers should bear the cost.

Texas has seen the queue

Texas faces the same issue on a scale that has begun to look almost surreal.

ERCOT was tracking roughly 226 gigawatts of proposed large electricity loads late last year, with data centres accounting for almost 73 per cent. Nobody expects all of those projects to be built, and that uncertainty is itself part of the difficulty.

Developers can seek connections for sites that may never proceed. Several proposed locations may represent different versions of the same planned investment. Projects can be delayed or cancelled as quickly as commercial assumptions change. The grid operator is nevertheless required to decide which of these demands are credible enough to influence billions of dollars of infrastructure planning.

The old method of considering large projects largely one at a time became increasingly difficult to sustain. Texas moved towards examining them together, judging how much demand the grid could realistically accommodate and identifying which upgrades would be required.

By September, Abbott had gone further and stopped data-centre permits while their impact was audited.

Few developments illustrate the physical scale of the AI boom more clearly. The constraint is no longer simply whether technology companies can obtain enough advanced chips. Increasingly, it is whether they can find somewhere capable of supplying all those chips with electricity.

Data centres do not have to make electricity dearer

There is an important counterargument, and leaving it out would distort the economics.

A large new electricity consumer can actually benefit the customers already connected to the system.

Electricity networks have enormous fixed costs. Power stations, substations and transmission systems must be financed whether they operate at half their potential or much closer to full capacity. If a data centre locates in an area where spare generating and network capacity already exists, buys enormous quantities of electricity and pays an appropriate tariff, those fixed costs can be spread over a larger volume of sales.

In those circumstances, average costs for other customers may fall.

Data centres can also finance their own connections, contract directly for new generation or reduce their electricity consumption at times when the grid is under exceptional pressure. Some operators are considering building power generation beside their facilities rather than relying entirely on the public network.

The problem, therefore, is not that computers consume electricity. Modern economies have always turned energy into productive activity. Aluminium smelters, chemical plants, factories and railways can all consume extraordinary amounts of power.

The difficulty arises when demand reaches the electricity system faster than supply can respond and when part of the cost of serving that demand is borne by somebody other than the customer creating it.

Recent economic research demonstrates why a simple answer is difficult. One study examining data-centre entry between 2010 and 2024 estimated that residential electricity prices rose by about 2.1 per cent following their arrival. Another study, using a different methodology, concluded that data centres had historically produced modest reductions in average retail electricity prices by spreading the fixed costs of the network across more electricity sales.

Those findings are not necessarily contradictory. They suggest that the result depends heavily on location, spare capacity, regulatory structure and the pace at which new infrastructure has to be constructed.

A data centre arriving beside an underused grid is one economic event. A cluster requiring new power stations, transmission lines and substations is another.

The argument is therefore not really about whether data centres are good or bad for electricity customers. It is about the terms on which they connect.

GDP is not a household ledger

There is a temptation during every investment boom to confuse the scale of the spending with the scale of the eventual benefit.

The railway booms of the nineteenth century created fortunes and bankruptcies while leaving behind infrastructure that transformed economies. The telecommunications boom produced extraordinary overinvestment in fibre, some of which initially appeared wasteful before becoming immensely useful as internet traffic grew.

The AI buildout may eventually acquire its own version of that history.

If artificial intelligence produces the productivity gains its advocates anticipate, today’s extraordinary spending on computing and electricity may look entirely rational in retrospect. A more productive economy would be able to support a much larger energy system, while investments that appear extravagant today could become the foundations of industries not yet visible.

But there is a distinction between measuring economic activity and deciding who benefits from it.

A billion dollars spent on a data centre raises investment. A billion dollars spent reinforcing the electricity network can raise investment as well. Both contribute to economic activity, but neither figure tells us how the gains and costs are distributed between technology companies, investors, electricity utilities, workers and households.

That is why the argument developing in Texas, Virginia and other data-centre states matters beyond the electricity industry.

Artificial intelligence is beginning to encounter the oldest constraint in economics: scarcity.

There may be almost unlimited appetite for computing power, but there is no unlimited stock of generating capacity, transmission lines and transformers waiting beside the grid to satisfy it.

The first phase of the AI boom was dominated by the struggle for chips. The next may be shaped just as much by the struggle for megawatts.

America may build enough generation and transmission to accommodate the data centres without imposing greater costs on everyone already connected to the network. Technology companies may finance enough of that expansion themselves that households eventually benefit from a larger and more efficient electricity system.

But if electricity demand continues to arrive faster than supply, power becomes more valuable. And when the companies creating that demand are not charged for the infrastructure built to serve them, the cost does not disappear into the grid. It is redistributed through it.

That is the economic choice now taking shape beneath the AI boom: not whether America should build more computing power, but how much of the physical infrastructure required to support it should be paid for by the companies demanding it and how much should be carried by everyone else.