Artificial intelligence has a power problem.

The United States can manufacture more chips.

AI companies can build more servers.

Investors can provide more money.

But none of that matters if the electric grid cannot deliver enough power to run the machines.

That problem is becoming especially visible on PJM Interconnection, the largest U.S. power-grid operator, serving roughly 67 million people across 13 states and the District of Columbia.

PJM is now preparing extraordinary measures as data centers and other large electricity users drive demand far faster than the grid's supply pipeline can keep up. The grid operator has warned of widening electricity shortages and is working with federal regulators on mechanisms to secure new generation while managing the load created by huge new facilities.

This is the point at which the AI story stops looking like software.

It starts looking like heavy industry.

A modern AI data center is an enormous electricity consumer.

Advanced processors run continuously.

Cooling systems operate around the clock.

Networking equipment consumes power.

Storage systems consume power.

And unlike many traditional industrial facilities, an AI data center can require a huge amount of electricity concentrated in one location.

That concentration creates a challenge for grid operators.

A new office building adds demand gradually.

A giant AI campus can add hundreds of megawatts or even gigawatts in a single development.

PJM estimates that prospective large electricity users could add roughly 70 gigawatts of demand by 2038, with data centers representing a major share of that increase.

To put that number into perspective, 1 gigawatt represents 1,000 megawatts.

A 70-gigawatt load increase is therefore comparable to adding dozens of major power plants' worth of demand.

And that demand is arriving at precisely the moment when portions of America's generation fleet are retiring.

PJM's latest capacity auction provided one of the clearest warnings yet.

The grid operator came up roughly 6.8 gigawatts short of the reliability requirement for the 2028-29 delivery year, even after the auction cleared at a price cap of approximately $325 per megawatt-day. Reuters reported that the shortfall was the second consecutive auction in which PJM failed to procure enough capacity to meet its reliability requirement.

That is a remarkable result.

The whole purpose of a capacity market is to ensure that enough electricity-generating resources will be available when the grid needs them.

If the auction reaches its maximum allowed price and still does not obtain enough capacity, the problem is no longer merely pricing.

It becomes a physical infrastructure problem.

The system needs more generation.

It also needs transmission.

And it needs those assets quickly.

The consequences are not limited to future blackouts.

They are showing up in electricity costs.

PJM's wholesale power costs averaged $136.53 per megawatt-hour in the first quarter of 2026, compared with $77.78 during the same period a year earlier, a roughly 76% increase. Bloomberg attributed the jump partly to intense data-center demand.

That creates one of the most contentious questions in the AI infrastructure boom:

Who pays for the additional power capacity?

Data-center companies argue that they are bringing investment, jobs and tax revenue.

Utilities need to build infrastructure to serve them.

Grid operators need more generation.

Consumers worry that they will be asked to pay for infrastructure they do not personally use.

That debate is already spreading across the United States.

In July, PJM announced plans to establish a reliability backstop procurement process to fill the 6.8-gigawatt gap.

The process is designed to attract new power resources that can begin supplying electricity to the grid.

PJM also proposed a registry to track large new electricity users, including data centers, and to better understand where demand is developing and how much power individual projects expect to consume.

That sounds bureaucratic.

It is actually a major change.

Grid planners need accurate information about data centers because large-load additions are no longer small enough to treat as normal demand growth.

A hyperscale campus can change the electricity requirements of an entire region.

PJM has also proposed rules under which certain large new customers could be required to secure their own generation or face curtailment during grid emergencies.

Reuters reported in August that the operator was considering emergency procedures under which data centers could be disconnected before residential customers if those facilities had not arranged sufficient independent power.

That is an extraordinary possibility.

For decades, large industrial customers generally expected utilities to provide electricity from the common grid.

AI data centers are changing that expectation.

The argument is increasingly becoming:

If you want to build a massive new electricity demand center, you may need to bring some of the supply with you.

That is a significant shift in how infrastructure may be financed.

PJM is not the only place where this pressure is visible.

Across the United States, developers are planning data-center campuses with electricity requirements measured in gigawatts.

Business Insider recently highlighted projects including Project Matador in Texas, which is pursuing as much as 17 gigawatts of private power capacity, and an Ohio campus targeting as much as 10 gigawatts, much of it contracted for OpenAI use.

Meta, Google, Oracle, xAI and other major technology companies are all participating in this buildout.

The projects are often being located in rural areas because large plots of land are cheaper and electricity may be easier to secure.

But rural locations do not eliminate the infrastructure problem.

They simply move it somewhere else.

New transmission lines may be needed.

New substations may be required.

Gas pipelines may need expansion.

Nuclear plants are being reconsidered.

Battery storage is becoming more important.

And companies are exploring behind-the-meter generation to reduce their dependence on the public grid.

The electricity demands of AI are already changing the conversation around nuclear power.

Microsoft has been pursuing nuclear arrangements.

Google has invested in advanced nuclear technology.

Amazon has also explored nuclear-powered data-center strategies.

The attraction is straightforward.

Nuclear reactors can provide large quantities of electricity continuously without relying on sunlight or wind conditions.

AI data centers need exactly that kind of reliability.

But nuclear construction is slow.

New reactors can take years to plan, finance, permit and build.

That creates a timing mismatch.

AI companies want computing capacity now.

Power projects can take years.

This explains why natural gas is playing such an important role in the near-term AI power story.

Gas-fired plants can generally be built faster than nuclear facilities.

PJM's restructured interconnection queue has attracted significant interest from gas projects as data-center demand grows. Reporting from Inside Climate News noted that natural gas is dominating parts of PJM's first reformed queue because developers are looking for relatively fast ways to add dispatchable generation.

That creates another tension.

AI companies need reliable power.

Gas can provide it.

But gas infrastructure also creates questions around fuel availability, emissions and long-term investment risk.

Grid planners therefore face a difficult choice between speed and long-term system design.

PJM's CEO David Mills warned in May that the current structure was not adequate to accommodate the surge in demand while simultaneously protecting households from rising electricity costs.

Bloomberg reported that Mills said PJM could not ensure sufficient future electricity supply while shielding residential consumers from rising bills under the existing structure.

That is an unusually blunt assessment from the head of the country's largest grid operator.

The problem is not only that data centers consume power.

It is that they can consume so much power so quickly that traditional planning models struggle to keep up.

Grid infrastructure was built around relatively predictable changes in electricity demand.

AI is creating a demand curve that can move much faster.

This is one of the most important developments in the entire AI industry.

Investors often compare companies based on processors, models and revenue.

But electricity is becoming just as important.

A model can be extremely profitable in theory.

If the company cannot secure enough power to run it, however, the economics stop mattering.

That is why AI companies increasingly sign long-term electricity agreements and pursue their own generation.

The competition for AI infrastructure is becoming a competition for electricity contracts.

The speed of data-center construction has also created political and social resistance.

Residents in several states have complained about noise, water consumption, land use and the potential effect of large power infrastructure on local communities.

The Financial Times reported this week that more than 1,000 U.S. communities have been involved in efforts to impose moratoriums or restrictions on data-center development, affecting more than $200 billion in proposed projects.

That is a sign that the debate is no longer confined to utility regulators.

It is becoming a community issue.

People want the economic benefits.

They also want to know whether they are being asked to subsidize the infrastructure required to deliver those benefits.

The cost-allocation debate is particularly sensitive.

If a utility builds a new transmission line to serve a data center, should the cost be paid primarily by the data center?

Or should it be spread across millions of households and businesses connected to the grid?

There is no universally accepted answer.

Utilities typically operate under regulatory frameworks that distribute infrastructure costs across customer classes.

But the scale of AI projects is challenging traditional assumptions.

When one new customer adds demand equivalent to a small city's electricity use, the economics look very different.

That is why regulators are increasingly considering special rules for very large electricity users.

The technology sector has historically moved according to Moore's Law, faster chips and better software.

Electricity works differently.

Power plants take time.

Transmission lines take time.

Transformers take time.

Permitting takes time.

Construction takes time.

That creates a physical bottleneck that software cannot simply optimize away.

AI can make a model more efficient.

It cannot instantly build a substation.

AI companies are working on more efficient chips, models and data-center designs.

Better software can reduce the number of computations required for certain tasks.

Liquid cooling can improve energy efficiency.

Chip improvements can deliver more AI performance per watt.

Those gains matter.

But if demand for AI computing grows faster than efficiency improves, total electricity consumption can still rise.

That is the paradox.

More efficient AI can make AI cheaper.

Cheaper AI can increase usage.

More usage creates more demand.

The efficiency gains can therefore become fuel for further growth rather than a substitute for additional electricity.

The United States has entered a stage where AI policy and energy policy can no longer be separated.

The country needs more chips.

It needs more data centers.

Those require more electricity.

More electricity requires more generation and transmission.

And that infrastructure needs capital.

The system therefore becomes circular.

AI creates demand for power.

Power constraints limit AI.

Those constraints encourage new generation.

New generation requires investment.

And investment creates another layer of the AI economy.

The biggest beneficiaries may eventually include utilities, power developers, nuclear companies, natural-gas generators, battery manufacturers, grid-equipment suppliers and transmission businesses.

For years, Nvidia's GPUs were considered the ultimate bottleneck in artificial-intelligence development.

Now memory is becoming a bottleneck.

Data-center construction is becoming a bottleneck.

And increasingly, electricity is becoming a bottleneck.

PJM's 6.8-gigawatt capacity shortfall and projected 70-gigawatt growth in large-load demand illustrate the scale of the challenge.

That does not mean America's grid cannot support the AI boom.

It means the grid must change to support it.

The traditional model of gradually adding electricity capacity may no longer be sufficient.

AI is forcing utilities and regulators to plan around giant customers that can appear almost overnight and consume vast quantities of electricity continuously.

That may be the most important lesson from the current power crunch.

Artificial intelligence is often presented as an invisible technology living inside software.

But the physical reality is enormous.

It requires buildings.

Chips.

Cooling.

Water.

Transmission.

Transformers.

Power plants.

And thousands of workers.

The companies that win the next stage of AI may therefore be determined partly by something that has nothing to do with model intelligence.

They need electricity.

Lots of it.

And on America's largest power grid, the race to secure that electricity has already begun.

The AI boom may have started with a chatbot.

It is increasingly becoming a contest over who can secure the power to keep the machines running.