There is a remarkably simple test for whether the AI boom is actually as economically transformative as its boosters say it is: Make the companies building it pay their own damn electricity bill.
Not just the meter at the server warehouse. The whole bill.
If a hyperscale data center requires a new power plant, transmission line, substation, distribution upgrade, water system, or grid-reliability backstop, put that cost on the project that caused it. If the economics still work, excellent. Build it. If they don't, then the public has just learned something extremely important about the business model.
Pennsylvania moved sharply in that direction on August 18. Gov. Josh Shapiro's new executive order requires data-center developers seeking state permits to meet Responsible Infrastructure Development standards. Among them: Developers must cover the cost of new generation, transmission, distribution, and related infrastructure needed for their projects without shifting those costs to households and businesses.
If the AI boom is real, it can survive a full-cost test.
The order also requires local approval, ends the use of nondisclosure agreements for data-center projects, demands energy and water reporting, and ties state tax benefits to compliance. Pennsylvania's special counsel for energy affordability is also supposed to work with utility regulators so data centers, rather than ordinary customers, absorb the cost of reliability measures created by their demand.
This shouldn't be treated as an anti-AI policy. It's a pro-accounting policy.
The AI industry has been allowed to narrate electricity as though it were weather: Demand is "surging," capacity is "tight," the grid is "strained," and somehow billions of dollars of wires, turbines, transformers, and land just need to appear around the technology.
But demand isn't weather. Somebody made a decision.
And there is another reason to force the accounting now: A lot of the supposed demand isn't load yet. It's proposals. Pennsylvania says more than 100 data-center projects have approached the state, while the governor says only a handful currently look viable enough to have the necessary permits. Across the country, regulators have worried about speculative projects showing up in multiple interconnection queues and inflating forecasts before anybody has proved the customer, financing, or power plan is real.
That matters because utilities build decades-long assets against forecasts. If a speculative 500-megawatt project gets counted as inevitable, everybody can end up paying for infrastructure long after the spreadsheet that justified it has disappeared.
So add another rule: Prove you're real before the public builds around you. Put down financial security. Identify the end user. Show the power source. Accept a minimum bill that covers infrastructure reserved for you. And if you walk away, you don't get to leave a transformer-shaped hole in everybody else's rates.
The International Energy Agency (IEA) reported in April that electricity consumption by data centers jumped 17% in 2025, while electricity use at AI-focused facilities rose even faster. Its 2026 outlook still sees total data-center consumption roughly doubling by 2030 and AI-focused consumption tripling.
In the United States, data centers accounted for roughly half of total electricity-demand growth in 2025, according to the IEA. The Department of Energy had already estimated that data centers consumed about 4.4% of US electricity in 2023 and could reach 6.7-12% by 2028.
There is nothing inherently illegitimate about using that much power. Aluminum smelters use enormous amounts of electricity. Steel mills use enormous amounts of electricity. Electrifying transportation and heating will use enormous amounts of electricity too. A serious industrial society needs to know how to build power.
The illegitimate part begins when one customer's expansion becomes everyone else's compulsory investment.
That risk isn't theoretical. In Virginia, the world's largest data-center market, Reuters reported this month that Dominion's fuel costs have risen nearly 90% in five years as data-center growth leaves the utility more exposed to expensive wholesale electricity. Across PJM, which serves 67 million people, the grid operator is now proposing rules that would put some new data centers first in line for curtailment during shortages unless they bring adequate power with them.
Good.
If an AI company wants the upside of scale, it can own the downside of scale too.
That means at least five things.
Make large-load customers pay for the incremental grid infrastructure they cause. Require financial security so households aren't stuck with stranded assets if a project disappears. Make large-load forecasts public enough to detect speculative projects being counted multiple times. Give communities access to energy, water, and ownership information before approvals. And establish emergency curtailment rules before the emergency arrives.
This is what mature industrial policy looks like. It doesn't worship growth, and it doesn't panic about growth. It prices the actual machine.
The alternative is a familiar political trick: Call the upside private innovation and the downside public infrastructure.
No.
If the AI boom is real, it can survive a full-cost test.
If a company needs a new power plant, let it pay for the whole damn thing.
