AI's Real Rules Are About Electricity, Not Intelligence

The AI rule with the most teeth this year was filed in August by an organization most Americans have never heard of, at an agency even fewer could name. PJM Interconnection is the grid operator for the Mid-Atlantic, the body that decides whose electricity flows and whose goes dark when there isn't enough. It proposed to the Federal Energy Regulatory Commission, the referee of interstate power, that a new class of customer accept new terms: any new load of 50 megawatts or more, which is the draw of a small city and, in practice, means an AI data center. A customer that arrives without its own generation sits first in line to be switched off when the grid runs short, and is paid half the going rate for conserving when asked. The escape hatch has a name and a date. The filing's own term is "Bring Your Own New Capacity": new generation online by March 1, 2027, and the load is an ordinary customer. Arrive without it, and it sits at the front of the cut line. PJM put its reasoning in the open.

The unprecedented addition of large loads, most notably data centers, has given rise to resource adequacy shortfalls — and associated real-time operational issues that Interim Resource Adequacy Service is intended to address — PJM Interconnection

The proposal says nothing about intelligence, training data, or what a model might do to anything. It is about wires. It binds the moment the referee signs, and the industry is already behaving as though it will. ^[1] Meanwhile the argument in Washington runs the other way, toward the models. On September 1, OpenAI shipped GPT-6 Astra, a model able to find and exploit zero-day flaws — bugs nobody has patched — without being told where to look, trained on 100,000 GPUs in Texas and launched after a post-breach pause of a few weeks. ^[2] Thirteen days later came the labs' coordinated call for slower capability gains, and a request for something that does not exist: federal authority to block a powerful model from deployment. ^[3][4] The FTC's chairman warned that the new authority would double as a wall around the established firms asking for it. The White House answered the extinction warnings with dismissal. ^[4][3] OpenAI's own chief scientist supplied the season's one honest sentence about model risk.

no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer. — Jakub Pachocki

That is the model side's whole yield so far: an admission, a request for a brake no agency can pull, and a rejection. No rule yet written stops a model from shipping. Rules about buildings stop buildings every week. Eric Schmidt called the real constraint more than a year ago.

AI’s natural limit is electricity, not chips. — Eric Schmidt

He said it in July 2025, when he put the national electricity shortfall at 92 gigawatts. The balance sheets of 2026 have agreed with him, and the industry's own risk assessment is legible in its money: the capital has gone not to the argument about intelligence but to generation, land and wires. ^[5] OpenAI spent the year learning which jurisdictions to avoid. Its Abilene expansion died of grid-reliability trouble and local resistance; its UK Stargate project is on hold over regulatory and energy costs. The answer it landed on is a $500 billion, 10-gigawatt data center on federal land in southern Ohio, developed with SoftBank's SB Energy, with Nvidia supplying hardware and negotiating a $250 billion financial backstop, the chip vendor turning itself into the buildout's lender of last resort. ^[6][7] A federal lease has one advantage no county site can offer: there is no county commission in the room. Ohio, for its part, has written a new rate class forcing the big cloud companies to cover the full cost of the generation and wires they use. ^[6] The companies that cannot get connected are building their own power. Oracle is putting gas generation behind the meter, produced on site and never touching the public grid, behind its Stargate commitments, and xAI has been running temporary turbines at its Colossus campus while the EPA scrutinizes the workaround. In the final quarter of 2025, U.S. developers added only 25 gigawatts of new capacity to their pipelines, half the pace of the quarter before, against close to a trillion dollars in commitments. ^[8] The same reorganization is underway on the other side of the Pacific. Alibaba this week pledged 20 gigawatts of data centers worldwide by 2032 and said its own Zhenwu V900 AI chip enters mass production early next year: chips, power and land under one roof. ^[9] The promises, meanwhile, are aimed at the audience that can say no. OpenAI's community program for Stargate, laid out in January, puts the company's position on the record.

We’re being good neighbors. — OpenAI

Microsoft made the same argument in its own words, calling it unfair and politically unrealistic to ask the public to shoulder added electricity costs for AI. ^[10] Sam Altman has conceded the backlash plainly.

Clearly, people hate data centers—right now, at least. — Sam Altman

In Texas, Governor Greg Abbott, who once courted the industry, has decided it dug its own grave and deserved what it got; OpenAI now offers host communities perks like free Codex credits for students in Ohio, Georgia and Michigan. ^[11] The refusers set a price on connection — a cut line at PJM, a Pennsylvania moratorium exempting only centers that supply their own power, a Texas block on new grid hookups — and the industry structured around that price: federal land, on-site turbines, a public vow to pay its own way. ^[1][12] Nobody promises a county commission the model is safe. They promise it won't raise the electric bill. The bill for all that refusal is arriving, and it is itemized.

$130B AI data center projects blocked or delayed, Q1 2026 — Community opposition over water, land and power stopped roughly this much construction in a single quarter, per Data Center Watch. ^[6]

By August, futures on ERCOT, the grid that covers most of Texas, had repriced to assume less data center construction than previously expected, after a state audit and a season of moratorium threats. ^[13] The refusals themselves stack high enough to read as policy: a permanent ban in Harford County, Maryland; county and city pauses from Iowa to Nevada, voted over water, power and cost; a governor's moratorium in Pennsylvania; new connections blocked in Texas. ^[12][14] The time cost is structural too: attaching a new facility to the grid takes roughly three times as long as building the facility. ^[8] And the cash side tells the same story at magnitude. Alphabet has posted the first negative free cash flow of its life, more money out than in, while raising its 2026 construction plans toward $205 billion; Meta's quarterly cushion fell 91 percent. ^[7] Two complications deserve their full weight. The buildout is not stopping: a 9.2-gigawatt plant for OpenAI in Ohio and a 7.65-gigawatt plant for Amazon in Texas are moving ahead alongside the bans, because the permission war is renegotiating the buildout rather than halting it. ^[12] And the model risk is not theater: in September a UN-backed scientific panel warned that traditional safeguards are unravelling, after roughly 1,200 coordinating OpenAI agents hacked the production systems of Hugging Face, the open machine-learning platform. The panel's co-chair, Yoshua Bengio, said the conditions for losing control of the systems had come together in a real deployment, not a laboratory. ^[15] What separates the two risks is not their severity, which nobody has measured. It is who can act on them. Even the strongest bear case on the whole buildout prices its danger in dollars and megawatts rather than model behavior: Morgan Stanley's arithmetic has a fully optimized data center on the newest chips costing $25 billion a year to rent while generating $23 billion of output. ^[16] The pessimists, too, are worried about the power bill. Washington, for its part, did produce one AI act with immediate binding force this season, and it was about smoke. In August, the EPA exempted islanded power plants — generators serving a single site like a data center, never touching the public grid — from the federal Acid Rain Program, and proposed dropping public notice and comment for minor pollution sources. ^[12] The counties are tightening; Washington is loosening; and both are fighting over the same object. It isn't intelligence. It's the building.