PJM Just Put a Ceiling on the AI Boom

PJM Interconnection, the grid operator covering thirteen states from Chicago to the mid-Atlantic, said this week it will start cutting power to large data centers during shortages rather than let the whole system brown out. The Read: this is the first time a major U.S. grid has told the AI industry, in writing, that demand has outrun supply.

The largest electrical grid in the U.S. has struggled to cope with an onslaught of data centers, and after an auction to add more generating capacity fell short, the grid's operator, PJM Interconnection, has said it will cut off data centers and other large users during power shortages. The numbers behind the decision are the story. By 2035, data centers are expected to use 4x more electricity than they do today. The rule itself is narrow and dated: PJM won't start curtailing supply until June 2027, and the cuts will only apply to data centers that are 50 megawatts or larger. The grid operator is running another auction for new generating capacity to try to close the gap before the clock runs out.

Nobody is being punished here. PJM is doing what a well-run utility does: telling its biggest customers the physics of the system before the lights actually flicker. YOUR READ: this is the grid, not Washington or Wall Street, setting the real speed limit on the AI build-out, and it is a limit measured in megawatts, not dollars.

Three stories landed this week that look like they belong in different sections of the paper. Read together, they are one machine.

Story one: the hyperscalers just told Wall Street how much they intend to spend chasing AI compute. Microsoft's capex is projected to nearly triple to $190 billion by fiscal 2027, and Meta's capex for fiscal 2026 rose by over $10 billion to $136.7 billion. Alphabet went further: Google's parent reported $44 billion in capital expenditures during the latest quarter, bringing spending over the past 12 months to $132 billion, roughly double the level of a year earlier, with management guiding for about $200 billion in capex over the next 12 months. Zoom out and hyperscaler AI capital spending has been raised to $750 billion in 2026, up from $670 billion, and is set to cross $1 trillion in 2027. That capital does not buy chips alone. It buys the switchgear, transformers, gas turbines, and prime-power gensets that turn electrons into trained models. Story two: the same companies are quietly becoming nuclear utilities. Every major tech hyperscaler has signed at least one nuclear power deal for AI data center capacity, with 13 announced projects committing over 9.8 GW of nuclear capacity. Microsoft secured a $16 billion, 20 year power purchase agreement for the Three Mile Island Unit 1 restart at 835 megawatts, while Meta leads the pack with up to 6.6 gigawatts across TerraPower's Natrium reactor, Oklo's Aurora, Vistra, and Constellation. Google and Amazon are in it too, with commitments to Kairos Power and X-energy's small modular reactors. Story three is the one from the Loud Thing above: the grid saying, politely, that it cannot connect everything fast enough.

THE THROUGH-LINE, our read, not fact: these are not three stories about tech, energy, and infrastructure. They are one story about a queue. The capex numbers are a bet that compute will be scarce and valuable. The nuclear deals are hyperscalers trying to jump the queue for power by building their own generation rather than waiting on the grid. PJM's curtailment rule is the queue itself, finally speaking. Every dollar of that trillion-dollar 2027 capex number is a claim on electrons that do not yet exist. The company that has already locked in gigawatts, not just GPUs, is the one whose spending actually turns into trained models on schedule.

Start with the grid, not the earnings call, and the AI story reads differently. The industry did not just discover it needs more power. It discovered that power is now the binding constraint, ahead of chips, ahead of capital, ahead of talent. That is why Microsoft, Meta, Google, and Amazon are all, in effect, becoming power companies with software divisions attached.

The machinery under this week's news is not a data center or a rate decision. It is a queue for electrons, and everyone with a trillion dollars to spend just got told there is a line. Watch PJM's next capacity auction: if it clears, the queue moves. If it fails again, expect more hyperscalers to announce their own reactors before they announce their next model.