The Power Wall

The Thesis

On 18 July 2026 a grassroots group called HumansFirst ran 142 protests across 42 US states against data-center construction. Four days earlier New York’s governor had signed a one-year moratorium blocking new environmental permits for any data center above 50MW — the first state to do it. That same week The Information reported that Oracle, unable to get gas turbines permitted for its Project Jupiter site in New Mexico, had switched to costlier fuel cells at a cost of billions more than planned.

None of those three stories is about money. The AI buildout has no trouble raising money — that is the other pillar, and the loop described there keeps tightening. What it has trouble with now is electricity, and more precisely with permission to use electricity. Developers face grid connection queues of up to seven years. So they leave the grid: on-site turbines, fuel cells, small reactors, anything that can be built inside a fence line. But leaving the grid does not remove the constraint. It converts a utility problem into a permitting problem, and permits are granted locally, and locally is exactly where consent has collapsed. The escape route runs into the same wall from the other side.

That wall is what this page tracks — how fast it went up, which inputs are actually binding, and which listed names own something on the scarce side of it. It does not forecast where any of these stocks go.

The Evidence

The cleanest measure is not a dollar figure. It is how much of the record is now about fights rather than plans.

Coverage in the TEXXR archive that pairs data centers with permits, zoning, moratoriums, ordinances or organized opposition ran at 2 articles in 2024 Q2. In 2026 Q2 it ran at 22 — eleven times as much in eight quarters. In the first three weeks of 2026 Q3 alone it is already at 11, on pace to pass the record quarter again. Coverage pairing data centers with power, grid, megawatts, turbines or nuclear grew on the same clock but from a bigger base: 11 articles in 2024 Q1, 79 in 2026 Q1.

That growth rate matters because it outruns the money story. Coverage of AI data-center capex grew 6.3x over nine quarters, as the capex pillar sets out. The constraint is being written about faster than the spending.

Prices made it political. Power on PJM, the largest US grid, rose 76% year over year to an average $136.53/MWh in the first quarter, which Bloomberg attributed directly to data-center demand. In July the grid’s own market monitor calculated that PJM’s latest capacity auction would add $6.3 billion to customer bills across 13 states and Washington DC through 2029. PJM had already told large data centers in January they would have to bring their own generation or curtail, and expects demand to grow 4.8% a year for a decade. A bill arriving in thirteen states is what turns a zoning argument into an election argument, and Politico reports the industry now fears exactly that: New York’s pause, it wrote, could embolden more state restrictions heading into the midterms.

Consent fell first, and fast. In Virginia, home of the largest data-center cluster on earth, the share of voters comfortable with new construction fell to 35%, from 69% in 2023. A national Quinnipiac poll in March found 65% opposed building data centers in their own community. The geography explains the friction: Pew data cited by the Financial Times found 67% of planned data centers are rural while 87% of existing ones sit in urban and suburban areas. The industry is moving into places that have never hosted one and have no local precedent for saying yes.

The blocking is now measurable. A June study reported by NBC News found opponents had blocked or delayed at least 75 US projects worth roughly $130 billion in a single quarter, with organized opposition groups doubling to 833 across 49 states. Seattle’s council voted 9-0 for a one-year moratorium in June. New York was already at least the sixth state to introduce pause legislation when it filed its bill in February. Even a rejection is instructive: Maine’s governor vetoed a similar bill in April, but on the narrow ground that it failed to exempt one project in a distressed mill town — the objection was to the drafting, not the premise.

The workaround is generation you own. The Wall Street Journal was documenting on-site power plants built to bypass the grid as early as October 2025. By that December, developers were reaching for aeroderivative turbines and diesel generators. Musk took it a step further and quietly bought APR Energy, a mobile-turbine operator, for an implied $1 billion — buying the generation company rather than waiting for a utility. And it is precisely there that the second wall shows up: Reuters found xAI had installed far more turbines without permits at its Memphis site than it acknowledged, with the pollution falling hardest on Black neighbourhoods. Escaping the interconnection queue put the company in front of the EPA and a civil-rights lawsuit instead.

Even the equipment is short. A domestic shortage of transformers and switchgear has forced US builders to rely on Chinese imports, delaying construction — a supply chain that sits awkwardly beside the export-control politics running through the China stack.

The record carries its own check. In January, PJM cut its summer 2027 peak demand forecast to about 160GW from 164GW, because some announced projects lacked firm service or construction commitments. Announced gigawatts and contracted gigawatts are not the same number, and the grid operator is the party with the least incentive to confuse them. Anyone reading the wall as a guaranteed shortage should read that revision first.

The Companies

The names on the scarce side of this wall are not the ones spending the most. They are the ones holding a grid connection, a signed permit, or a way to make power without either.

Applied Digital and TeraWulf are the archetype: bitcoin miners that already held interconnection agreements and energised shells, converting them to AI compute. Their asset was never the rigs. It was queue position — a thing capital cannot buy forward. TeraWulf’s July lease with Anthropic, a 20-year deal worth about $19 billion for roughly 400MW in Kentucky, prices that position explicitly. CoreWeave sits one step downstream, leasing capacity rather than owning the interconnect, which is why the Abilene site it lost to a rival mattered: Microsoft entered talks for hundreds of megawatts there only after Oracle walked away from its expansion. The same building, two tenants, no new megawatts.

Two names join the registry with this pillar. $IREN, another miner conversion, announced a deal with Nvidia in May to deploy up to 5GW of AI infrastructure, with Nvidia holding the right to invest $2.1 billion. $BE — Bloom Energy — is the fuel-cell maker Oracle turned to when the turbines would not clear: their partnership expanded in April to as much as 2.8GW of capacity, days after Oracle received a warrant to buy $400 million of Bloom stock. That warrant is the AI-capex loop showing up inside the power layer: the customer taking equity in the supplier it depends on.

Around them sit the category baskets, each answering the same question from a different side — who sells the electrons, who makes the gear the grid is short of, who is building the twenty-year answer to a two-year problem, who holds the land and the interconnect, who converted the rigs, and who fills the gaps in the load. Treat every one of them as a single bet wearing several tickers; each basket page reports its own internal correlation for that reason.

The demand side is unchanged in its ambition. Zuckerberg announced Meta Compute in January with a target of “tens of gigawatts” this decade. France booked €110 billion of proposals in a single week, about 10GW — as the FT put it, roughly ten nuclear reactors’ worth. The gap between that ambition and a county planning board is the whole subject of this page.

The Lenses

The Helmer lens reads a grid connection as a textbook cornered resource — one of the seven powers, defined as preferential access to a coveted asset on terms unavailable to rivals. Most claimed cornered resources dissolve on inspection, because capital eventually buys an equivalent. An interconnection agreement is the rare one that does not: the queue is administered by a party with no profit motive to clear it faster, and the wait is measured in years no balance sheet can compress. That is the strongest structural argument the miner-conversion names have, and it is worth stating plainly what would break it — a federal or state process reform that shortens queues, which is precisely what the White House was drafting toward in February and what Trump and Northeastern governors asked PJM to do in January. A cornered resource that policy can uncorner is contested, not held.

The Shiller lens points at the other side of the ledger. Narrative economics treats a story about an asset as contagious, spreading person to person and changing behaviour as it spreads. Opposition groups doubling to 833 across 49 states in a year is contagion with a number attached, and the mechanism is visible: an electricity bill arrives, a local paper connects it to a building, a template ordinance travels to the next county. Microsoft read this early, pledging to pay its own power costs and refuse local tax breaks in January — an attempt to break the story rather than argue with it. Whether that works, or whether the narrative has already passed the point where any single builder’s conduct matters, the record has not yet answered.

What Moved

Sources