The Signal
On July 31, 2025, PJM Interconnection cleared its 2026/2027 base residual auction at $329.17 per MW-day. Two years earlier, the same auction cleared at $28.92. That is a 1,038% move in the price of capacity — the right to draw power when the grid is stressed. No model weights changed. No wafer was fabbed. The single most important input to AI expansion repriced by an order of magnitude, and the equity market filed it under "utilities footnote."
The article that triggered this analysis came from Crypto Briefing. Its claim: US utilities "may cut data centers during peak demand due to AI power strain." One factual assertion. Three unsourced inferences. No named source, no jurisdiction, no number, no date. It is a headline with a verb attached.
That does not make it false. It makes it unverifiable. And in a bear market, an unverifiable claim is a liability, not a signal. So I went to the ledger. The ledger does not lie, only the interpreters do.
What follows is a teardown of the actual mechanism the headline gestured at. Not the panic. The plumbing.
Context: Demand Response Is Not New — That Is the Entire Point
Before anyone prices a "regulatory shift," establish what already exists. Interruptible load is not a 2025 invention. American grids have run demand-response programs for decades. Aluminum smelters, cement kilns, chlorine plants, and — critically — Bitcoin mines have all signed contracts trading interruptibility for a discount. A large industrial user accepts that the utility can shed its load during peak hours. In exchange, the user pays a lower tariff and often collects capacity payments for standing ready.
So the phrase "utilities may cut data centers during peak demand" describes an existing commercial tool being pointed at a new class of customer. The tool is old. The customer is new. The fracture is in the mismatch between them.
Here is the scale of the new customer. Lawrence Berkeley National Laboratory estimates US data center consumption at roughly 176 TWh in 2023 — about 4.4% of national electricity. By 2028, LBNL projects 325 to 580 TWh, or 6.7% to 12% of the total. The International Energy Agency put global data center demand near 460 TWh in 2022, with a 2026 range of 620 to 1,050 TWh. That is high-double-digit compound growth against a transmission system whose new-build cycle runs five to ten years.
That gap is the entire story. Demand compounds annually. Supply is built in five-year increments. The mismatch is arithmetic, and arithmetic does not negotiate.
Now the part the Crypto Briefing piece never mentioned. Crypto miners were the textbook interruptible load of the last decade. They chased cheap power, signed curtailment agreements, and monetized their flexibility. In 2024 and 2025, they were displaced — not by regulation, but by a higher bidder. AI and HPC data centers pay more per megawatt-hour than hashing does. Core Scientific, once a pure Bitcoin miner, now leases capacity to CoreWeave. The miners are being evicted from their own power contracts, and they are inheriting the very curtailment exposure they once exploited. That inversion is the real crypto-relevant signal. The headline buried it.
Core: Three Fractures the Headline Compressed Into One Word
Fracture One — The Interruptibility Mismatch
Here is where the existing contract template breaks. A smelter can be cut for four hours and lose four hours of aluminum. A Bitcoin miner can be cut and lose four hours of hashes. Both are linear, loss-tolerant loads.
AI load is not linear. Split it in two.
Training is quasi-interruptible, in theory. Modern distributed training writes checkpoints. A cluster can, with orchestration, be paused and resumed. In practice, an abrupt shed of a multi-thousand-GPU job wastes synchronization work, risks corrupted states, and burns expensive wall-clock time. The tolerance is not zero, but it is far below a smelter's.
Inference is effectively uninterruptible. A customer-facing model serving an enterprise SLA cannot go dark because the grid hit its afternoon peak. That is a breached contract, not a scheduling inconvenience.
So the two halves of the AI workload have opposite interruptibility profiles, and the legacy demand-response contract was never designed for either. A single "cut the data center" lever treats a loss-tolerant industrial process and a latency-bound revenue service as the same object. They are not the same object.
This is the systematic flaw. The grid's scheduling logic assumes load can be shed. The AI revenue model assumes load cannot be shed. The contract that binds them has not been rewritten to reflect that the two parties disagree about what the load even is.
Fracture Two — The Queue, Not the Peak, Is the Binding Constraint
The headline fixates on curtailment. Curtailment is the small problem. The large problem is the interconnection queue.
A new data center does not merely draw power; it must be granted the right to connect. In several US regions, that queue now runs into the multiple-year range. This is the true geographic ceiling on AI compute expansion — not the peak-hour cut, but the multi-year wait to be permitted to plug in at all. A curtailment is a temporary tax. A queue position is a structural gate.
The article's framing — "utilities may cut data centers" — is directionally correct and operationally trivial. Of course a stressed grid sheds load. The question that matters is which data centers get built, where, and how fast. That answer lives in the queue, in the PPA market, and in FERC's docket. None of it is in the headline.
I have audited incentive systems before — the Curve gauge model, where I showed mathematically that reward distribution favored whale wallets absent slippage protection. The lesson transfers. When you cannot see the queue, you cannot price the asset. The curtailment is the visible symptom. The queue is the hidden variable.
Fracture Three — Who Pays for the Fixed Cost
Then there is co-location. Amazon contracted directly with Talen's Susquehanna nuclear plant. Microsoft signed to restart Three Mile Island via Constellation. These "behind-the-meter" arrangements connect a large load straight to a generator, sidestepping the public grid's shared cost structure.
This has landed at FERC as a live rulemaking question: who pays for the transmission and capacity that the grid maintains, if the biggest loads bypass it? If a hyperscaler buys a nuclear plant's output directly, the fixed cost of the public network does not disappear. It reallocates. To whom? To residential and small-commercial ratepayers.
That reallocation is the political fuel underneath every "moratorium" headline. Virginia's "data center alley," Georgia, Indiana, Ohio — all have seen local pause motions and rate cases over who bears the load-growth cost. The PJM capacity price spike does not stay in an auction print. It flows to the consumer bill. And consumer bills are politically radioactive.
So the fracture is not technical. It is distributive. The grid is being asked to socialize the fixed cost of infrastructure whose returns accrue to a handful of firms. That is not an engineering problem. It is a cost-allocation fight, and cost-allocation fights are settled by regulators, not by physics.
The Value Transfer Nobody Priced
Step back and the shape is clear. This is not a story about AI being punished. It is a story about value moving from the compute side to the power side.
Owners of dispatchable generation, grid capacity, and interconnection rights are acquiring pricing power they have not held in twenty years. A nuclear plant that was a stranded asset in 2020 is now an AI-infrastructure growth asset in 2025. A gas peaker that existed to meet rare peaks now has a hyperscaler willing to pay for firmness. A transformer manufacturer that was a boring industrial is now on the critical path.
Meanwhile, the colocation operator that depends on the public grid faces rising power costs and curtailment risk with weak ability to pass them through. The hyperscaler that signs long-term PPAs and builds its own generation absorbs the shock. The gap between those two positions is the new moat. It is a power moat, and it is deeper than a chip moat because it takes longer to dig.
History repeats, but the gas fees change. In 2021 the scarce resource was block space and the winners were those who controlled it. In 2025 the scarce resource is firm megawatts, and the winners will be those who control the interconnection.
The Mining-to-AI Arbitrage
I want to name the commercial opportunity the Crypto Briefing piece walked past, because it sits at the exact intersection of crypto and this power question.
A Bitcoin miner holds an already-interconnected power contract, often with curtailment rights, often in a power-rich region. That contract is now worth more as an option on AI capacity than as a hashing operation. The miner's real asset was never the ASICs. It was the queue position and the interconnect. Selling or converting that position to HPC is a pure capacity arbitrage — the same power, repriced against a higher-value workload.
This is why the story surfaced on a crypto outlet. The subject matter is not coincidental; it is a miner-displacement narrative wearing an AI headline. But the piece never states the mechanism. It presents a conflict and omits the arbitrage that resolves it. Code is law; intent is irrelevant — and the code here says the megawatt flows to whoever pays the most per unit of firmness.
Contrarian: What the Bulls Got Right, and Where the Bears Overreach
The bears are directionally right and structurally lazy. They read "data centers may be curtailed" and conclude AI expansion stalls. That conclusion overreaches.
First, interruptibility can be engineered, and engineering it creates revenue. A data center that installs storage, on-site generation, and orchestration can sell its flexibility back into the capacity market rather than merely suffer curtailment. The same asset that looks like a liability under the old industrial contract becomes a demand-response product under a new one. The bears priced the cost and ignored the offsetting payment.
Second, utilities have a strong financial incentive to serve these loads, not to shed them. Load growth expands the regulated rate base, and the rate base is how a regulated utility earns its return. A utility that curtails its largest new customers is a utility that is strangling its own growth. The incentive gradient points toward accommodating load, not rejecting it.
Third — and this is the bull case the headline buries — the political backlash is a feature for incumbents, not a bug. Moratoria and rate cases raise barriers to entry. They favor the player who already holds the interconnect and the PPA over the new entrant still waiting in the queue. The constraint does not kill the market. It concentrates it.
Where the bears are right is narrow and sharp: the binding constraint is not physics, it is allocation politics. The real risk is not that AI gets cut off at peak. The real risk is that the cost of guaranteeing AI's firmness gets socialized onto retail ratepayers, and that the resulting political recoil — not the grid — becomes the thing that slows deployment. The bulls who model AI power as a pure supply-and-demand curve miss that the clearing mechanism is a PUC hearing.
Compliance Checklist
Before you underwrite any name in this theme, verify the following, in order:
- Interconnect status: Does the asset hold an executed interconnection agreement, or a queue position? The difference is the difference between an asset and a hope.
- Contract type: Is the power supply a firm PPA, a behind-the-meter co-location, or spot exposure? Spot exposure is not an AI moat.
- Curtailment terms: Who bears the curtailment risk, and what compensation attaches? Read the actual clause, not the summary.
- Regulatory exposure: Which PUC jurisdiction governs, and is there a pending rate case or moratorium motion?
- Pass-through ability: Can the operator pass power cost to the customer, or is margin fixed?
If any answer is "not disclosed," treat the position as narrative, not fundamentals. Trust is a bug, not a feature.
Takeaway
The headline said utilities may cut data centers at peak. The ledger says the peak is a distraction. The real constraint is the interconnection queue, the contract template that has not been rewritten for AI's interruptibility profile, and a cost-allocation fight that will be settled by regulators rather than by megawatts.
The forward question is not whether AI gets curtailed. It is who captures the rent on firm power — and whether the public is asked to pay for the firmness that private returns depend on. Watch the queue. Watch the rate cases. Watch the FERC docket on co-location. The curtailment is noise. The allocation is the signal.
