The entire article is eleven words. "US utilities may cut data centers during peak demand due to AI power strain." No line number. No commit hash. No named utility, no region, no megawatt figure, no threshold, no compensation clause. Three of those words — "may," "peak," "strain" — are the only load-bearing elements, and all three are unquantified. I have spent four weeks tearing apart source code with more verifiable internal structure than this paragraph. Code does not lie, but it often omits the truth. And yet the signal underneath the omission is real, because what it describes is not a rumor. It is an arithmetic inevitability that one market is pricing in real time — a market almost none of the article's readers have ever opened.
To understand why a five-line industry brief matters, you have to understand what it is standing on top of, and what it is carefully stepping around. The AI buildout has been narrated for two years as a semiconductor story. GPUs are the scarce good; Nvidia is the toll booth; the constraint is lithography and CoWoS packaging. That narrative is now stale. The binding constraint has migrated from the fab to the substation, and the migration is visible in a single, boring, public number that no hype cycle can massage.
Start with scale. US data centers consumed roughly 176 TWh in 2023 — about 4.4% of national electricity. Lawrence Berkeley National Laboratory projects 325–580 TWh by 2028, or 6.7% to 12% of the grid. The IEA estimates global data center demand at roughly 460 TWh in 2022, potentially reaching 620–1050 TWh by 2026. These are not forecasts in the marketing sense. They are load forecasts, the kind utilities are legally required to file, and their compound growth runs in the high double digits annually. Transmission line construction runs five to ten years. The mismatch is not a sentiment problem; it is a scheduling problem, and schedules do not negotiate.
Now the price signal, which is the only part of this story that cannot be spun. PJM, the largest US grid operator, cleared its capacity auction at $28.92 per MW-day for 2024/25. For 2025/26 it cleared at $269.92 — roughly nine times higher. For 2026/27, $329.17. Capacity markets are not vibes. They are procurement, and the procurement is screaming that the system has entered structural tightness. When I built a discrete-event simulation of Impermax's reward distribution in 2020, I was not predicting a mood — I was computing the crossover point where impermanent loss outpaces farming yield. The PJM auction is the same class of artifact: a number that tells you where the crossover sits, whether or not you want to hear it.
Here is where the headline gets lazy. "Cutting data centers during peak demand" is not a new regulatory posture. Interruptible load and demand response have been standard commercial arrangements in the US for decades. Aluminum smelters, cement kilns, and — critically — Bitcoin mines have long traded the right to be curtailed for a discount on their power rate. The mechanism is not the news. The news is that AI load is structurally incompatible with the mechanism, and nobody has redrawn the contract.
A traditional interruptible industrial load degrades gracefully. Cut an electrolysis line for ninety minutes and you lose throughput; you do not lose the asset. AI load splits into two species with opposite properties. Training jobs can, in principle, checkpoint and resume, but a large cluster dropping to zero abruptly wastes compute and state, and the orchestration overhead of curtailment is itself a cost. Inference — especially anything behind an enterprise SLA — is effectively non-interruptible. You cannot tell a customer that their fraud-detection endpoint was paused because a heat wave hit the mid-Atlantic. So you have a class of load that signed interruptible-style economics while behaving like firm load. That gap is the entire story, and the brief reduces it to "may cut."
I have watched this exact structural blind spot before. When I analyzed TerraUSD seventy-two hours before collapse, the flaw was not a bad actor; it was a circular dependency — LUNA and UST each defining the other's value in a feedback loop that any control engineer would flag as unstable. The AI-power relationship has the same shape. Data center expansion is priced on the assumption of firm power. Grid operators are pricing capacity on the assumption of curtailment rights. Both cannot be true simultaneously. When two contracts assume mutually exclusive states, the resolution is not a compromise — it is a repricing.
The true bottleneck is even further upstream than peak curtailment. It is the interconnection queue. In several PJM regions, a new large load can wait years to energize. That queue, not the GPU allocation, is the geographic ceiling on US AI capacity. And it is why co-location has become a federal question: Amazon's behind-the-meter arrangement with Talen's Susquehanna nuclear plant, Microsoft's restart of Three Mile Island with Constellation — these are attempts to bypass the queue by attaching load directly to generation. FERC is now litigating who pays for the grid's fixed costs when a hyperscaler exits the shared pool. The brief does not mention any of this. Trust is a variable; verification is a constant, and the constant here is that somebody always pays for the wire.
There is a quieter chapter the crypto-native source of this brief should have flagged and did not. Bitcoin miners were the original interruptible load, the flexible buyer of last resort that made marginal power economics work. They are now being displaced by AI and HPC data centers paying multiples of their power rate — Core Scientific's pivot with CoreWeave is the template. AI is not merely consuming power; it is expropriating the grid position miners spent a decade acquiring, and inheriting the curtailment exposure miners once accepted as a feature. This connects to a structural observation I have made before about hash power: after the fourth halving, miner revenue collapsed, and the economics that once justified distributed, flexible mining now concentrate capacity toward whoever can repurpose it. Decentralization rhetoric hollows out precisely when the electricity bill comes due.
Contrarian Angle
Here is what the bulls, for once, got right, and what the brief's conflict-driven framing hides. Every participant in this story treats the power constraint as a cost. Almost none treats it as a revaluation event. When electricity shifts from a controllable operating expense to a strategically scarce asset, the value does not disappear — it migrates to whoever holds the scarce side. Power, not compute, becomes the rent-extracting layer. The independent power producers — Constellation, Vistra, Talen, NRG — and the regulated utilities sitting on data center corridors, Dominion foremost among them, are being repriced from low-growth defensive holdings into AI-infrastructure growth assets. The electrical supply chain — transformers, switchgear, turbines, the GE Vernova and Quanta layer — faces inelastic demand with no substitute. Data centers that own generation and long-term PPAs gain a moat; third-party colocation that depends on the public grid loses optionality. The correct reading of "utilities may cut data centers" is not that AI is under threat. It is that the party with the switch acquires pricing power over the party with the workload. That is a transfer, not a loss.
Kill Switch
I include this in every review, and the conditions are specific. This thesis fails if: (1) interconnection reform materially shortens queue times, dissolving the scarcity premium faster than capacity repricing holds; (2) regulated utilities see returns compressed by political backlash against residential rate increases, capping the revaluation; (3) storage and demand-response technology matures to the point where AI load becomes genuinely interruptible at low cost, restoring the old contract equilibrium; or (4) the "AI power theme" detaches from fundamentals into a concept-stock bubble, where the repricing is narrative rather than cash flow. Watch the PJM, ERCOT, and MISO auctions quarterly. Watch the state PUC dockets on data center tariffs. Watch FERC's co-location ruling. If those move against the thesis, the switch trips.
The brief gave you eleven words and no variables. The variables are public, quarterly, and priced. Hype builds the floor; logic clears the debris. The debris here is the assumption that AI's constraint is silicon. It is not. The next bottleneck is a wire, a queue, and a rate case — and the reader who treats those as the real signal, rather than the headline, will be positioned before the repricing finishes.