PJM Interconnection's capacity auction for the 2025/26 delivery year cleared at $269.92 per megawatt-day. The prior year cleared at $28.92. That is a ninefold repricing of the right to be served by the largest organized electricity market in North America, and the auction's own documentation attributes the jump primarily to data center load. Forget the ETF flows. Forget the halving. The most consequential number printed in the American energy complex this cycle was a capacity price, and almost nobody in crypto read it as a crypto story. They should have. The mechanism now squeezing AI data centers is the identical mechanism that once squeezed Bitcoin miners, and the outcome of that squeeze will determine which of them keeps the lights on.
The trigger for this analysis is a short item from Crypto Briefing reporting that US utilities may curtail data centers during peak demand because of AI power strain. The piece is thin. One factual claim, a handful of unsourced inferences, no named sources, no geography, no dates. Treat it as a topic signal, not an evidentiary document. The signal, however, is real, and it points at something the market still prices incorrectly.
Before the assertion, the arithmetic. US data centers consumed roughly 176 TWh in 2023, about 4.4% of national electricity. Lawrence Berkeley National Laboratory projects 325 to 580 TWh by 2028, or 6.7% to 12% of the grid. The IEA puts global data center consumption near 460 TWh in 2022, with a 2026 range of 620 to 1,050 TWh. Those are compound growth rates in the high double digits. Transmission line construction runs on five-to-ten-year cycles. The mismatch is not a forecast. It is a schedule conflict, and schedules do not negotiate.
The mechanism itself is not new. Demand response and interruptible load have been standard commercial arrangements in the United States for decades. Large industrial users โ aluminum smelters, cement kilns, cold storage โ sign contracts that allow the grid operator to curtail them at peak in exchange for discounted power. The Crypto Briefing item frames this as a possible regulatory shift. It is not a shift. It is an old tool being pointed at a new class of customer. What matters is not the tool. It is the customer.
That customer is the AI data center, and the customer's defining property is that it cannot be interrupted without destroying value. This is the whole story, and it is the part the coverage misses.
The interruptibility gap is the real structural conflict, not the curtailment headline. Traditional interruptible load can absorb a curtailment order because the product is a commodity with no time value. An aluminum smelter that loses four hours of power loses four hours of aluminum. The metal is still worth the same tomorrow. It has no expiration. An AI training cluster that loses four hours loses four hours of compute, which it can partially recover from a checkpoint, at the cost of wasted GPU-hours and synchronization overhead across thousands of accelerators. An inference endpoint serving an enterprise SLA loses four hours of contractual availability, which is a refund, a churn event, and a reputation that does not come back. The product is not a commodity. It is a continuous promise, and the promise is the revenue.
Power, like liquidity, flows to the path of least resistance until the path is saturated. The grid was built to route electrons to loads that tolerate interruption. The AI load does not tolerate interruption. So the grid has to choose between two incompatible optimization functions: the physical requirement that supply equal demand in real time, and the commercial requirement that the largest new class of demand never be switched off. The existing interruptible tariffs were never designed for loads that are simultaneously enormous and inelastic. That is the gap. The headline says utilities may cut data centers. The structural truth is that the contract architecture for doing so does not yet exist in a form that AI customers will sign.
I spent 2017 auditing the liquidity reserves of ICO tokens and treating them as balance sheet instruments rather than technology experiments. The same discipline applies here. A data center is not a technology asset. It is a balance sheet with a power contract attached, and the power contract is now the dominant line item. Power typically runs 30 to 40 percent of data center operating cost. AI training racks push power density to 30 to 100 kilowatts per cabinet, against 5 to 10 kilowatts for the traditional enterprise load. That density compresses the tolerance for any interruption and inflates the cost of the backup systems required to survive one. The balance sheet does not care about the model architecture. It cares about the megawatt.
The crypto bridge here is not incidental. It is the core of the story. Bitcoin miners were the archetypal interruptible load of the last decade. They signed cheap power contracts precisely because they agreed to be curtailed at peak, and they built their economics around the fact that a stopped ASIC loses almost nothing but time. That property made miners the ideal grid customer during a supply glut and the ideal scapegoat during a supply squeeze. Now the AI and high-performance computing buildout is outbidding them. Core Scientific's pivot toward CoreWeave hosting is the template. Miners with firm interconnection rights and energized sites are being acquired, converted, or displaced. The AI complex is taking the megawatts, and in the process it is inheriting the curtailment exposure the miners used to own. Crypto Briefing's interest in this story is not random. The publication's audience holds the assets that are being repriced out of the power stack.
The conventional read of this dynamic is that AI is stealing power from miners. The accurate read is that AI is inheriting the miner's grid position while losing the miner's flexibility. A miner can be curtailed and shrug. An AI data center cannot. So the entity that best tolerates interruption is being replaced by the entity that worst tolerates it, on the same interconnection queue, under the same grid rules. That substitution is the source of the entire conflict.
Here is what I mapped during the 2022 Terra collapse. Contagion does not announce itself. It travels through the collateral and the counterparties, not through the headlines. The Terra de-pegging was a stablecoin event that became a lending event that became a centralized exchange event. The power constraint is the same shape. A capacity auction clears nine times higher, which reprices every power purchase agreement, which reorders every data center's site selection, which reorders the interconnection queue, which reprices the land and the transmission rights, which reprices the utilities and the equipment suppliers. The curtailment headline is the visible tip. The repricing underneath it is the contagion.
The more binding constraint than peak curtailment is the interconnection queue. A new data center does not simply connect to the grid. It applies, and it waits. In several US regions the wait for a large load to be studied, approved, and energized now runs into multiple years, and the queue itself is congested with gigawatts of proposed load that will never be built. This is the true geographic ceiling on AI compute expansion. The bottleneck is not silicon supply. It is the permission to connect. You can buy a GPU on a lead time of months. You cannot buy an interconnection right on a lead time of months. You inherit one or you wait. That asymmetry is what makes existing energized capacity the scarcest asset in the buildout.
Which brings the co-location fight into focus. Amazon's arrangement with Talen Energy at the Susquehanna nuclear plant, and Microsoft's agreement tied to the restart of Three Mile Island through Constellation, are behind-the-meter structures that connect a hyperscaler directly to a generator, bypassing the public grid for the bulk of the load. These deals have triggered active controversy at the Federal Energy Regulatory Commission over a single question: who pays for the fixed costs of the transmission system when the largest customers leave it? If a hyperscaler takes a nuclear plant's output directly, the plant's capacity is no longer available to serve the public grid, but the transmission network that everyone else depends on still has to be maintained. The cost does not disappear. It redistributes, usually onto residential ratepayers.
The co-location debate is where the technical and the political fuse. It is not really an engineering dispute. It is a cost-allocation dispute dressed as an engineering dispute. And cost allocation is where the political economy of this entire trend becomes unavoidable.
Electricity prices are the most politically sensitive prices in any democracy. A capacity auction that clears nine times higher does not stay inside a market report. It flows into household bills, and household bills produce regulators, and regulators produce rate cases. Ohio, Georgia, and Virginia have already seen contested proceedings over large-load tariffs and who bears the cost of serving data centers. Local moratoriums on new data center construction have appeared in Virginia's data center corridor, in Georgia, and in Indiana, driven by noise, water consumption, land use, and the fear of rate increases. The Crypto Briefing item's reference to moratoria is grounded in real local politics, even though the item provides no examples.
The water dimension is the blind spot the power framing hides. Data center cooling consumes water, and in water-stressed regions that consumption is a more sensitive local issue than electricity. A community can tolerate a transmission line. It is far less tolerant of a cooling tower drawing down an aquifer. Any analysis that treats this as purely an electricity story is reading half the constraint. The local externalities โ water, noise, land, rates โ are the political fuel, and the curtailment mechanism is the exhaust.
I ran the design of a cross-border B2B settlement pilot in Seoul in 2024, moving $50 million of test transactions from T+2 to T+0 on a hybrid CBDC and tokenized deposit model. The lesson that transferred was not about settlement speed. It was about how quickly an institutional rail reconfigures once the constraint moves. When you remove a friction, the value does not stay where it was. It migrates to the next bottleneck. Power is now the friction, and the value is migrating toward whoever controls the megawatt.
Centralization is the inevitable entropy of scale, and the power constraint accelerates it. A hyperscaler with the balance sheet to sign a fifteen-year power purchase agreement, self-generate, and participate in demand response can absorb a ninefold capacity price increase. A third-party colocation provider renting space on the public grid cannot. The same shock that pressures the smaller operator hands the larger one a moat. The constraint does not level the field. It tilts it. The entities that already hold interconnection rights, energized sites, and long-dated power contracts are the ones who benefit from a narrative that makes those assets scarce. That is not a side effect. It is the mechanism.
The value transfer runs in one direction and it is legible on a chart. Independent power producers with nuclear and gas fleets โ Constellation, Vistra, Talen, NRG โ have been repriced as AI infrastructure rather than low-growth utilities. Regulated utilities sitting on data center load growth โ Dominion in Virginia, AEP, Southern Company โ see their rate base expand as load grows. The electrical equipment supply chain โ transformers, switchgear, cable, turbines, the GE Vernova and Eaton and Quanta Services layer โ is a hard requirement for any grid expansion and benefits regardless of which hyperscaler wins. Storage and demand response providers sell the very interruptibility the AI load lacks, which makes them the arbitrageurs of the gap.
On the other side of the ledger sit the pure colocation real estate investment trusts, exposed to power cost inflation and curtailment risk without the generation assets to hedge them, and the Bitcoin miners who fail to convert their energized capacity into AI or HPC hosting. The miner that holds a firm interconnection and a dispatchable load is no longer just a miner. It is a power asset with a hashrate attached, and the market has only begun to price it that way.
The contrarian read is that the bearish framing of this story is exactly backwards. The headline says AI is being curtailed, which sounds like a negative for AI. The structural truth is that this is a repricing of the power supply side, which is a negative for nobody who owns generation and a windfall for whoever holds the interconnect. The decoupling thesis is not that AI and power move together. It is that power has quietly decoupled from the utility index and re-coupled to the AI capex cycle, and the market is still trading it as a defensive bond proxy.
I have seen this narrative construction before. In DeFi, "liquidity fragmentation" was not a problem discovered in the wild. It was a problem manufactured to sell the product that claimed to solve it. The same pattern is visible here. The "AI power crisis" is partly a constructed narrative, and the construct serves whoever already holds the scarce asset. A crisis narrative that makes interconnection rights and power contracts more valuable is a crisis narrative written by their owners. The scare reprices their balance sheets upward. When you see a problem framed as existential, ask who holds the solution and whether they wrote the framing.
The crowded-trade warning belongs here. By the middle of 2025, the AI power theme is no longer undiscovered. The independent power producers have been repriced. The equipment names have been repriced. The marginal investor now has to separate genuine exposure โ a signed power purchase agreement, a self-generation asset, a firm interconnect โ from concept exposure, which is a press release and a stock chart. That separation is the same discipline I applied in 2020 when I wrote that the yield-farming incentive structures were unsustainable and predicted a 70 percent drop in the headline APYs of the major farms within six months. The yield was real until it was not. The power premium is real until it is priced. The question is not whether the trend is true. It is whether you are early or late.
The 2026 AI-agent payment layer I helped build for Seoul Blockchain Week is the forward edge of this. We deployed a testnet where autonomous agents negotiated and settled data transactions, processing more than ten thousand transactions a day. Every one of those agents consumes compute, and every unit of compute resolves to a watt. When machines become primary economic actors, the demand they generate is not elastic to price the way human demand is. An agent does not decide to use less power because the capacity price rose. It executes its objective. That inelasticity is the reason the interruptibility gap is structural rather than cyclical, and it is the reason the grid, not the GPU, is the binding constraint on the next phase of the buildout.
The market still prices compute as the scarce input. The balance sheet says otherwise. The scarce input is the firm, dispatchable, interconnected megawatt, and it is scarce precisely because it cannot be manufactured on a chip fab's schedule. The value transfer from the compute supply side to the power supply side is not a trade. It is a repricing of the physical layer that everything else runs on.
Watch three signals to know whether this is early or late. The capacity auction results in PJM, ERCOT, and MISO set the price of the constraint each season. The state public utility commission rulings on large-load tariffs and cost allocation determine who absorbs the political cost, and Ohio, Georgia, and Virginia are the leading indicators. The FERC rulings on behind-the-meter co-location decide whether the hyperscalers can keep building the moat or whether the public grid forces them to pay for it. Those three data streams are the true order book for this trade, and they update on a schedule no exchange publishes.
A cycle is a transfer of scarcity from one layer to another. The last cycle moved scarcity from capital to block space. This cycle is moving it from block space to the megawatt. The miners who understood interruptibility are positioned to sell it. The utilities that understood it are positioned to reprice it. Everyone else is standing in a queue, waiting for permission to connect, watching a capacity price they did not read until it hit the headline. When the grid becomes the ledger, the only question that matters is the one nobody in the market is asking: who holds the interconnection, and who is paying the fixed cost of the wire it hangs on?