The PJM capacity market has sent one of the clearest price signals in modern power markets—and one of the least interrogated. Over the past two auction cycles, prices for committing generation during peak demand periods rose from $28.92 per megawatt-day in 2024/25 to $329.17 in 2026/27—more than elevenfold in two years. The 2027/28 auction pushed further still: prices hit the statutory cap at $333.44, and the market remained 6,623 megawatts short of its reliability target, the first shortfall of that kind since 2007. Absent the cap, clearing prices would have approached $530 per megawatt-day.
Article Summary
Those numbers appear to tell a simple story: demand is overwhelming supply. But what they actually encode is a very specific version of the future—a massive, on-schedule buildout of data centers powering an AI economy that scales in a straight line. The market is not just pricing scarcity; it is pricing belief.
That belief is now under pressure from three directions at once. Power consumption per unit of AI output is falling faster than assumed, as efficiency gains compound across hardware, cooling, and model design. Political, regulatory, and community resistance is accelerating, moving from isolated friction to something that increasingly shapes where and whether projects proceed. And the physical supply chain—gas turbines, transformers, semiconductor inputs—is running into hard limits that capital alone cannot resolve. None of these forces has fully worked its way through the lagged system that converts demand signals into capacity prices.
This is the tension at the center of the market. Prices reflect a future that is internally consistent but externally exposed. The question is not whether AI will drive large-scale electricity demand—it will—but whether the timing, scale, and execution embedded in today’s price signal survive contact with these emerging constraints. What follows is an examination of that gap: how expectations become prices, where those expectations are soft, and what happens when the system begins to recognize it.
The Market Is Always Looking Two Years Backward
AI investment sentiment does not flow directly into capacity auctions. It moves through a documented multi-step process, each stage with its own timing, before it reaches a clearing price.
PJM's annual load forecast has two distinct layers. The base layer is top-down and econometric — historical consumption patterns, macroeconomic indicators, electrification trends. It is standard utility modeling, and it is currently projecting mild contraction. Everything riding above it is a second layer: large load adjustments, submitted each July by PJM's member utilities, reporting specific known data center and industrial additions not captured in the base forecast. As Modo Energy's analysis of PJM's forecasting methodology documents, this bottom-up layer now accounts for more than 100% of PJM's projected peak demand growth over the next five years. The entire incremental demand signal driving capacity prices is announcement-based, not modeled.
Why Electric Customer Agreements are softer than they sound
The instrument underlying those announcements is an Electric Customer Agreement — a contract between a developer and a load-serving utility establishing the right to electric service. An ECA is not a construction commitment. It imposes no meaningful financial penalty for cancellation or delay. It is closer to a reservation than an obligation, and developers have a rational incentive to file them for speculative projects because queue position has option value. Microsoft signed agreements, entered PJM's queue, and canceled up to 2 gigawatts of capacity reservations anyway. No significant penalty followed.
PJM applies vetting criteria — haircuts for non-firm projects, utilization rate assumptions, double-counting checks — but the firmness classification itself depends on whether an ECA exists, and ECAs are softer than the label implies. The January 2026 downward revision reflects PJM tightening that vetting after the fact, discounting projects without executable construction timelines. It is not a macro call. It is a belated judgment that the announced pipeline was less real than submitted.
Every actor in the chain has an incentive to shade upward
The incentive structure underlying this process compounds the problem. Developers over-commit because cancellation is cheap and queue position is valuable. Utilities report bullishly because load growth justifies rate base investment that earns a regulated return. PJM is institutionally more accountable for under-forecasting reliability shortfalls than for over-forecasting costs that get socialized. Every actor in the chain has an incentive to shade toward the upside — a dynamic examined in depth from the ratepayer and grid planning perspective in prior work on this topic. What this piece adds is the investor consequence: the generators and infrastructure funds financing new capacity against $329 per megawatt-day are the ones holding hard capital commitments when the soft commitments unwind.
The lag makes the exposure precise. The 2027/28 auction, held in December 2025, cleared on a forecast built from large load adjustment submissions made in July 2024, finalized in January 2025. From utility submission to delivery year start: roughly thirty months. PJM's January 2026 downward revision will not reach a clearing price until the June 2026 auction for delivery year 2028/29 — more than two years away. The lag that inflated these prices works exactly the same way on the way down.
AI Is Learning to Do More With Less Power
The large load adjustment submissions that drove capacity prices upward were built on assumptions about the relationship between AI output and electricity consumption. Those assumptions are under pressure from two directions simultaneously.
Total energy per unit of AI output is falling. Liquid cooling is cutting power usage effectiveness — the ratio of total facility energy to compute energy — from the traditional range of 1.5 to 1.8 down toward 1.05. Inference efficiency is improving as models are compressed and optimized for deployment rather than training. The DeepSeek demonstration in January 2025, which delivered frontier model performance at a fraction of assumed compute cost, was one illustration of a broader pattern: the compute required to produce a given AI capability level is declining, and the forecasts embedded in interconnection queue requests did not assume that rate of decline.
Peak demand is softer than the queue numbers imply
Peak demand — the specific variable that drives capacity prices — faces additional pressure from workload flexibility. Companies including Emerald AI, working with NVIDIA and piloting with major utilities including NextEra and Constellation, have demonstrated in peer-reviewed EPRI tests that AI compute clusters can reduce power draw during grid stress events without degrading performance on mission-critical workloads. The LBNL uncertainty range for data center electricity consumption in 2028 already spans from 325 to 580 terawatt-hours — a band wider than Florida's entire annual consumption. The efficiency trajectory suggests the market has been pricing the top of that range.
The Political Backlash Is Now Being Written Into Law
The political and regulatory environment surrounding data center development has shifted materially over the past eighteen months, and it is shifting from three directions at once.
The state legislative wave is the most visible. More than 300 data center bills were filed in 30-plus states in the first six weeks of 2026. Moratorium legislation is advancing in at least 12 states. Ohio, Georgia, Texas, California, and Illinois are each pursuing different mechanisms — minimum usage charges, cost-shift studies, mandatory grid contributions, ratepayer protection requirements — to make data centers bear more of the infrastructure costs their load creates. The political fuel is not hard to identify. Average residential electricity bills rose nearly 30% between 2021 and 2025. The Union of Concerned Scientists estimates that ratepayers in seven PJM states are already on the hook for $4.3 billion in infrastructure projects approved in 2024 solely to connect data centers.
A voluntary pledge is not a policy
The White House Ratepayer Protection Pledge, signed in March 2026 by Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI, is a voluntary commitment with no interconnection mechanism, no financing facility, no supply chain strategy, and no enforcement. It signals political direction without changing the underlying economics. As Jane Flegal's March 2026 analysis for the Searchlight Institute documents in detail, the primary barriers to building a modern grid require durable policy solutions that vague pledges cannot substitute for.
Community opposition is canceling projects at four times last year's rate
Community opposition is a separate phenomenon from legislation, and it is accelerating. At least 25 data center projects were canceled in 2025 due to local resistance alone, four times the number canceled for the same reason in 2024, according to the Ad Hoc Group. The friction has grown serious enough that hyperscalers and utilities are exploring a fundamentally new model to work around it: funding heat pumps, rooftop solar, and distributed batteries in nearby homes to unlock interconnection capacity and reduce opposition simultaneously. The Ad Hoc Group calls this "bring your own distributed capacity." Google has already piloted a version through Xcel Energy's Capacity*Connect program in Minnesota, funding distributed battery investments alongside a 1.9-gigawatt clean energy agreement. That an approach this structurally novel is receiving serious attention from regulators and major utilities is a reasonable measure of how much conventional friction has accumulated.
Each of these layers operates through a different mechanism and is unlikely to reverse quickly. State legislation, once enacted, does not disappear between rate cases. Community opposition, once organized, tends to persist and spread. Together they represent a category of demand headwind being written into law and local planning processes across the country.
The Equipment to Build This Future Doesn't Exist Yet
The supply-side constraints on executing the announced buildout are structural, and two of them deserve to be kept distinct.
The generation equipment bottleneck is the first. Large frame gas turbines — the machines that actually add grid-scale dispatchable capacity, distinct from the smaller aeroderivative units used for onsite or peaking applications — are manufactured by a small number of global suppliers at a combined rate of perhaps a few dozen units per year. Backlogs already stretch to 2028 and beyond. High-voltage transformers, 85% of which are imported, carry lead times of two to four years. These are manufacturing constraints, not permitting constraints. Cheaper financing and faster permitting cannot move them. A generator that wins a capacity contract — whether through the normal auction or a backstop procurement — still has to source equipment from the same constrained assembly lines.
The Hormuz cascade hits a different part of the stack
The Hormuz supply chain cascade is a separate and compounding problem for the semiconductor stack that data centers actually run on. As covered in a prior piece in this series, HBM and DRAM costs rose 80 to 90% in the first quarter of 2026. The helium shortage at Ras Laffan is structurally different from an LNG disruption: helium is co-produced with LNG at the facility itself, meaning a Strait reopening does not restore production at a damaged plant. The supply constraint runs through the wafer fabrication process for the chips that go into the servers, not through the power lines that connect them to the grid. Bloomberg reported earlier this month that of the 12 to 16 gigawatts of data center capacity planned for 2026, only 5 gigawatts is currently under construction, with a third to a half of planned projects likely to be canceled or delayed. The reasons span both categories described here.
This Bet Requires All Three to Go Right
Capital financed at $329 per megawatt-day is a simultaneous bet that efficiency gains stay slow, that political and regulatory friction stays manageable, and that the physical buildout executes on announced timelines. Monitoring Analytics, PJM's independent market monitor, put the demand-side exposure directly in its January 2026 report: data centers accounted for 40% of the December 2025 auction's $16.4 billion cost, and $6.2 billion of that — nearly all of it — is attributable to data centers that have not yet been built. Across PJM's last three auctions, data center forecasts above existing load have driven 45% of the $47.2 billion in total capacity costs. S&P Global's analysis of the revised variable resource requirement curve suggests prices could decline roughly $38 per megawatt-day by 2028 under a revised demand picture — and that estimate does not account for the efficiency and friction effects described here.
PJM's backstop proposal confirms the diagnosis but not the cure
On April 10, PJM proposed a one-time backstop procurement of 14.9 gigawatts of new resources, citing a projected decade-long capacity gap of 50 to 60 gigawatts. It is worth being precise about what that proposal does and does not change. The 50 to 60 gigawatt shortfall is itself derived from the same announcement-based load forecast this piece has examined — built on ECAs that carry no meaningful cancellation penalty and large load adjustment submissions from utilities with their own incentive to report bullishly. The generators asked to fill that gap through a backstop contract would be making hard capital commitments against a demand signal built from instruments that carry no equivalent obligation. And those generators, once contracted, would still face the same large frame turbine backlogs, the same transformer lead times, and the same semiconductor supply constraints as every other project in the queue. A 15-year contract with PJM as counterparty does not move an assembly line. The backstop proposal is PJM confirming the diagnosis. It is not a cure.
None of the three headwinds described in this piece needs to be decisive on its own. They need only be sufficient, in combination, to shift the load forecasts that feed the next auction's curve. That process appears to be underway.
What Goes Up on a Lag Comes Down on One Too
What PJM’s recent auctions have revealed is not simply a tight market, but a system pricing a narrative faster than it can verify it. The elevenfold increase in capacity prices rests on a chain of assumptions—about demand permanence, execution certainty, and infrastructure readiness—that have not yet been tested against reality. Each link in that chain is now under quiet strain. Efficiency is improving. Friction is accumulating. Supply is constrained. And yet the price signal, by design, still reflects a version of the future formed two years ago.
This is the central asymmetry. Capital is being committed today against yesterday’s expectations, while the conditions shaping tomorrow’s demand are already shifting. The lag that amplified the upside has not disappeared; it has simply delayed the moment of recognition. When that recognition arrives, it will not announce itself as a single event, but as a gradual erosion of the assumptions embedded in the curve.
A few caveats are necessary. The efficiency trajectory could slow. Political and regulatory friction could ease as frameworks mature and reliability pressures reassert themselves. Markets have a way of clearing at unexpected price levels when the stakes are high enough. And the lag itself is not guaranteed to unwind symmetrically—what inflated prices over two auction cycles may correct over more than two. The point is not that the bullish scenario is wrong. It is that capital is being priced as though it is the only scenario, and that carries unacknowledged risk.
The structural lag that allowed AI sentiment to drive capacity prices upward can operate in reverse. Demand revisions, efficiency gains, regulatory and community friction, and supply-side ceilings are already in motion. Capacity prices will not reflect those forces until the lag runs its course. Investors financing generation at today’s clearing levels are not just buying megawatts. They are buying a specific version of the future—and that future requires multiple conditions to hold, simultaneously and on schedule.
Markets do not correct because they are wrong; they correct because reality catches up. The question now is not whether AI will drive enormous electricity demand—it will—but whether the timing, scale, and shape of that demand align with the commitments being financed today. That distinction is where risk lives. And in a market where soft commitments drive hard capital decisions, the most dangerous position is not being wrong. It is being early, leveraged, and certain at the same time.
