When a single class of electricity customer starts behaving like a city – ordering power equipment years in advance, locking up turbine supply for the rest of the decade, and driving capacity-market charges measured in the billions – the question of who pays for the grid to serve that customer stops being abstract. It becomes the central fact of the U.S. power economy.

The numbers behind the spiral

The most direct measure of what AI demand is doing to construction costs comes from Lazard’s annual electricity-generation cost report. Large-scale solar projects now cost an average of $69 per megawatt hour, up from $58 a year ago – a 19% increase. Adding battery storage pushed costs to roughly $109, up from $91. Combined-cycle natural-gas plants, the workhorses of baseload generation, climbed to roughly $90 per megawatt hour, up from $79 – the highest level in about fifteen years. At the household level, electricity prices rose 4% in June from a year earlier, according to the Bureau of Labor Statistics, outpacing overall inflation.

These are not small moves. Demand growth is “dominating everything in energy,” as George Bilicic, Lazard’s vice chairman of investment banking and global head of power, energy and infrastructure, put it – pushing costs up across exploration-and-production companies, pipeline operators, utilities, and infrastructure-services firms.

The downstream picture is equally stark. Fifty-one investor-owned utilities have now committed to an estimated $1.4 trillion in capital spending over the next five years, up more than 20% from last year’s projections, according to PowerLines, a consumer-education group. On the equipment side, GE Vernova chief executive Scott Strazik said many customers are ordering gas turbines for delivery into the 2030s, and by the end of this year, half of the company’s 2031 turbine output will be under contract. “This is just a larger market for a longer period of time,” Strazik said. The backlog is not a temporary inventory squeeze. The heavy-equipment pipeline is booked out five years, and the company is boosting output to match.

Roughly 90% of new generation projects likely to start producing electricity this year are renewable energy and battery storage, according to the Energy Information Administration. But that composition has not contained costs – solar, gas, and solar-plus-storage are all climbing.

The price signal that broke

A functioning market uses price to signal producers to build. In the PJM Interconnection – the grid operator serving roughly 67 million people across the mid-Atlantic and Midwest, with data-center growth concentrated in Northern Virginia – that mechanism produced a price that satisfied no one.

PJM’s latest capacity auction fell short of the reserves it hoped to procure. The operator plans another auction for additional capacity this fall. But the mechanism has a structural problem: the auction hit a price cap, and capacity prices are not high enough to entice developers to build the additional power plants the market needs, particularly natural gas. “The auction has pulled off making everyone unhappy,” said Peter Gardett, chief executive of market-data platform Noreva.

The arithmetic explains why the price cap matters. Of approximately $16.4 billion in capacity-market charges, $6.3 billion was caused by data-center demand, said Joe Bowring, president of Monitoring Analytics, the market’s independent monitor. That is 38% of total charges from a single concentrated demand source. In a market where that demand surge should push the price high enough to bring new supply online, a hard price cap prevents the signal from clearing. The result is a supply shortage that the market mechanism, by design, cannot correct. Bowring is calling for new data-center load to be removed from the capacity market and procured separately – a structural change that would let capacity prices rise to a level where developers actually build.

The equilibrium: blended cost recovery

The higher construction costs will eventually work their way into monthly utility bills. That is how the regulated model works: utilities recover capital investments over time through rates approved by state regulators. The question is how those costs are allocated across customer classes.

A structural analysis of the dynamics at work in the U.S. power system helps explain why the answer is likely to be: spread across everyone. Four main players populate the game. AI data centers want guaranteed, scalable megawatt capacity at the lowest marginal cost – and their forward ordering of GE Vernova turbines reveals willingness to pay for reliability regardless of fuel source, despite stated clean-energy preferences. The 51 investor-owned utilities, operating under regulated-return frameworks, maximize profit by expanding ratebase – every dollar of approved capital investment generates a guaranteed return. State regulators must balance utility returns against consumer rate shock. Households want minimal monthly bills but have no direct seat at the allocation table.

In this sequential game, data centers locate first (heavily concentrated in Virginia), utilities invest to meet demand, and regulators then choose how to recover costs. The regulator’s decision is constrained by administrative feasibility: more than 190 separate utility filings wrestling with contract provisions, special rates, or charging mechanisms for large data-center customers are currently being contested across the country, according to utility filings aggregated by AI startup Halcyon. Untangling which grid upgrades serve only data centers versus the broader system is structurally difficult, as many upgrades benefit all users and are traditionally applied to all customer bills. Given that constraint, the regulator’s feasible choice set collapses to a single branch.

The subgame-perfect equilibrium, derived by backward induction from the players’ payoffs and the administrative-cost constraint, is blended cost recovery across all ratepayers. The outcome is stable – no single player can unilaterally improve its position – but Pareto-inferior to a coordinated investment with explicit cost allocation. The PJM auction outcome provides empirical confirmation: the price cap held, the reserve target was missed, and the market produced no new gas generation.

Structural note: the “regulator” player in this analysis, for analytical tractability, represents multiple distinct institutions – state rate-case commissions, the FERC/PJM auction mechanism, and Monitoring Analytics – each with separate authority and incentive structures. The analysis treats them as a unified agent; the actual institutional fragmentation is itself a structural constraint on the equilibrium.

For a regular household watching its electricity bill climb 4% year over year – faster than inflation – the practical meaning of that equilibrium is that grid infrastructure built to serve the AI boom is being socialized across all customers.

The feedback loops that lock it in

The strategic positions produce a blended-rate equilibrium; the systems dynamics show why that equilibrium is structurally self-reinforcing – the feedback loops that maintain it operate independently of any player’s strategic choice. Two reinforcing loops are currently dominant, and three balancing loops are either slow to respond or impaired.

Reinforcing loop R1 – the demand-capacity cost spiral. AI data-center load increases appear price-inelastic in the short term. This drives up utility capital plans and construction costs. Higher costs flow into rates, raising the breakeven for data-center operations, but demand continues growing, feeding further capital-plan increases. The loop is reinforcing and currently dominant, with delays of two to four years before throttling mechanisms fully engage.

Reinforcing loop R2 – cost and rate pass-through. Construction cost escalation triggers utility rate cases, which raise electricity bills for data centers and all customers. Costs passed through to cloud and AI consumers remain price-inelastic in the short term, reinforcing further cost increases. This loop amplifies the initial cost shock through the regulated rate structure.

Balancing loop B1 – the equipment bottleneck throttle. Rising turbine orders swell GE Vernova’s backlog, extending delivery lead times. Half of 2031 turbine output is already under contract, CEO Scott Strazik said, orders reaching into the 2030s. The delay between contract and turbine delivery is four to six years. This throttles new capacity additions and eventually relieves price pressure, but only on a multi-year timeline.

Balancing loop B2 – the permitting and capacity-market throttle, impaired. Demand growth triggers the PJM capacity auction, but the price cap prevents the price from reaching a level that would incentivize new gas construction. The cap leaves supply tight and data-center demand unmet by new physical generation. This loop would normally balance supply and demand through price, but the price-cap impairment means it cannot clear the market. PJM fell short of its reserve targets and plans another auction this fall. The price cap creates a causal-chain deadweight loss: the willingness to pay for gas generation (by data centers) exceeds the construction-cost threshold (by builders), but the binding price cap prevents clearing – this is a structural uneconomic equilibrium, not merely a price-level distortion.

Balancing loop B3 – political cap intervention, weak. Rising retail rates generate political pressure, leading to voluntary pledges. But the loop is weak: tracking costs specifically caused by data centers is difficult, and the 190-plus pending filings indicate the cost-allocation question remains unresolved. The loop may not materialize meaningfully.

The system exhibits a Limits to Growth archetype: the reinforcing growth engine (R1, R2) meets balancing constraints (B1, B2) acting on the same stock – grid generation capacity. The limit is not demand exhaustion but supply-side bottlenecks and a regulatory price cap that prevents the capacity-price signal from reaching developers.

A Shifting the Burden pattern is also present: the voluntary pledge (B3) targets ratepayer perception rather than the underlying supply shortage, and the structural cost-allocation question – who pays for grid infrastructure serving a single customer class – remains unresolved across the 190-plus filings.

The political placebo

The White House has persuaded tech companies and a group of electric utilities to pledge to avoid shifting the cost of new grid equipment needed by data centers onto regular households and businesses. That pledge, as the source coverage itself notes, is voluntary and unenforceable – “it is not clear how voluntary pledges will work in practice” and “tracking costs specifically caused by data centers will be difficult.” The pledge functions as cheap talk: state regulators, not the White House, control final rates, and the administrative burden of isolating data-center costs makes the pledge structurally unenforceable.

The Shifting the Burden archetype describes what happens next. The pledge targets the visible concern: public worry about cost-shifting. It does not alter the structural system: the absence of any binding cost-allocation mechanism. Over time, it may dampen the political pressure that generated it. But the more than 190 open utility filings confirm that the cost-allocation question remains unresolved. The symptomatic fix works temporarily; the underlying condition worsens.

The Virginia concentration

Data-center growth in the PJM region is concentrated in Northern Virginia, creating a local constraint that the regional capacity-market design cannot efficiently address. Of roughly $16.4 billion in capacity-market charges, $6.3 billion was caused by data-center demand – 38% of total charges driven by a single concentrated load in a single sub-region. The concentration amplifies every bottleneck in the system: equipment booked out five years is being competed for by hyperscale projects in one corridor, permitting delays are most acute where demand is densest, and the gap between data-center growth and available grid capacity is widest precisely where the political pressure is sharpest.

What’s missing from the picture

The player inventory in both the game-theory read and the systems map excluded siting and environmental actors. Their opposition drives permitting delays that restrict supply-side expansion. Including them would shift the equilibrium toward slower build-outs and higher clearing prices – meaning the cost trap is likely deeper than the analysis captures.

The forward turbine-ordering behavior is credible revealed preference, but may reflect venture-funded land-grabbing rather than durable demand. If grid costs erode data-center unit economics, operators may curtail or relocate, potentially altering the reinforcing R1 loop.

Open questions

  • Can state regulators devise a workable mechanism for isolating data-center grid costs before blended recovery across all ratepayers becomes the entrenched default?
  • Will PJM’s planned additional capacity auction this fall clear at a level that actually brings new generation online, or will the price cap again prevent the market from doing its job?
  • Whether Monitoring Analytics’ proposal to remove data-center load from the PJM capacity market and procure it separately alters the auction equilibrium before the next PJM capacity auction, not whether “structural levers are pulled” in abstract.
  • Will the equipment bottleneck – half of GE Vernova’s 2031 output already contracted – ease as the decade progresses, or does the forward-ordering behavior signal a structural shortage that no amount of regulatory reform can fix in the near term?

The cost figures – Lazard’s 10%+ construction-cost increases, the $1.4 trillion utility capital plan, the 4% year-over-year electricity price rise – are not disputed. The analytical question is who will absorb them. The current equilibrium distributes them across all ratepayers. Whether that equilibrium holds depends on whether the structural levers that could alter it are pulled before the cost escalation becomes politically untenable.

Analytical techniques used in this piece

This analysis applies the methods below. Each links to a short, plain-English explainer you can read and reuse.

Root-Cause Analysis
Traces a symptom back along its causal chain to the conditions that actually generated it.
Strategic Interaction (Game Theory)
Models a situation as a game — players, moves, payoffs, and likely equilibria.
Systems Dynamics (Structural)
Maps a system’s structure — stocks, flows, and the architecture that shapes its behavior.