PJM's capacity price went from $28.92/MW-day in 2024/25 to $329.17/MW-day in 2026/27, with data centres assessed as responsible for 63% of the increase in the 2025/26 auction. That converts a national argument about AI into a specific dollar figure on a household bill across 13 states and DC, which is why opposition polls bipartisan and keeps rising. The decisive fights are about large-load tariff design and interconnection rules, not about whether AI is good.
In the PJM region — 13 states plus the District of Columbia, roughly 67 million people — the price the grid pays to guarantee generating capacity went from $28.92 per megawatt-day in the 2024/25 delivery year to $329.17 per megawatt-day in 2026/27. That is more than a tenfold increase, and it flows to residential bills. Analysis by the Institute for Energy Economics and Financial Analysis assessed data centres as responsible for 63% of the increase in the 2025/26 auction, worth roughly $9.3bn recovered from customers in higher rates.
That is the whole story, and it explains something the dominant coverage frame does not. Opposition to data centres is usually written as AI backlash — anxiety about jobs, unease about the technology, ordinary NIMBYism. If that were the mechanism, the politics would sort roughly the way other technology politics sort. They do not. They sort the way utility bills sort, which is to say not at all.
Polling across multiple independent instruments now shows majority opposition and, more strikingly, a fast-moving trend. The Annenberg Public Policy Center at Penn surveyed 1,320 US adults between 16 June and 19 July 2026 and found 61% opposed to construction of new data centres in their area — up from 49% in March. A Gallup poll found seven in ten opposed to an AI data centre in their local area, including 48% strongly opposed. A Heatmap Pro survey conducted by Embold Research from 8 to 13 August put opposition at roughly three-quarters, with more than six in ten strongly opposed.
The cross-party breakdown is the part that should stop you. In the Annenberg data, majorities of Democrats (69%), Republicans (54%) and independents (53%) all oppose. Opposition runs highest among adults under 30, at 70%, and falls to 57% among those 65 and older — the opposite of the age gradient you would expect from a story about technological unfamiliarity.
It has also crossed into campaign spending, which is the clearest available signal that a story has left the trade press. The National Republican Senatorial Committee has internally described data centres as a sleeper issue for the entire midterm cycle. NPR covered it as a live primary-season concern. CNBC ran two separate pieces on the political dimension inside ten days.
Most people have never heard of a capacity market, so here is the mechanism without jargon.
PJM does not just buy electricity. It separately buys a promise — a commitment from generators to be available to produce power on the highest-demand day three years from now. That promise is procured in an annual auction. Generators bid, PJM buys enough to meet its forecast peak plus a reserve margin, and everyone who clears is paid the same clearing price. The cost is then recovered from load-serving utilities, who pass it into retail rates.
Two features make this arrangement explosive when a large new class of customer appears. The first is that the auction procures against a forecast, not against actual consumption — so projected data-centre load raises the quantity PJM must buy years before a single server is racked. The second is that it is a uniform-price auction: everyone clears at the marginal bid. Push the demand curve right along a steep supply curve and the price does not rise proportionally, it rises violently. Hence 28.92 to 329.17.
The 2026/27 auction cleared at $329.17/MW-day in every zone — and would have cleared higher still had PJM not applied an administrative price cap. The observed number is a censored one.
The scale of the forecast revision is worth stating concretely. PJM's 2022 load forecast projected about 5,700 MW of growth by 2037 in the Dominion Zone, which contains northern Virginia's data centre alley. Its 2025 forecast projected more than 20,000 MW of growth from data centres alone in that zone over the same horizon. Three years of forecasting changed the input to a multibillion-dollar procurement.
Now the crucial part: the cost is recovered from load broadly. A household in a PJM state pays a share of the capacity charge that reflects its contribution to peak demand — not a share that reflects who caused the capacity to be procured. That is the entire dispute, stated in one sentence. The capacity was bought for a new class of very large customer; the bill is allocated across all customers.
Once you see it as cost allocation, the political shape stops being puzzling. A rate increase is not an opinion about artificial intelligence. It arrives monthly, it is the same size for a Republican household and a Democratic one, and it has an identifiable beneficiary who is not the person paying it. That is the oldest and most reliable political grievance there is, and it has nothing to do with whether large language models are good.
It also explains the durability. Technology backlashes usually decay as the technology becomes familiar. A cost allocation does not decay; it recurs, and it compounds as more capacity is procured against higher forecasts. One projection circulated in coverage of PJM costs, from the NRDC, puts the average family's monthly bill up by roughly $70 by 2028. That is a projection rather than an outcome, and should be read as such — but the direction is not seriously disputed even by people who dispute the attribution.
The 63% attribution figure is contested, and burying that would be dishonest.
SemiAnalysis has argued that the causal weight assigned to data centres in popular framing is too high, and that the increase reflects a supply-side story at least as much as a demand-side one: generation retirements outpacing new interconnection, transmission underbuild, and a queue process that has taken years to clear new resources. The Rutgers State Policy Lab reached a broadly similar conclusion, arguing data centres are not yet the primary driver of most consumers' bill increases. PolitiFact examined a version of this claim in June and found the causal picture more mixed than campaign rhetoric implies.
The steelman is straightforward: in a uniform-price auction with a steep supply curve, any demand increment at the margin appears to be responsible for the entire price move, and so does any supply decrement. Attribution in that setting is a modelling choice, not a measurement. If 20 GW of generation had not retired, the same data-centre load would have cleared at a much lower price — and you could just as defensibly write the headline about retirements.
What is not contested: capacity prices rose roughly tenfold, the cost is recovered from ratepayers broadly, and the load growth driving PJM's forecasts is overwhelmingly data centres. Those three facts are sufficient for the political dynamic regardless of how the attribution argument resolves.
If the reframing is right, then most of what is being fought over publicly is not where the outcome gets determined. Ballot measures and campaign advertising express the grievance; they do not allocate the cost. Three things do.
Each of those is a technical proceeding at a public utility commission or a regional transmission organisation. None of them will be decided in November. But they determine whether the next capacity auction's cost lands on households or on the customers whose demand summoned it — which is what everybody is actually arguing about.
The useful prediction from all this: expect the politics to keep intensifying even if AI sentiment improves, and expect the effective resolution to come from tariff dockets that almost nobody covers. A grievance about a bill is settled by changing who pays the bill.
Partly, and the size of that part is genuinely disputed. IEEFA assessed data centres as responsible for 63% of the capacity price increase in PJM's 2025/26 auction, worth about $9.3bn recovered from customers. But SemiAnalysis and the Rutgers State Policy Lab both argue the attribution is weaker than headlines suggest, because generation retirements and transmission underbuild contributed to the same price move. In a uniform-price auction any marginal change can be made to look responsible for the whole increase. What is not disputed: capacity prices rose more than tenfold, and data-centre load is what is driving PJM's demand forecasts.
DrafterDaily covers the infrastructure and rules that decide where AI's costs actually land.
Read more Technology analysisCVE-2026-0768 is an unauthenticated root RCE in Langflow. It was disclosed in January, the fix has shipped through seven releases, and attackers are hitting it in September — because low-code AI middleware became critical infrastructure without acquiring a patch owner.
OpenAI's first published benchmarks for its custom inference chip are real, from a public benchmark, and normalised on nameplate TDP rather than measured power. OpenAI disclosed that itself, in a sentence nobody is quoting.
IDC has cut its 2026 PC forecast three times in nine months — from −2.4% to −8.9% to −11.3%. Thirty-two million fewer machines, and the market is worth more money. The revision trail is the story.