The AI Power-Bill Map Is Really a Market-Design Map

A map showing where AI data centers collide with rising electricity bills made the Hacker News front page this week, and the reaction was predictable: the machines are eating our power bills. The map is real. The bills are real. But if you trace the money instead of the vibes, the geography tells a more specific story — one about a single grid operator's market design, and one that developers building on AI infrastructure should read as a forecast of where their compute will and won't get built.

Start with the numbers that survive fact-checking. Bloomberg's analysis of nodal wholesale prices, built on data from Grid Status, found spots where wholesale electricity cost as much as 267% more in April 2025 than five years earlier — and more than 70% of the nodes with price increases sat within 50 miles of significant data center activity. Nationally, data centers consumed 4.4% of U.S. electricity in 2023, and the Department of Energy's Berkeley Lab report projects 6.7% to 12% by 2028. That's a demand shock the U.S. grid hasn't seen since air conditioning.

The receipts are in PJM's capacity auctions

The clearest paper trail runs through PJM, the grid operator covering 13 states from Virginia to Illinois — including Northern Virginia, the largest data center concentration on Earth and the physical home of us-east-1. PJM runs annual capacity auctions that pay generators to exist, and those auctions have gone vertical. The 2025/26 auction cleared roughly ten times higher than the year before. July 2025's auction hit a record $329/MW-day. This July's auction, for the 2028/29 delivery year, slammed into the FERC-approved price cap of $325/MW-day — and would have cleared at $555 without it, $777 in the Chicago zone.

Monitoring Analytics, PJM's independent market monitor, has done the attribution work: data centers drove $9.3 billion of the 2025/26 cost increase, $6.3 billion of this year's $16.4 billion auction, and $29.4 billion — 46% — of the $63.6 billion in total capacity charges across the last four auctions. Those charges land on every ratepayer in the territory. Baltimore households saw bills jump more than $17 a month after the 2025 auction alone.

So yes: if you live in PJM territory, AI is on your power bill, with receipts.

Same GPUs, different market, different bill

Here's what the "AI broke your power bill" framing skips: Texas absorbed a comparable data center buildout and its prices stayed comparatively boring. SemiAnalysis made this argument at length — ERCOT runs an energy-only market with no capacity auction, and its forward prices rose modestly while PJM's capacity prices went up roughly 9x. Same AI boom, radically different consumer outcome.

That's because PJM's capacity auction converts demand forecasts into prices years ahead of delivery, and those forecasts are stuffed with speculative load — data center projects that request interconnection in five states and build in one. PJM has repeatedly revised its own data center forecasts downward, but the auction prices in the phantom gigawatts anyway. Meanwhile supply can't answer the price signal: this year's auction attracted about 525 MW of new resources against a 6.8 GW shortfall, because the interconnection queue takes years to traverse. A market that prices in imaginary demand while blocking real supply will produce spectacular numbers regardless of what the demand is for.

My read: both stories are true, and the second one matters more. The demand shock is genuine, but the size of the consumer bill is a policy variable, not a law of physics. Which is exactly why the policy is now changing fast.

The regulatory bill arrives

The political response has outrun most people's mental model. Over 300 data-center bills landed in 30 state legislatures in the first six weeks of 2026. By May, 23 states had approved at least one special "large load" tariff. Oregon's POWER Act forces facilities over 20 MW into their own rate class with mandatory 10-year power purchase agreements. Ohio regulators approved an AEP tariff making big data centers pay for 85% of their subscribed capacity whether they use it or not. Texas now requires new loads over 75 MW to be remotely curtailable during grid emergencies. PJM itself is floating a backstop capacity auction and a "connect and manage" fast lane for large loads, and its market monitor wants data centers pushed into separate procurement entirely — bring your own generation, or sign 15-year capacity deals that wall off residential ratepayers.

The era of data centers as passive grid customers is over. The new deal is: pay for what you reserve, generate your own, or be interruptible.

What this means if you build on this stuff

If you're a developer, this reshapes your infrastructure in three concrete ways.

First, region economics are diverging, and capacity will follow the friendly markets. Take-or-pay tariffs and PPA mandates change hyperscaler buildout math state by state, which is why so much new construction is heading to ERCOT territory, the Gulf, and the Midwest rather than piling deeper into Virginia. Expect new cloud regions — and new GPU capacity — to show up in places your latency budget hasn't considered yet. If your architecture assumes us-east-1 grows forever, it's aging poorly.

Second, curtailable compute is becoming a first-class primitive. A training job that checkpoints cleanly and yields for a four-hour grid event is worth real money to an operator negotiating interconnection — Texas has made it law, and PJM's proposals point the same way. Carbon-aware scheduling was the warm-up; demand-response compute is the actual product. If your batch workloads can't tolerate preemption, that's now a cost problem, not just a resilience one.

Third — and this is the part worth sitting with — developers currently pay for AI twice. Once per token on the API bill, where efficiency gains keep pushing prices down, and once per kilowatt-hour at home, where the infrastructure surcharge has been quietly socialized onto everyone in PJM's footprint. The 23-state tariff wave is an attempt to move that cost back onto the industry, and if it succeeds, some of it will resurface in compute prices. That's the correct outcome. Subsidized infrastructure produces distorted architecture decisions, and the current per-token price of inference is mildly fictional in the regions where ratepayers are covering the buildout.

The map, in other words, isn't really a map of data centers. It's a map of which markets pass AI's infrastructure costs to households and which make the industry carry its own weight. Over the next two years, that second category is going to get a lot bigger — and where your GPUs live will depend on it.