Somewhere on a spreadsheet inside PJM Interconnection — the grid operator that keeps the lights on across thirteen states, from Chicago to the DC suburbs — sits a number that unsettles me more than any AI doomsday headline: 1,066 gigawatts. That’s how much electricity data center developers have formally requested from utilities nationwide. It’s also, in my view, the most misleading statistic in the entire AI-infrastructure conversation, because most of it is phantom data center power demand — capacity that gets filed, forecast, and priced into your electric bill, and then quietly never gets built.
Here’s the part that should actually bother you: you are already paying for it. Grid operators and utilities are on pace to genuinely commit to serving only about 298 GW of that 1,066 GW mountain of requests, according to Wood Mackenzie research reported by Bloomberg. That leaves roughly 72% — some 768 GW — as phantom load that will statistically never draw a single watt. Capacity auctions don’t wait around for concrete to get poured, though, so your bill is climbing today to underwrite data centers that may exist only on paper.
The Phantom Data Center Power Demand Problem, By the Numbers
Start with a term nobody explains at parties: an interconnection queue is simply the waiting line a utility keeps of every project — solar farm, factory, data center — that wants to plug into the transmission grid. Filing into that queue costs a developer almost nothing: some paperwork, a modest deposit, maybe a feasibility-study fee. There’s no requirement that you have financing, permits, or a signed power contract before you get in line. That low bar is precisely why the topline number is so distorted.
Wood Mackenzie’s data center forecast, cited in Bloomberg reporting from August 12, 2026, found utilities are on pace to actually commit to that 298 GW figure out of the full 1,066 GW requested. Subtract one from the other and you get 768 GW of demand — larger than the entire generating capacity of Japan — that analysts expect to quietly evaporate. It’s the electricity-market equivalent of everyone in town calling every restaurant to hold a table for the same Saturday-night party, then no-showing at four out of five.
Source: Wood Mackenzie data, reported by Bloomberg, Aug. 12, 2026
How Interconnection Queues Get Gamed
None of this is fraud, and I want to be clear about that. It’s exactly how you’d behave if you were racing to lock down a $10 billion campus before a competitor grabbed the same substation capacity. Developers routinely file interconnection applications with several different utilities, at multiple candidate sites, for the exact same planned facility. Once one site “pencils out” — financing closes, permits clear, a power contract gets signed — the rest get quietly abandoned or simply never activated.
I spent years around CERN’s grid-computing allocations, where every physics collaboration requests more compute time than it will ever use, because the penalty for asking too little dwarfs the penalty for asking too much. Utilities are now living that same incentive problem, just with turbines instead of GPUs. A real project like the OpenAI-Nvidia Ohio buildout can coexist in the queue data with five phantom copies of itself filed elsewhere, and grid operators have no reliable way to tell which is which — some developers are even chasing offshore, floating sites just to dodge congested queues entirely.
What makes planners take even bloated numbers seriously is that real load really has exploded. U.S. data center grid demand is running near 42 GW today, up from 23 GW in 2023 — genuine, metered growth, not a filing. Grid operator load forecasting teams look at that trajectory and reasonably assume some of the queue’s mountain must be real. The trouble is separating the signal from 768 GW of noise, and right now nobody has a working filter.
Inside PJM’s Capacity Auction: Pricing a Mirage
A capacity auction is basically an insurance market: power plant owners get paid years in advance to guarantee electricity will be available on the grid when needed, whether or not it’s actually called on. PJM runs the largest one in the country, and it uses long-term load forecasts — the same forecasts fattened by duplicate data center filings — to decide how much capacity utilities must buy and at what clearing price.
The results are not subtle. PJM’s forward capacity auction cleared at $2.2 billion for the 2024-25 delivery year. For 2027-28, it cleared at $16.4 billion — roughly a 7.5x jump, substantially driven by data center load baked into the forecast. That cost doesn’t stay inside a utility boardroom; it flows straight through to the capacity charge on every residential bill in the PJM footprint.
Source: Bloomberg reporting on PJM Interconnection capacity auction results, cited in nuclear-news.net, Aug. 17, 2026
⚡ PHOTON’S TAKE
I’ve watched grid planners try to tell real megawatts from vaporware for years, and PJM’s capacity auction just proved you can price a mirage. Developers file the same gigawatt-scale project with multiple utilities at multiple sites because it’s free to do and expensive not to, and nobody polices it. The result isn’t an AI power shortage — it’s a forecasting failure that ratepayers are underwriting today. Fix the queue rules, require financial skin in the game before a request counts, and half this crisis disappears overnight. Until then, your bill is subsidizing spreadsheets, not servers.
Why Phantom Data Center Power Demand Hits Your Bill Anyway
Utilities and grid operators, as Power Magazine has argued, aren’t wrong to plan for growth — they’re wrong to treat every unverified request as equally credible. Some states are pushing back: Texas regulators recently ordered a sweeping audit of data center power requests specifically to separate financed projects from speculative ones. That’s the right instinct, and every PJM state should be running the same exercise before the next auction clears.
Rising AI electricity prices are real, but the honest story isn’t “the grid can’t keep up with AI.” It’s that a market built for a world of slow, verified requests is now being flooded with cheap, disposable ones, and the auction mechanics can’t yet tell the difference. Until interconnection queues charge developers something closer to what a reservation is actually worth — real deposits, real deadlines, real penalties for walking away — this pattern will repeat with every fresh wave of hyperscale announcements.
My prediction: within eighteen months, at least one major grid operator formally adopts a “show me the money” filing standard, requiring proof of site control and financing before a request counts toward long-term forecasts. Until that happens, phantom data center power demand will keep clearing PJM’s capacity auction as if it were real, and households will keep fronting the bill for a boom that, statistically, is mostly a rounding error waiting to happen.
