Data Center Dreams, Grid Reality: A Tale of Missed Connections
The Wild World of Proposals
Ever noticed how the U.S. power grid feels like a crowded wedding reception with an endless line of guests who never actually show up? That’s exactly what Brian Martucci of UtilityDive is describing: a flood of data center plans that end up as paper‑backs instead of concrete reality.
Why It’s a Mess
When utility folks and grid operators try to map out their future, a maze of “interconnection requests” turns into an overgrown tangle. Astrid Atkinson—former Google tech lead turned CEO of Camus Energy—puts it bluntly: “You’re seeing five to ten times more requests than data centers that ever get built.” That’s a recipe for planning headaches.
Forecasts—A Wild Card Game
Even short‑term projections for how fast data center loads might grow feel like a game of “Where’s Waldo?”
- RAND Corporation put a “upper confidence” price tag on 347 GW of AI‑sector power consumption by 2030.
- Conversely, Schneider Electric slapped that figure with a “blowing‑up butterfly” label in their latest whitepaper.
- Other solid voices? They’re clamping it under 100 GW—a more grounded view, if you asked.
What It Means for Us
Picture this: you’ve got a giant buffet (the grid) that needs to feed millions (power demand), but half your guests never arrive. Utilities are having to stay on their toes, juggle too many “what if” questions, and shuffle resources like a circus act.
Bottom Line
It’s a reality check: proposals without build‑out mean more paperwork, less certainty, and a grid that’s left guessing where the next surge will pop up. For a smooth ride, we need better data, smarter projections, and a dash of patience.
When AI Grows, Power Grids Get a Reality Check
Schneider Electric ran a fun little experiment, dreaming up how much juice the next decade of AI might need. The numbers rolled in from a modest 16.5 GW to a whopping 65.3 GW, with the sweet spot landing at 33.8 GW. That’s their “golden mean” – enough energy to keep the brain‑y machines humming without crashing the grid.
Why the Wild Ride Matters
These wildly divergent forecasts look a little like a toddler’s drawing of a “future city” – full of big, bright ideas but missing the practical details. For the folks running the power plants, that means a real headache: phantom data centres. Bianca Giacobone from Latitude Media coined that term in March, and it’s hit home for utilities, grid operators, and regulators alike.
What the Experts Say
- Atkinson. He tells power‑supply teams to treat big‑eye predictions with a dash of skepticism. Take Exelon’s forecast, for instance: a tidy 11 GW of “high‑probability” data‑centre load over a decade. Sounds tidy, but the real world tends to be messier.
- In 2018, the Lawrence Berkeley National Laboratory threw a curveball, comparing mid‑2000s predictions with actual growth across 12 Western U.S. utilities. Spoiler: many overestimated the demand.
Why It’s Hard for the Grid Crew
Utilities can’t just read a crystal ball. They’re stuck trying to tease apart which interconnection requests will actually make it to the grid and which are just sugar‑plated hopes. The fallout? Oversized requests chew up the limited study time and push other projects down the line, messing with long‑term planning and making overbuilding a pricey nightmare.
Getting Their Act Together
To steer clear of these pitfalls, utilities have rolled out a few clever tricks:
- Standardised, large‑load interconnection frameworks that shave off guesswork.
- Demanding more upfront financial commitment from data‑centre developers – because if you’re going to hard‑wire a lot of power, you better be serious.
- Teaming up with state policymakers for a little extra backing or regulation.
Bottom line: as AI keeps shooting up, the power game is shifting from guesswork to a more disciplined, partnership‑driven approach. And if you’re a data‑centre developer, a little extra cash in your pocket might be the ticket to keep your electric dreams alive.
A problem of transparency
Phantom Power: Why Data Centers Keep Mysteriously Disappearing
Stealth Mode: The Transparency Gap
The so‑called phantom load issue isn’t just about missing wattage—it’s also a cloak‑and‑dagger problem. Developers and their agents hide land purchases and early‑stage plans in cryptic LLCs and blanket NDAs. Atkinson comments that while they trim projects ruthlessly, they stop short of making every announced plan a guaranteed ticket to the data‑center graveyard.
Examples striking the scene:
- Microsoft scrapped up to 2 GW of promised capacity since January.
- Tract pulled the plug on a 30‑building Phoenix proposal last year amid local pushback.
Over‑planning to Be Safe
Around the industry, even seasoned giants—Microsoft, Meta, Amazon, Google—submit a flood of proposals, often far exceeding probable demand. The reason? Uncertainty about power availability and the labyrinth of permits. Less experienced developers burn through an even larger percentage of their ideas, adding to the phantom count.
Grid Jitters and the Rise of “Behind‑the‑Meter” Power
Long waits for grid interconnection are nudging operators toward on‑site generation. Take Elon Musk’s Memphis xAI hub: the Grok model runs on 35 gas turbines right at the edge of the grid. A lawsuit filed in April by an environmental group surfaced that fact.
Chris Wright’s Liberty Energy could slot a 1 GW, off‑grid, gas‑fired plant into a planned Pittsburgh business park, feeding both data centers and other heavy industrial loads.
GE Vernova’s 21‑GW reservation pipeline is about a third owned by data‑center customers, according to CEO Scott Strazik in April.
Field‑Day in the Midwest
Great River Energy (GRE) in Minnesota has seen “more than a handful” of large‑load interconnection requests recently. Zachary Ruzycki, the director of resource planning, worries about draining staff time on projects that might not come to life.
Still, the promise of investment drives new generation plans. GRE will spend an $812 million federal grant to acquire roughly 1.3 GW of renewable power to feed future loads, CEO David Saggau said in January.
GRE isn’t alone; its state hosts investor‑owned utilities like Xcel Energy. Together, they’ve received at least 11 data‑center proposals since 2020—Amazon, Microsoft, Meta, and three 500‑MW schemes from Tract. Some may duplicate each other, but there’s no reliable way to tell which is which. Of those 11, only Meta’s has claimed it has begun construction as of earlier this year.
Industry‑Wide Juggling Act
“This is a challenge across the industry,” says Patricia Taylor, director of policy and research at the American Public Power Association. Data‑center developers survey both within their own communities and in neighboring ones, constantly adjusting to new information and shifting baselines.
When it’s cheaper, “You’ll buy queue positions all day long”
Data Centers and the Power Grid: A Tug‑of‑War
Long‑haul trains of servers are on the move, and the guys who keep the lights on are trying to keep the ride from crashing.
What the Power Research Group Found
In September 2024, EPRI asked 25 big utilities how they saw the future of data‑center demand. Nearly half of the respondents (48%) said they’d be responsible for at least 10% of peak grid use by 2030. A smaller slice, 26%, pitched that number at twice that share. But the tone of the answers was less optimistic than the numbers: most folks doubted that every project would actually hit the grid.
When the utilities that already had data‑center requests topping 50% of today’s peak load peered into the future, none saw the extra five‑year load surge past 35% of the peak. Even those that had the biggest current load shares were cautious.
How Utilities Tame the Beast
- Fast‑track‑ers (about 30%) treat the projected number at face value but expect it to crescendo gradually.
- Cautious doers (another 30%) strip the number down according to how mature the project looks—think land deals, permits, signed agreements, and whispered “sold‑out” rumors.
- By “dialling down” the numbers, they check whether the giant server farms are really the real estate heavyweights they claim to be.
Ruzycki from Great River Energy said, “We have to serve the grid, not the dream.” But he added, “If a data center has the land and the muscle to build, we’ll put the gears in motion to make sure we handle them.”
Great River also bills the folks who vet large requests for their time, so members don’t bear the cost. Meanwhile, APPA’s Taylor argues that utilities should grab deposits for interconnection studies, lock in service agreements that make data centers pay their share for upgrades, and guarantee a minimum load—because the world can’t be run on a ping‑pong of last‑minute decisions.
One‑Size‑Fits‑All? A Fairly Bold Idea
Allison Clements, a former FERC commissioner, and Peter Freed, ex‑Meta energy strategist, suggested in a Utility Dive op‑ed to put a standard process on the table. Think: interstate “relay” queues, anonymous status updates, and a system that forces developers to prove readiness, pay stage‑by‑stage fees, and drop dead projects that don’t make the cut.
But the horse still runs on a rails you never see.
Reality Check: The Phantom Load Problem
Karl Rábago from Rábago Energy points out that developers are savvy at tugging utilities to one another. “It’s cheaper to sit in a queue than to abandon it,” he says. That makes them buy seats in the queue all day long, hoping the grid will be generous enough to invite them once they’re ready.
State bills like Texas’s hot‑ticket are nice in theory, but Rábago argues they’re vague. He favors a “reverse auction” that hunts for the fewest sweeteners a data center will need to plug into the grid.
Virginia’s Leading the Charge
The state that’s hosting the largest data‑center market has begun testing the ideas. Dominion Energy, Appalachian Power, and Rappahannock Electric Cooperative have all mapped out new large‑load rate classes aimed at server farms. Dominion and Appalachian will make sure data centers pay at least 60% and 80% of contracted demand, keeping existing ratepayers out of the cold.
Rappahannock’s proposal is all about collateral and direct dealings with special subsidiaries to keep the grid calm. They’re pushing for data centers to put up collateral, pay upgrades, and might even cover up to 100% of contract load. That’s a lot of responsibility, but it could keep the whole system running without a punch‑drunk server chiming.
As we shuffle toward the next decade, the solution may lie in how we shape the rules—like giving everyone a fair shake after all, while forgiving the grid while a giant digital playground is under construction.
