Aggregator

More Than Two-Thirds Of The Power Sought For US Data Centers Will Never Materialize

Zero Rss
2 months ago
More Than Two-Thirds Of The Power Sought For US Data Centers Will Never Materialize

It was almost a year ago, in November 2025, when we first calculated that the AI supercycle was facing a huge problem: namely, a power shortfall of (at least) 44 nuclear power plants, or GW or energy, by 2028.

A Huge Problem For The AI Revolution: A Power Shortfall Of 44 Nuclear Power Plants By 2028 https://t.co/2ciiUudp8T

— zerohedge (@zerohedge) November 12, 2025

Fast forward to last week when we made another stunning discovery: the reason why Texas Gov Abbott froze the rollout of new data centers in his state is because the number of interconnection requests facing ERCOT had risen to a mindblowing 474 GW (of which 90% were data centers), when the state's highest ever power demand just hit 91GW. Needless to say, this will never happen in (maybe) a hundred years, but it does prompt the question we have been asking for years: where will the energy come from.

How mathematically impossible is the energy math behind the data center rollout? Just Texas alone is facing 474GW of interconnection requests (ERCOT), of which 90% is data centers. Which is why gov Abbott froze rollout of new data centers in Texas.

For context, ERCOT all-time… pic.twitter.com/sdl26FPmwL

— zerohedge (@zerohedge) August 8, 2026

Well, as it turns out, the answer is nowhere. According to new projections from Wood Mackenzie, zmore than two-thirds of the electricity sought for the artificial intelligence boom in the US isn't likely to materialize due to "phantom" projects and long-shot pitches.

The consultancy, which was inspired by not only our recent use of the term "bragawatts", but the Bernstein report it came from (link here for pro subs) which reached the same conclusion...

Think it's just about tokens? Think again: the next AI crisis is real compute vs "bragawatts"

"We certainly don’t have a crystal ball, but we have spent a lot of time in the data center industry and think we’re pretty good at sniffing out the real builds from the “bragawatts”.…

— zerohedge (@zerohedge) July 21, 2026

... said it sees US grid operators and utilities likely committing to about 28% of the 1,066 gigawatts requested for data center projects (as most know by now, a gigawatt is the equivalent generating capacity of a traditional nuclear reactor: enough energy to power about 750,000 households for a year, or about 10-30 data centers).

While certainly no one expects every proposed project to be built, Bloomberg notes that the data reveals how inflated projections of future power demand complicate planning and budgeting by utilities and grid operators. That can translate into higher utility bills, since capital costs for upgrading infrastructure are typically covered through rate hikes.

"Everyone's trying to figure out the rules of the road in order to make this the most efficient process without increasing grid prices," said Alex Klaessig, co-founder of energy market intelligence firm Halcyon.

Developers have been taking a shotgun approach to pitching projects to utilities, seeking to get ahead in the AI build-out. That's overwhelming US grids and creating even more data center bottlenecks. The explosion of requests is stretching approval timelines, forcing more applications to be vetted than ever before, and threatening to undermine US efforts to compete in the global AI race. It's what prompted Texas to put a pause on all data center construction.

Of course, the counterfactual by the pro-AI lobby immediately emerged, to wit: "If we don't allow these data centers to come online as fast as we can, then we might lose the future benefits of AI," Klaessig said, oblivious that the future benefits of AI are already as good as lost when one considers that at this very moment China is building 37 nuclear reactors while the US is building... zero.

Meanwhile, just like double (and triple, and quadruple) ordering in the semiconductor industry, some developers are pitching the same project to multiple utilities, with plans to push ahead with the application that lands the best deal and speediest approval. That creates what the industry calls "phantom" applications, complicating efforts to accurately forecast true power demand.

"Grid operators don't know which ones are real and which ones aren't," said Glenn Schwartz, who heads energy policy at consulting firm Rapidan Energy Group. He estimates that only 20% to 30% of the power that developers are seeking will go toward projects that get built, due in part to phantom requests.

Flooded grid queues aren't specific to data centers: renewables projects are regularly hampered by long wait times to connect to electrical networks across the world.

While it's impossible to determine how many requests are duplicates, Reid Ramdathsingh of Rystad Energy estimates about half of the applications are credible on the biggest US grid, operated by PJM Interconnection LLC, which serves 67 million Americans from Illinois to Virginia. He sees 14% of applications as legitimate on the main grid in Texas, a state that has seen the fastest data center growth.

PJM said in a statement that it has taken steps to improve the forecasting of large loads.

Of course, the 1,066 gigawatts of applications cited by Wood Mackenzie would require a staggering jump in US electric capacity. The amount represents 83% of the nation's total utility-scale generation capacity at the end of last year, according to the US Energy Information Administration. 

Making matters worse and adding to the chaos, many applications are from firms with no prior experience building facilities with energy needs that can rival mid-sized cities. The projects of those first-time developers are often disproportionately large, though generally less likely to come to fruition than those from well-established big technology firms, said Wood Mackenzie analyst Caitlin Connelly.

To crack down on the barrage of requests, many utilities have introduced steep upfront application costs, demanding big-money collateral and near-perfect credit ratings. Such requirements are squeezing small to mid-sized developers that often pay project costs up front and then sell completed facilities to well-capitalized AI companies.

Getting through a queue used to be straightforward with utilities offering up electricity to developers, according to Brad Richter, senior vice president of energy at Hut 8 Corp., which manages 11 data centers. Now, for the most part, utilities are saying, 'I don't have it anymore,'" Richter said, noting that those power providers are increasingly "closed for business."

As we reported previously, Texas Governor Greg Abbott ordered regulators on Aug. 3 to audit every data center seeking access to the state's main grid, effectively pausing those projects' approvals until reviews are complete. His surprise call puts $13 billion in industry revenues at risk and threatens to delay almost a fifth of America's data center pipeline in the medium term, according to BloombergNEF.

The Electric Reliability Council of Texas is tracking about 474 gigawatts of connection requests, with around 90% of those from data centers, according to Abbott. That total, Bloomberg notes paraphrasing verbatim what we said earlier this week,  "is more than five times the system's record peak demand."

Ercot uses multiple screening stages to distinguish between speculative and credible projects, a spokesperson said.

Exelon Corp., a utility owner serving customers from Illinois to Delaware, in July slashed its pipeline of energy demands from data centers by nearly 40%, leaving a queue of about 11 gigawatts. Exelon is prioritizing connection requests that are likely to come to fruition.

"Utilities are using what they'll refer to as a 'first-ready, first-served' model, as opposed to a 'first-come, first-served' model, to weed out folks that really don't have the capability to deliver," said Brian Janous of Cloverleaf Infrastructure LLC, which works with utilities to develop ready-to-build sites for data center users and providers. 

And for a more detailed follow up on what this lack of energy means, we urge readers to go through the latest report by Bernstein's Madison Rezaei, titled "Data Center Pipeline Probabilities: Separating the credible developers from dudes with PowerPoints," (available to pro subscribers), in which the bank writes that "of the 492GW of capacity currently in our dataset, we anticipate 33% (135GW) are credible builds. There are a series of haircuts in progress here, the most meaningful of which is the stage - we consider physical construction progress to be a major indicator of credibility, whereas shiny press releases only go so far. The tenant haircut is the next most pronounced - for a high-quality, IG tenant, certainty goes up; whereas for some of the miners or longtail neoclouds, credibility is lower. The sponsor haircut is also meaningful, though less so than the prior two - this is a fast-moving space and we didn’t want to overpenalize new entrants who are showing clear signs of development and tenant signing. And finally, while there is lots of talk of NIMBYism and legislative hurdles, it is not a big loss for DC builds today (in fairness, we are grandfathering in builds that are already in progress)."

Rezaei continues: "We’re infrastructure people, so we tend to skew conservative and not give meaningful credit for announcements before we see progress of builds. That’s how we end up at the ~33% build rate. Importantly, we are not taking a stand (today) on the actual timing of that pipeline. The dataset would suggest target delivery of ~88GW by the end of 2028 (highly unlikely). We would consider the 135GW number to be a 5-10 year build (almost certainly undercounting for the out years)."

Bottom line: very little of the bombastic projections you see and hear every single day about the pace and size of the AI rollout will ever materialize. That also means that so much of the capital already invested in said rollout will be wasted. The question is how big the negative impact on stocks will be when people finally do the math. 

More in the Bernstein report available to pro subs.

Tyler Durden Fri, 08/14/2026 - 09:20
Tyler Durden

zero rss