How ‘Phantom’ Data Center Projects Are Making It Impossible to Forecast AI Power Demand — ‘Grid Operators Don’t Know Which Ones Are Real’

The phenomenon of "phantom" data center projects – those that are announced but never built, or announced far in excess of what physical constraints will actually allow – has emerged as one of the most disruptive forces undermining accurate forecasting of AI-related power demand. 

The scale of announced spending is staggering, with hyperscalers alone expected to deploy roughly $800 billion in 2026 and over $1 trillion in 2027. 

Household electricity prices have already risen 10.1% over two years, faster than overall inflation, as the anticipation of data center loads drives infrastructure investment. 

Nebius (NBIS) raised its year-end 2026 contracted power forecast to 5 gigawatts, CoreWeave (CRWV) expanded active power by nearly 500 megawatts in a single quarter to reach 1.5 gigawatts, and numerous Bitcoin (BTCUSD) miners are converting facilities totaling multiple gigawatts to AI use – and yet the aggregate of all announced projects vastly exceeds what can plausibly be built within stated timeframes.

In fact, many of the projects behind those investment decisions exist only as announcements, interconnection queue entries, or financial commitments that may never materialize at their stated scale or timeline.

“Grid operators don’t know which ones are real and which ones aren’t,” said Glenn Schwartz, head of energy policy at consulting firm Rapidan Energy Group, in comments to Bloomberg.

The result is that utility planners, grid operators, and policymakers face an impossible task.

Until the gap between announced phantom capacity and physically deliverable capacity narrows – something unlikely before 2028 at the earliest – reliable forecasting of AI power demand will remain essentially impossible.

Labor Shortfalls Complicate Forecasting, Too

In addition to the “shotgun”-style requests described by Bloomberg’s reporting, where developers pitch the same project to multiple utilities, there are other factors that could create an enormous gap between planned and realized capacity.

The skilled labor shortage represents perhaps the most fundamental constraint. The US mechanical, electrical, and plumbing labor pool qualified for data center work is far smaller than headline figures suggest, with only 30% of MEP labor residing where 70% of projects are located. 

At recent peak recruitment rates, the US has added approximately 60,000 combined MEP craft laborers annually, a figure that likely represents a ceiling rather than a floor for growth. Training bottlenecks mean that data center capacity additions cannot grow materially faster unless labor productivity improves dramatically or data centers cannibalize workers from the rest of the construction economy.

This labor constraint creates a cascading forecasting problem. Companies announce gigawatt-scale projects and secure power interconnection agreements, but the physical construction timeline stretches far beyond initial projections. 

Circular Financing vs. End-User Demand

There are also financing structures to consider. 

Nvidia (NVDA) has arranged more than $500 billion in dedicated capital pools to finance customer projects, while CoreWeave alone carries a $104 billion revenue backlog and has raised over $10 billion in debt in a single quarter. 

These financial commitments create the appearance of certain demand, but as analysts have noted, this circular financing – where suppliers help fund purchases of their own products – raises questions about how much represents genuine end-user demand versus speculative capacity that may never draw electrons from the grid.

Supply chain constraints add another layer of opacity to power demand forecasts. 

Super Micro Computer's (SMCI) CEO attributed revenue shortfalls to customer delays in power, cooling, and networking infrastructure, while Foxconn cautioned that AI server rack volumes depend on how much advanced chip packaging capacity can actually be secured. 

Arista Networks (ANET) explicitly stated its outlook is set by what it can build, as opposed to what customers are requesting, with the industry expected to remain supply-constrained until 2028.

What the Fed Has Said About AI Spending

The Federal Reserve itself is struggling with this uncertainty, unable to determine whether AI infrastructure spending represents a temporary inflationary impulse or a permanent structural shift in electricity demand. 

“AI will be a significant disinflationary force, increasing productivity and bolstering American competitiveness,” wrote Fed Chair Kevin Warsh in 2025, before assuming the top job at the central bank. 

More recently, Warsh took a somewhat measured tone in his testimony before the Senate.

"Will [AI] increase measured prices over the course of the next 12 months? I suspect it will. Whether that's inflationary or not, that's up to the Federal Reserve — and we're going to have something to say about that.”

While that doesn’t add much clarity to the central bank’s stance on the AI buildout, it’s fair to say that the underlying data – muddled by phantom bids and circular financing – isn’t providing policymakers with very reliable information to analyze.

This article was created with the support of automated content tools from our partners at Sigma.AI. Together, our financial data and AI solutions help us to deliver more informed market headline analysis to readers faster than ever.