Phantom AI Data Centers: How to Verify Real Demand

On paper, the AI data center buildout has never looked bigger. Hyperscalers are expected to deploy roughly $800 billion in new infrastructure in 2026 and more than $1 trillion in 2027. Yet a wave of "phantom" projects, announced with great fanfare and never built, or announced at a scale no physically constrained timeline can deliver, has quietly broken the industry's most important planning tool: the ability to forecast power demand.

Glenn Schwartz, head of energy policy at consulting firm Rapidan Energy Group, put it bluntly this week: "Grid operators don't know which ones are real and which ones aren't." That single sentence explains why utilities keep building speculative substations, why regulators sit on interconnection requests they may never honor, and why your electricity bill is rising even before most of these facilities have a single server inside.

Why Planning Is Breaking Down

Start with the numbers. Nebius raised its year-end 2026 contracted power forecast to 5 gigawatts. CoreWeave expanded active power by nearly 500 megawatts in a single quarter to reach 1.5 gigawatts. Bitcoin miners are re-tooling multi-gigawatt sites for AI workloads. Add it all together and the aggregate of announced projects vastly exceeds what can plausibly be built within stated timeframes.

Part of the problem is the "shotgun" request. Developers pitch the same project to several utilities at once, reserving interconnection slots they have no intention of using, and grid operators are left holding phantom reservations. The queue fills with entries that will never consume a single megawatt.

Circular financing muddies the picture further. Nvidia has arranged more than $500 billion in capital pools to finance customer purchases of its own chips. CoreWeave alone carries a $104 billion revenue backlog and raised more than $10 billion in debt in a single quarter. When suppliers fund the demand for their own products, analysts warn, some of that "certain" demand is really speculative capacity that may never draw power from the grid.

The labor market tells the same story. Only about 30% of the mechanical, electrical, and plumbing workforce qualified for data center work lives in the regions hosting 70% of projects. At peak recruitment, the United States adds roughly 60,000 craft laborers a year, a ceiling that caps how fast any of these projects can physically be built. Supply-side warnings from Super Micro, Foxconn, and Arista Networks all point the same direction: the industry stays supply-constrained until at least 2028.

A Checklist for Separating Real from Phantom

You do not need to be a grid operator to apply basic due diligence. The next time you hear about a new AI campus, run it through this checklist:

• Check the construction timeline. A gigawatt-scale build takes three to five years of physical construction after permits. Any announcement promising delivery in under two years is almost certainly a placeholder.
• Ask who the end customer is. If the tenant is a supplier financing its own buyer, treat the demand signal as speculative until firm contracts exist.
• Compare the site with the labor pool. A mega-campus in a region with no experienced mechanical and electrical contractors will slip its schedule no matter how much money backs it.
• Verify the interconnection status. A slot in the queue is not a commitment. Most queued capacity never gets built.
• Watch for double-booking. The same project pitched to three utilities is three queue entries but only one campus.

None of these checks separates a good project from a bad one with certainty, but together they flag the phantoms early, which is more than the current system manages to do.

What This Means for Your Power Bill

Household electricity prices have already climbed 10.1% over two years, faster than overall inflation, as utilities front-load grid investment on the back of uncertain AI demand. The Federal Reserve is caught in the same fog. Chair Kevin Warsh called AI a "significant disinflationary force" before taking office, but told the Senate he suspects AI demand will raise measured prices over the next 12 months, adding that whether that counts as inflation is up to the central bank.

Analysts do not expect the gap between announced and deliverable capacity to narrow before 2028 at the earliest. Until then, treat every headline-making megawatt as a hypothesis rather than a fact.

1. Audit your power contracts for escalation clauses tied to data center load, and price the risk before you renew.
2. Demand real interconnection data from your local utility instead of queue counts, which conflate genuine builds with shotgun applications.
3. Discount capacity-based AI stock theses unless the company discloses firm contracts, tenant names, and construction milestones.

The AI boom is real. The question is which parts of it will still be standing when the grid, the labor pool, and the supply chain finish their own reality check somewhere around 2028. Learning to spot the phantoms now is the cheapest insurance you can buy.