The grid can't tell the difference between a real bottleneck and a hallucination.
The Summary
- Bernstein Research estimates only 35% of announced US data center capacity will actually get built, with the industry now tracking "Bragawatts" (dubious megaprojects) and "ghost projects" (duplicate grid connection requests)
- In Texas alone, electricity connection requests have exceeded five times the state's existing supply, creating phantom congestion that makes real power shortages look worse than they are
- The largest real projects now breaking ground include a 17-gigawatt campus in Texas and a 10-gigawatt facility in Ohio contracted mostly to OpenAI, each requiring more electricity than most American cities
- Even Amazon, among the world's most experienced builders, is struggling to secure adequate power for its AI infrastructure
The Signal
The AI data center boom has a credibility problem layered on top of its power problem. Thousands of announced projects will never materialize, and the gap between what gets announced and what gets built is making it nearly impossible to separate infrastructure planning from infrastructure theater.
Start with the numbers. The US has over 3,000 operating data centers and 1,500 more under development, most in rural areas. But Bernstein's analysis suggests two-thirds of the announced capacity is smoke. Some developers are announcing projects without securing electricity, construction permits, or customers. Others are filing multiple grid connection requests for the same project at different locations, hoping one sticks.
"In Texas, these requests have reached more than five times the state's existing electricity supply."
The scale of what's actually being built is still staggering. Consider the real megaprojects:
- A Texas campus pursuing 17 gigawatts of capacity
- An Ohio facility targeting 10 gigawatts, most contracted to OpenAI
- 1 gigawatt of continuous supply can power roughly 835,000 average American homes
US data centers consumed about 313 terawatt-hours in 2025, representing nearly 40% of global data center electricity use. That's before the current wave of AI-specific infrastructure comes online.
The phantom projects create a perverse dynamic. Grid operators see connection requests that suggest apocalyptic demand. Regulators see bottlenecks everywhere. Communities mobilize against development that was never viable to begin with. Meanwhile, serious builders like Amazon face real constraints trying to power equipment they've already bought.
Key distinctions between real and ghost projects:
- Real projects have multi-year power purchase agreements with utilities
- Real projects have identified customers or anchor tenants
- Real projects break ground within 18 months of announcement
The Bragawatt phenomenon isn't new to infrastructure. Every boom cycle produces a layer of speculative announcements designed to attract capital, partnerships, or attention. What's different here is the urgency. AI compute demand is real and growing fast. Companies need to signal they're building capacity to reassure investors and customers. But the gap between announcement and reality has never been wider or more consequential.
The Implication
If you're evaluating companies in the AI infrastructure stack, the power question is now the first question. Who has contracted capacity, not who announced plans. The difference between a company with 2 gigawatts of permitted, powered data center space and a company with a press release about 10 gigawatts is the difference between building the future and betting on it.
For grid operators and policymakers, the challenge is learning to filter signal from noise without blocking legitimate projects. Texas receiving connection requests for five times its grid capacity doesn't mean Texas needs to quintuple generation. It means the request process has become a cheap option on future development. Smarter jurisdictions will start requiring earnest money or milestone commitments before counting projects as real demand.