The AI industry’s growth is facing a core challenge: the gap between its bold claims and actual delivery. Bernstein Research’s review of over 7,000 planned US projects—tracked through satellite imagery, permits, and construction records—found only about 35% of those plans will reach completion. That means roughly 170 gigawatts of potential capacity out of the announced 500 gigawatts may remain unbuilt, earning the label “bragawatts” for announcements lacking real-world follow-through.
Who Gets Credibility Scores?
Bernstein’s evaluation distinguishes sharply between experienced operators and newcomers. Tech giants like Amazon, Google, Microsoft, and Meta received perfect credibility scores of 5 out of 5. Specialized data center firms such as Digital Realty, Equinix, and CoreSite also scored highly. In contrast, O’Leary Digital, backed by investor Kevin O’Leary, garnered just 0.35 out of 5, with Bernstein estimating only 3% of its announced 15 gigawatts will materialize.
The term “bragawatts” has become industry shorthand for overstated capacity claims, though its exact origin remains unclear. Madison Rezaei, a Bernstein analyst, first encountered it from an Equinix executive. The concept captures skepticism toward developers with limited construction experience.
O’Leary Digital Pushes Back
O’Leary Digital rejects the findings, arguing its projects are in early stages and that the analysis overlooks foundational work. The company cites 11,000 acres secured in Utah and 13,000 acres in Alberta, along with a 30-year development agreement in Utah. CEO Paul Palandjian emphasizes progress in planning, engineering, and permitting, noting projects designed for self-generated electricity. He prefers judging the company by milestones rather than speculative projections.
Other low-scoring developers include Vermaland and Joule Capital Partners. Joule has proposed 50 projects totaling 11.7 gigawatts, but Bernstein estimates only 11% will be built. Nearly all of Joule’s capacity remains at the announcement stage, with no visible construction or customer commitments. Vermaland’s $33 billion La Osa project in Arizona, initially planned for 2.2 gigawatts, was scaled back and rejected by regulators in late August. Bernstein assigned Vermaland a credibility score of 0.75 out of 5.
Sergio Toro, CEO of Aterio, confirms the trend in his firm’s data. While 30 to 40 gigawatts of new capacity are announced monthly, only 7 gigawatts begin construction each month, and just 2 to 3 gigawatts ever become operational. Many projects are abandoned before spending significant funds, often after multiple filings in different locations to test regulatory approval odds. Satellite imagery reveals little ground activity for most announcements, highlighting the disconnect between promises and progress.
Texas Grid Overrun: Speculative Demand
The mismatch between announcements and reality extends beyond capacity. Data center companies have requested 450 gigawatts of electricity demand on Texas’ main grid, more than five times the grid’s current 85 gigawatts. Many requests are speculative, clogging approval processes and distorting perceptions of actual need. Texas has paused new applications to filter out non-serious projects, a move expected to favor established players over speculative builds.
Even top-tier companies face delays. Amazon, for instance, has encountered challenges securing electricity and supplies for its AI-focused expansions. Major tech firms often commit to three or four times the capacity they need over two years, acknowledging that roughly a third of unstarted projects will fail or be delayed. Bernstein analysts note that the most credible developers over-commit by only about double what they realistically require.
Despite the overstatement, the AI infrastructure boom is not entirely unfounded. Bernstein estimates nearly 80% of capacity developed directly by the largest tech companies is credible. Most speculative projects fail before borrowing significant funds, limiting unpaid debt risks. The greater concern lies in stranded capacity, projects canceled, delayed, or never approved, leaving a trail of unmet promises.
Regulators Crack Down Nationwide
This issue spans beyond Texas. States like Virginia, Nevada, and Georgia have seen surges in speculative data center proposals, often from companies with no prior construction experience. Local regulators now demand proof of financing, engineering feasibility, and pre-leased space before granting permits. In Arizona, Vermaland’s $33 billion La Osa project collapsed after regulators rejected it in August, citing insufficient demand evidence and environmental concerns. The company had initially proposed 2.2 gigawatts, but Bernstein’s assessment suggested only 4% would materialize, a pattern seen across other low-scoring developers.
Credible projects continue to progress. These firms, Amazon, Microsoft, Google, and Meta, consistently earn 5-out-of-5 credibility scores due to their track records.
