Two-Thirds of US AI Data Center Pipeline Is ‘Bragawatts,’ Bernstein Warns
TREE NEWS reports: Of the 586 gigawatts (GW) of announced, under-construction, or land-banked data center capacity in North America, only 199 GW—roughly 34%—is credible. The remaining 387 GW, or about two-thirds, is what the firm calls “Bragawatts”: capacity that exists only in press releases and PowerPoint presentations, not in steel and concrete. The findings challenge the market’s current obsession with political risks, which Bernstein says are a far smaller threat to the AI infrastructure buildout than project quality.
Market Implications
The report has sweeping implications for investors across asset classes. For equities, the AI trade—spanning semiconductor makers, power equipment suppliers, and cooling system providers—has been priced on demand forecasts that may include this phantom capacity. Bernstein estimates that the 387 GW of Bragawatts represent $3 trillion to $6 trillion in construction spending and roughly $19 trillion in total supply-chain expenditure that may never materialize. Companies with order books tied to these projects could face revenue shortfalls.
Within the credible 199 GW, hyperscalers and established data center operators dominate: they account for 76% of credible capacity despite announcing only 37% of the pipeline. By contrast, “other” developers—non-hyperscalers, non-colocation, non-miners, and non-neoclouds—announced 308 GW but are expected to build only 33 GW, an 11% survival rate. The top four hyperscalers—Amazon, Google, Microsoft, and Meta—alone account for 78 GW, or 39% of all credible capacity.
Tenant credit quality is a major discount factor. Leases signed by Google, Amazon, Microsoft, Meta, or Nvidia receive a 1.00 multiplier, while OpenAI standalone is rated 0.50—a coin flip—after the company cut its funding target from about $30 billion to $20 billion. Oracle stands at 0.75, and the Oracle/OpenAI combination (the Stargate tenant) at 0.80. Neoclouds like CoreWeave, Nebius, and Lambda are rated 0.45, below xAI at 0.60 and far below enterprise tenants like Jane Street and Bloomberg at 0.80.
Political risk, which has dominated recent market commentary, is the smallest of the four filters in Bernstein’s model, removing only 28 GW. Development stage (198 GW removed), tenant quality (113 GW), and developer track record (48 GW) are far more significant. Texas is the most penalized state, with about 43 GW of credible capacity at risk of delay, but Bernstein views this as a timing issue rather than cancellation.
For bonds, the risk is concentrated in utilities and power providers that have based capital expenditure plans on projected data center demand. If a significant portion of the pipeline fails to materialize, stranded asset risk could rise. Commodities, particularly copper and other construction materials, may also see demand forecasts revised lower. Crypto markets, which have traded in sympathy with AI infrastructure optimism, could face spillover effects if the AI trade cools.
Key Takeaways for Investors
- Focus on credible developers: Hyperscalers and blue-chip colocation providers (Digital Realty, Equinix, QTS, CoreSite, EdgeConneX, Compass, Aligned, NTT, Brookfield) have the highest survival rates. Bernstein assigns 16 developers a perfect 5.0 score.
- Discount tenant risk: Leases with OpenAI, Oracle, and neoclouds carry significant credit discounts. The market may be overestimating the durability of these commitments.
- Political risk is overhyped: While state-level pauses in Texas, Virginia, and elsewhere make headlines, they are more likely to delay than cancel projects. The bigger risk is project viability.
- Watch for downward revisions: Bernstein expects the 34% credibility ratio to fall further as tenant discounts increase and financing conditions tighten.
The AI infrastructure boom is real, but the market is mispricing the risk. Investors should separate the signal from the noise—and the Bragawatts from the real ones.
