AI's bottleneck is power and permits, and it pushes revenue out

The shortfall

Morgan Stanley analysts put the AI power shortfall at an estimated 80 gigawatts by 2029, according to the firm's podcast discussion, with utility transmission networks and substation equipment as the primary bottlenecks for data center power deployment. Data center developers are increasingly signing 15- to 20-year contracts for on-site power generation, the analysts say, and data center demand for battery storage is expected to exceed grid demand for new projects by 2030. U.S. re-industrialization is driving an estimated 10 gigawatts of incremental annual power demand. m:…1501

The constraint is not only generation. On Bloomberg Surveillance, commentators said AI infrastructure projects face delays from permitting, power and labor constraints, pushing revenue realization into 2027-2029. The same program noted 97 gigawatts of power production planned for 2028 that are at risk due to election outcomes. The through-line is timing: capacity that arrives late is revenue that arrives late. m:…0542 m:…9175

Storage is part of the workaround, but its scale is still modest: vehicle-to-grid programs in California currently dispatch between 500 and 1,000 megawatts during peak periods, Morgan Stanley's analysts note. Set against an 80-gigawatt shortfall, that figure shows why developers are signing decades-long generation contracts instead of waiting for grid capacity: the available fixes today are measured in megawatts, while the gap is measured in tens of gigawatts. m:…1501

The money side

Bank of America research discussed on its podcast puts AI capital spending on track to reach $1.2 trillion by 2027, with hyperscalers having issued $182 billion of debt this year, nearly 11 times the level seen in 2024, and hyperscaler free cash flow that could fall to negative $180 billion. Hyperscalers and semiconductor stocks lost more than $800 billion in market value during the recent earnings season, the discussion noted. Credit spreads for AI-related hyperscalers have widened by 25 to 30 basis points quarter-to-date, according to Bloomberg Surveillance, though non-tech AI-related bond deals were six times oversubscribed. m:…4186 m:…9175

One counterpoint in the same Bloomberg discussion: hyperscalers are seeing a 26% cash return on cash invested in AI, which the commentators said justifies continued high capital expenditure. The tension is plain: returns on deployed AI capital look strong while the cost of building the next wave of capacity keeps rising and arriving late. m:…0542

Buying power directly

WhiteFiber's filing shows one response: the company completed the acquisition of industrial property in Yadkin County, North Carolina, on October 28, 2026, for $60.0 million in cash, according to its 8-K. The properties are expected to be served with a combined minimum of 60 megawatts of gross electrical power supply, with potential for up to 198 megawatts within seven years. A $2.25 million earnest money deposit was applied toward the purchase price at closing. Unifi Manufacturing retained the right to occupy a portion of the property on a post-closing basis, the filing notes. m:…6013

This suggests the bottleneck is reshaping strategy: when grid power cannot be assumed, companies buy property with power attached, sign decades-long generation contracts, or repurpose other capacity. Morgan Stanley's analysts note automakers like Tesla and Ford repurposing underutilized EV capacity into energy storage businesses. The open question is whether these workarounds close an 80-gigawatt gap or simply move it: on-site generation and property deals add megawatts project by project, while the shortfall is measured in tens of gigawatts. m:…1501