70% of US grid interconnection requests are ultimately withdrawn, according to Berkeley Lab.1 That figure now sits at the center of the AI infrastructure investment thesis.
The industry has committed roughly $5.2 trillion in AI data center capex through 2030.2 Goldman Sachs projects global data center power demand to surge 165% by 2030 versus 2023 levels.3 The US power grid was built for 1% to 2% annual demand growth — not an exponential, decade-long surge.2
Investor Kevin O'Leary has argued publicly that 50% of planned US data centers will never be built due to grid constraints.4 Even a more conservative outcome — 30% to 50% reduction in commissioned capacity through 2030 — restructures hyperscaler return profiles.
Capex Without Capacity
Microsoft, Amazon, and Alphabet have each committed to aggressive multi-year AI infrastructure spending. Markets have priced in delivery. Grid delays break that assumption.
Power constraints compress returns in two stages. Capex deploys before facilities come online — land, permits, and hardware spend without revenue. Then delayed capacity delays AI service revenue. The gap between cash outflow and monetization widens quarter by quarter.
Actual data center completions versus quarterly capex guidance will become the sector's critical tracking metric. FERC interconnection queue approval rates are the leading signal. While withdrawal rates hold above 70%, build timelines slip.1
Valuation Exposure
Hyperscaler price-to-earnings multiples embed assumptions that AI monetization scales with infrastructure spend. A structural supply constraint breaks that model.
Current valuations imply capex translates into productive capacity on schedule. Grid bottlenecks mean capex is deployed but capacity is delayed by months or years. Return-on-invested-capital math shifts accordingly.
The 2027 to 2028 window is when AI infrastructure investment was expected to generate meaningful earnings contribution. That is also precisely when delayed facilities — from interconnection requests withdrawn today — would have come online.
The Binding Constraint
Chips are no longer the bottleneck. Power is. Semiconductor supply chains have improved substantially. Grid interconnection queues have not.
For hyperscaler stocks, the risk is not that AI demand disappoints. It is that infrastructure supply cannot keep pace with demand — and the capital committed delivers returns years late.
A 30% to 50% constraint on a $5.2 trillion buildout, sustained through 2030, is a structural downgrade to earnings visibility for the sector's largest names.2
Sources:
1 Berkeley Lab, US Grid Interconnection Queue Analysis
2 Industry Infrastructure Forecasts, AI Data Center Capex Projections Through 2030
3 Goldman Sachs Research, Global Data Center Power Demand Forecast 2030
4 Kevin O'Leary, Public Commentary on US Data Center Grid Constraints
