Building U.S. data centers is getting so expensive that AI companies are moving overseas

Rising U.S. power and water costs, voter backlash, political concerns, and federal regulatory obstacles are pushing AI hyperscalers to consider building data centers overseas. Big Tech is also seeking cooler climates abroad to reduce operating costs, creating potential headwinds for domestic AI infrastructure expansion and U.S. data-center investment.

Analysis

The binding constraint for hyperscalers is increasingly time-to-power rather than GPU availability. That shifts incremental AI economics away from pure semiconductor capacity and toward firms that can monetize power procurement, electrical equipment, cooling, and modular deployment across jurisdictions: Vertiv (VRT), Eaton (ETN), Schneider Electric (SU.PA), Quanta Services (PWR), and Caterpillar (CAT). Their addressable market expands if workloads become geographically dispersed, while hyperscaler returns on invested capital face pressure from duplicated network, compliance, and data-residency infrastructure.

The more important second-order effect is a potential reset in U.S. merchant-power and utility load-growth assumptions. CEG, VST, NRG, and selected regulated utilities have been awarded scarcity premia partly on projected data-center demand; a meaningful offshore diversion would not eliminate demand, but could defer contracted load and reduce the urgency of high-cost generation additions. The market should distinguish signed, deliverable interconnection agreements from speculative pipeline announcements—valuation risk is greatest where AI-driven load growth is embedded in 2027-29 EBITDA estimates without contracted power offtake.

Over 1-3 months, this is principally a narrative and permitting-risk issue rather than an earnings event. Over 6-18 months, foreign siting can favor Digital Realty (DLR) and Equinix (EQIX) if they convert their existing international campuses and enterprise connectivity into a premium for low-latency, sovereign AI deployments; however, overseas power prices, local permitting, and sovereign-cloud requirements can offset lower construction friction. The contrarian view is that political resistance raises the value of already-secured U.S. power rather than forcing a broad migration: evidence of hyperscalers signing long-dated domestic PPAs or taking equity stakes in generation would invalidate the load-deferral thesis.

Key Decisions for Investors

• Favor a 6-12 month long VRT / short CEG relative-value position rather than a directional AI trade: VRT captures geographically diversified electrical-and-cooling capex, while CEG is more exposed to a premium load-growth narrative. Reassess if CEG discloses incremental contracted data-center load or signed PPAs sufficient to raise 2027 EBITDA guidance.
• Maintain or add to ETN and PWR on 3-6 month weakness; both benefit from grid interconnection, switchgear, and transmission spend regardless of whether compute is located in the U.S. or abroad. Thesis fails if order backlog growth decelerates materially or lead times normalize faster than expected.
• Place DLR and EQIX on a watch list for evidence of AI leasing conversion at European, Canadian, and Nordic campuses; initiate only after disclosures show signed capacity and economics above conventional colocation returns. The missing data is the share of capacity pre-leased to AI customers versus merely reserved.
• Avoid adding to AI-driven U.S. utility exposure solely on announced data-center pipelines. Require a disclosed power contract, interconnection date, and utility capital-recovery mechanism; absent those, permitting delays can push cash-flow realization beyond current valuation horizons.