How AI concerns moved to the center of the midterm debate

Citi says political and regulatory opposition to AI data centers is rising ahead of the November U.S. midterm elections, with local moratoriums and at least 15 state legislatures proposing tighter restrictions. The effect has so far been limited largely to speculative early-stage projects, while late-stage developments with secured sites and grid access continue progressing amid sustained AI infrastructure demand. U.S.-China AI model performance gaps have narrowed to ultra-low single digits, but Chinese developers remain constrained by U.S. advanced-chip export controls; Citi sees model-safety restrictions and a potential resulting excess of computing capacity as a larger future risk.

Analysis

The key investable effect is not aggregate data-center demand but the widening value gap between projects with secured interconnection and merely announced capacity. Grid-connected campuses, transmission equipment, and dispatchable clean-power contracts should command scarcity premiums, while land banks and merchant developers without power rights face rising carrying costs and a higher probability of stranded development spend. This favors CEG, ETN, GEV, PWR, DLR and EQIX over early-stage private developers; it is less directly material to C despite its role as the research source.

For AMZN and META, political friction can perversely reinforce incumbent advantages: only the largest buyers can pre-fund generation, absorb multiyear permitting delays, and sign bespoke power contracts. Over the next 1-3 months, permitting headlines may pressure AI-capex beneficiaries indiscriminately, creating an entry point in power-constrained infrastructure rather than a reason to short hyperscalers. Over 6-18 months, higher all-in power and construction costs raise the minimum efficient scale for AI, supporting the dominant platforms if they can translate capacity into revenue.

The underappreciated tail risk is a demand-side rather than supply-side shock: a material safety restriction, a hyperscaler capex reset, or sharply improving model efficiency would turn contracted but unutilized compute into a liability. That scenario would first hit equipment order books and data-center REIT leasing spreads, then reduce the scarcity value embedded in nuclear and renewable PPAs. The thesis is falsified by two consecutive quarters of lower AMZN/META AI infrastructure capex guidance, weakening CEG contracted-power pricing, or evidence that major campuses are deferring energization rather than just delaying construction.

Key Decisions for Investors

• Accumulate CEG on broad utility or permitting-related weakness; use a 6-12 month horizon and size against XLU rather than as an outright duration-sensitive utility bet. Long CEG / short XLU targets continued repricing of firm clean power scarcity; exit if contracted-power economics weaken or CEG loses key renewal visibility.
• Initiate a 3-6 month long ETN / short XLI pair on pullbacks. Electrification equipment has higher exposure to interconnection upgrades and power-quality spending than the average industrial; risk is a broad hyperscaler capex pause, with a stop if ETN order growth or backlog conversion materially decelerates.
• Maintain AMZN and META as relative AI-capex winners, but do not add solely on infrastructure headlines. Add only after the next earnings reports confirm that incremental capex is accompanied by cloud, advertising, or AI-product monetization; a capex increase without revenue conversion is a multiple-compression risk.
• Avoid chasing speculative data-center construction exposure until project-level evidence shows secured power, permits, and signed customer commitments. Treat announced capacity without these milestones as an alert rather than investable backlog, particularly if local restrictions broaden from new approvals to operating conditions.
• Buy downside hedges on the AI-infrastructure basket through 6-9 month puts on SMH or a small short position in DLR if policy discussion shifts from siting restrictions to model-training limits. This hedge pays in the less-consensus scenario of compute-demand impairment rather than supply delay.