The hidden cost of AI: Why your town is negotiating with Amazon and Microsoft

PJM’s market monitor projects data-center demand could drive about $23B in higher electricity costs for mid-Atlantic and Midwest customers through at least 2028, forcing new “community contract” models. Local deals show the upside: Hobart, Indiana secured ~$200M in commitments from Amazon/AWS (plus hiring commitments) and Indiana’s LaPorte and Jasper County projects cite ~$1B and ~$98M upfront benefits, while utilities are increasingly shifting costs/financing to large loads (with agreements expected to return ~$1.4B to existing customers). Regulatory pressure is rising—FERC required all six regional grid operators to revise or justify interconnection/upgrades rules for large new users—supporting a broader consensus that data-center customers should bear upgrade costs while communities share benefits.

Analysis

This is less a pure AI-demand story than a financing-regime change. If large loads must prepay interconnects and generation, the binding constraint shifts from GPU supply to balance-sheet capacity and permitting, which favors the hyperscalers with the cheapest capital and the strongest political cover. That is a relative advantage for AMZN and GOOGL versus smaller AI/cloud competitors that cannot easily self-fund stranded infrastructure.

For utilities, the economic prize is not just incremental load but a cleaner regulatory bargain: ratepayer backlash declines when the load pays its own way, and the utility keeps the growth without the reputational hit. The bigger second-order winners are grid-capacity and backup-power suppliers, while the losers are data-center REITs, land-bank owners, and industrial users competing for the same constrained transformers, switchgear, and interconnect queues. In PJM-like markets, the first earnings impact is delay, but the 6-18 month effect is structurally tighter capacity and higher clearing prices.

The key catalyst path is regulatory, not operational: FERC implementation, state commission rulings, and local votes. If those bodies soften cost allocation or if AI capex slows sharply, the thesis reverses quickly; if more projects are blocked and capacity prices stay elevated, utilities with enforceable tariff structures should outperform. The market is still treating this as a simple 'AI capex up' narrative, but the real alpha is in who controls the megawatts and who gets paid to provide them.

Key Decisions for Investors

• Long XLU / short XLK over 3-6 months: utilities with enforceable large-load tariffs should re-rate as political risk falls, while the market starts discounting higher AI infrastructure capex; cut the trade if XLU underperforms XLK by more than ~8% after a pro-load regulatory ruling.
• Small tactical long NI vs. XLU on pullbacks: NI is an early proof point for the 'self-funded load' model and could outperform if more contracts are signed, but size modestly because the thesis depends on execution and commission approval; invalidation is a dilution of cost recovery in Indiana.
• Prefer AMZN and GOOGL on weakness over MSFT on any AI power-cost selloff: both have the scale to finance dedicated power solutions and convert scarcity into moat expansion; the relative trade breaks if MSFT announces comparable behind-the-meter or self-generation commitments.
• Avoid chasing data-center REITs and grid-constrained land banks until permitting velocity improves; if project delays continue, the near-term risk is deferred revenue rather than lost demand, but valuation compression can persist for multiple quarters.
• Watch PJM capacity-price prints and FERC implementation dates as the high-conviction alerts; if capacity prices roll over or the rules get watered down, reduce utility exposure and cover relative-value longs.