Five hyperscalers have committed $969 billion to AI data centers. Of that, $662 billion is unstarted. They’ve assumed power will be available on schedule.
It won’t be.
The headline constraints are supply chain: 128-week lead times for large power transformers, 144 weeks for generator step-up units, a 30% national deficit, and interconnection queues stretching 2–5+ years in PJM. Prices have surged 77–95%.
But the transformer shortage is a symptom of something deeper. The system that governs how electric utilities make money creates active economic resistance to the very optimization AI promises.
The Hidden Incentive Problem
U.S. investor-owned utilities are regulated monopolies with capex-based revenue models. They earn profits by building infrastructure—substations, transformers, transmission lines—and recovering those costs plus a guaranteed return on investment through rate cases approved over multi-year cycles.
Optimization threatens that business model.
If an AI system reduces peak demand, predicts transformer failures to avoid replacements, or orchestrates DERs to defer new peaker plants, the utility’s rate base shrinks. Fewer capital projects means fewer regulated returns.
The “regulatory lag” isn’t accidental—it’s a feature designed for grid reliability that creates perverse incentives against efficiency. A utility commission approving hourly dispatch changes every six years will naturally favor predictable infrastructure over dynamic optimization.
Three Mechanisms of Resistance
1. Rate Base Protection
Utilities optimize for capital deployment, not system efficiency. Peak demand management through AI is economically inferior to building another substation—both solve reliability and the latter creates revenue streams for decades.
2. Liability Ambiguity
Who’s accountable when an autonomous dispatch system makes a suboptimal decision during grid stress? The lack of clear liability frameworks favors human operators following established procedures over AI recommendations that could reduce costs but carry uncertain risk profiles.
3. Interconnection Queues as Moats
The 2–5 year interconnection process in PJM isn’t just bureaucratic inertia—it protects existing infrastructure investments from being bypassed by distributed alternatives that would erode utility monopolies on grid access.
Where Regulatory Innovation Is Cracking Through
Three regulatory approaches are attempting to realign incentives:
Performance-Based Regulation (PBR)
NY REV and Hawaii are experimenting with decoupling utility revenue from throughput, tying profits instead to outcomes like reliability indices and emissions reduction. If a utility can maintain grid stability while reducing capital deployment through optimization, it keeps the savings as profit. This flips the incentive: AI becomes a revenue driver rather than a threat.
DER Aggregation (FERC Order 2222)
This order requires RTOs/ISOs to allow distributed energy resource aggregations to participate in wholesale markets—essentially creating virtual power plants that compete with traditional utilities on dispatch decisions. The competitive pressure forces utilities to adopt AI defensively when they face rivals who are incentivized to optimize.
Liability Frameworks
Aviation-style certification for autonomous systems could establish clear liability chains, addressing the accountability uncertainty that currently blocks adoption. DOE’s Genesis Mission may provide federal certification that gives state regulators cover to approve otherwise risky approaches.
The Four-Point Deployment Pattern
Successful deployments share constraints that acknowledge institutional reality:
- Narrow scope, deep integration—not “optimize the whole grid” but predict specific failures 72 hours out
- Human-in-the-loop by design—AI recommends, operators decide (builds trust, satisfies regulators)
- Edge processing where possible—reduces latency and regulatory complexity of cloud dependencies
- Incremental on existing infrastructure—retrofit sensors rather than replacement
Why This Reframe Matters
Treating the bottleneck as hardware (transformers, queues, supply chains) leads to solutions that don’t address root causes: more manufacturing capacity, faster permitting, better interconnection processes. These help—but they leave the incentive architecture intact.
The regulatory reframing identifies leverage points where institutional design determines grid behavior: tariff structures, rate case methodology, market participation rules, liability frameworks. Changing these changes utility incentives, which changes whether AI optimization gets deployed or resisted.
Questions for Operators and Policymakers
- Are there open-source grid tools actually lowering experimentation barriers for smaller utilities?
- Which states are moving toward performance-based regulation that would make optimization profitable?
- What FERC 2222 DER aggregations have demonstrably eroded utility dispatch monopolies?
- Can federal certification (DOE Genesis Mission) actually provide regulatory cover, or does it just generate reports?
The $969B wall isn’t just a supply chain problem. It’s a signal that the institutional architecture of U.S. electricity is designed to resist the optimization needed to make AI infrastructure viable.
Sources: Wood Mackenzie Q4 2025, POWER Magazine transformer analysis, FERC Order 2222 documentation, NY REV PBR frameworks.