The Subnational Sovereigns

Position: The investment thesis around AI infrastructure is evolving from a national capex story to a geographically concentrated one driven by state-level regulatory geography. The fracture that follows will reverberate across the next decade influencing where and how quickly data center capacity expands or hits a growth ceiling.

Methodology: AI-assisted evidence infrastructure · Human-directed thesis · Primary-source verified

The investment thesis around AI infrastructure is evolving from a national capex story to a geographically concentrated one. Two events in the next eight months will significantly influence the geography of AI development for the decade. FERC's June 2026 ruling on PJM data center co-location rules sets the regime for the largest constrained market. The November 2026 midterm elections determine whether federal preemption appetite hardens or stalls.

By 2028, the US AI buildout will fracture along state lines. Five permissive energy-rich states (Texas, Louisiana, North Dakota, Iowa, Nebraska) will absorb 55-65% of net-new hyperscaler capacity. Five restrictive climate-mandate states (Virginia, California, New York, New Jersey, Oregon) absorb 13-20%, with Virginia's share dominated by grandfathered pipeline rather than new permits. The remainder lands in transition states deploying gas-plus-carbon-capture or other fossil-plus-decarbonized solutions as compromise.

Underlying structural forces matter more than the politics. Shale gas spent a decade looking for a long-dated, investment-grade demand vector after coal-to-gas saturation. AI is it. The geographic fracture is the physical expression of this codependence, not an accident of regulatory cycles. Federal preemption succeeds on AI model regulation but fails on data center infrastructure, because state environmental police power is the strongest federalism shield in US constitutional law. In other words, shale gas needed AI before AI needed shale, and the emerging U.S. AI map is just that pairing made visible.

The capital reallocation that follows is not cycle-dependent, but a more structural shift in AI development over the coming decade. The thesis fails only if permissive top-five share falls below 50% by end-2028 AND federal preemption successfully overrides state DC infrastructure authority. All else is sensitivity.

A year ago, the conversation about AI infrastructure focused on scale at all costs. Could hyperscaler capex sustain $400 billion, $500 billion, $600 billion annual run rates? Could the grid absorb 50, 70, 100 gigawatts of net-new data center load by 2030? Sell-side aggregation models smoothed national capacity into a single number. Federal AI policy framework signaled coordinated industrial policy. Hyperscaler capex velocity was assumed to overcome local friction.

That framing was rational. It also collapsed.

In the first six weeks of 2026, more than 300 data center bills were filed across 30 states. Twenty-seven states are advancing legislation. California, Ohio, and Utah have already enacted laws beyond the federal voluntary Ratepayer Protection Pledge. FERC ordered PJM in December 2025 to develop new rules for large-load co-location at power plants. The White House issued a National Policy Framework for AI in March 2026. Senator Tom Cotton introduced the DATA Act of 2026, creating a new "Consumer-Regulated Electric Utility" category that did not exist 18 months ago. The Trump administration's AI executive order of December 2025 explicitly preempts state AI model regulation while preserving state authority over data center infrastructure.

The fragmentation is no longer a forecast. The question is which way it breaks.

The conversation is moving from "how big" to "where." The directional resolution, the magnitude of geographic concentration that follows, and the binary events that determine outcomes are not yet fully developed in the AI infrastructure capital allocation models that will be tested in 2027 portfolio reviews.

This is where the math becomes specific. Bottom-up state-level capacity analysis, anchored to ERCOT's own forecast haircut methodology, produces a national net-new range of 50-80 gigawatts of hyperscaler AI capacity through 2028. The binary split across the country is not symmetric. It is structural.

The mechanism predates the politics by a decade. The political fracture playing out in 2026 across state legislatures, PUC dockets, and federal preemption initiatives is the visible surface of a structural codependence between two industries that needed each other before either knew it.

After the coal-to-gas conversion cycle saturated in the late 2010s, awash in supply, US shale gas faced a demand-side overhang. LNG export expansion absorbed some supply. Industrial demand grew modestly. But shale economics, with high upfront capex, long-lived production, and infrastructure intensity, needed a new long-dated, investment-grade demand vector. And then the AI boom emerged exposing a supply-demand imbalance across the US grid.

The codependence runs both directions. Hyperscaler AI workloads require firm power at densities and reliabilities that intermittent renewables paired with current battery storage cannot deliver at scale before 2030. Gas turbines fill the gap. The aeroderivative bridge fleet alone has scaled rapidly. FTAI Aviation's CFM56 conversion business launches commercial production in 2026. Crusoe ordered nineteen GE Vernova natural gas turbines for behind-the-meter deployment. VoltaGrid committed 2.3 gigawatts to Oracle's Stargate Texas project. Gas needs AI demand to clear the long-dated supply curve. AI needs gas to deliver compute at the scale and pace required to meet demand growth and aggressive growth aspirations.

The geographic fracture across US states is the physical expression of this codependence.

Permissive states have shale basin co-location. Texas sits over the Permian and Haynesville. Louisiana sits over Haynesville. North Dakota sits over the Bakken. Iowa and Nebraska are pipeline-served. They also have institutional reform appetite. Texas SB 6 establishes new large-load interconnection standards. ERCOT launched two new organizations in January 2026: Interconnection and Grid Analysis, and Enterprise Data and AI. Entergy's Integrated Resource Plan supports Meta's 2-gigawatt Hyperion campus in Richland Parish.

Restrictive states have codified the inverse. California's SB 100 mandates 100% clean electricity by 2045; SB 1168 specifically targets data center natural gas surcharges. New York's CLCPA would require all data center energy from renewable PPAs by 2040. Oregon's HB 2021 requires 100% clean by 2040. Massachusetts, New Jersey, and Washington carry parallel statutes. Gas-dependent AI load structurally violates these mandates. The collision is codified in statute, not discretionary regulatory judgment.

Three reinforcing axes (regulatory permissiveness, grid physics, resource geology) align on the same set of states. ERCOT's November 2025 large-load queue stood at 226 gigawatts, 77% of which is data center (Latitude Media; ERCOT November 2025 CDR). Applying ERCOT's own attested-request haircut of 49.8% to 55.4%, the realized 2030 Texas data center load arrives at 38 to 43 gigawatts. CAISO's median project sits in queue 5.5 years before withdrawal (LBNL Queued Up 2025). PJM's 2027/28 capacity auction added 5,250 megawatts of forecast peak load nearly all attributable to data center demand, but the queue physics constrains delivery.

ERCOT CFO Pablo Vegas, on the May 2025 Capacity, Demand and Reserves report explaining the haircut: "We can't plan a grid on every announcement that crosses our desk. We have to assume some of these don't show up." That sentence, and the institutional discipline behind it, is why Texas builds and other states do not.

State-level classification is the right analytical resolution despite sub-state heterogeneity, because state authority dominates sub-state policy fragmentation in both directions. In Texas, State Senator Paul Bettencourt's February 2026 letter explicitly warned counties they have no constitutional authority to impose development moratoriums; Hood County's vetoed moratorium and Hays County's stalled attempt are the operative examples. In California, Imperial Valley's 600-megawatt geothermal pipeline targeting data centers and CalEthos's 300-megawatt wholesale campus represent genuine sub-state pockets of permissiveness. But the Sierra Club's legal challenge, Imperial Valley city's CEQA lawsuit (December 2025), and the November 2026 Imperial County ballot initiative confirm that state-level mandates and the activist coalition reach into local-permissive corridors. The fracture between permissive and restrictive states holds because the state level authority almost always supercedes local politics, regardless of which direction sub-state pockets push.

Federal preemption will not reverse this. The Trump executive order of December 2025 preempts state AI model regulation. It explicitly carves out state authority over data center infrastructure. The carve-out is not accident. It reflects the doctrinal reality that environmental regulation is traditional state police power, protected by the strongest federalism shield in US constitutional law (Virginia Uranium v. Warren, 2019). Senator Cotton's DATA Act creates a new utility category for off-grid hyperscaler power. Even that does not preempt the state environmental statutes that bind on-grid AI load.

The historical pattern is familiar. From 1970 to 1995, US manufacturing relocated from the unionized, regulated, high-cost Northeast and Midwest to the right-to-work, permissive, low-cost South. Three structural conditions matched: state-level cost arbitrage on the binding input (labor then; electricity now), aggressive state industrial policy (BMW in South Carolina, Nissan in Tennessee; Meta in Louisiana, Oracle in Texas), and federal neutrality that permitted subnational specialization. The compute equivalent runs five to seven years instead of thirty, because hyperscaler capex velocity exceeds plant capex velocity by an order of magnitude.

Three structural reallocations are emerging as major drivers of AI infrastructure investment in the U.S. over the next decade.

First, the power sector valuation regime fragments along ISO lines. The decade-long mental model (regulated utilities deliver bond-like base growth, merchant IPPs deliver commodity volatility, renewables PPA developers deliver project finance returns) does not survive state-level fracture intact. ERCOT-zone merchant power rerates as the dominant data center load absorber; the institutional moat sits with operators holding first-mover access to ERCOT's new interconnection processes and SB 6 compliance pathways. PJM-zone regulated utilities face a binary regime resolution on FERC's June 2026 large-load co-location ruling. Favorable rules permit rate-base capture from data center growth; restrictive rules cap base growth and force structural multiple compression. CAISO-zone and NYISO-zone regulated utilities face a physics-locked growth ceiling regardless of regulatory direction. The queue takes 5.5 years; the climate mandates apply uniformly. The portfolio question that follows is geographic before it is company-specific. Allocators with diversified power sector exposure will need to rebuild positions by ISO before they rebuild by issuer.

Second, a new asset class forms between regulated utility, IPP, and corporate technology infrastructure. The proposed Consumer-Regulated Electric Utility category under the Cotton DATA Act creates federally-recognized off-grid power infrastructure serving captive AI loads. If enacted, it sits in capital structure terms between merchant power, regulated utility, and corporate tech infrastructure. There is no historical comparable for hyperscaler-financed, geographically concentrated, off-grid gas-fired utility serving sub-second compute load. Insurance underwriting frameworks, debt rating methodologies, equity sector classifications, all need to be built. The 18 to 36 months during which existing taxonomies systematically miscategorize this segment is the window during which early-recognition allocators capture rerating from miscategorized peripheral to investment-grade infrastructure subclass.

Third, risk repricing across credit, insurance, and municipal markets. Three under-priced vectors emerge. Infrastructure debt spreads diverge by regulatory regime; investment-grade spreads between AI-exposed permissive-state generation debt and AI-exposed restrictive-state utility debt likely widen 50 to 150 basis points by 2028, where current pricing treats the universe as homogeneous. Specialty reinsurance and catastrophe markets lack actuarial frameworks for hyperscaler-owned behind-the-meter gas concentrated in drought, heat, and hurricane corridors (Permian, Gulf Coast Louisiana, Bakken). The first major casualty event triggers a step-function repricing the premium-flow opportunity precedes. Blue-state municipal utility debt faces compound credit pressure: rate-shift politics, climate-mandate liability, and base-erosion combined.

Aroko analysis forecasts that by end-2028, 55-65% of net-new hyperscaler AI capacity concentrates in five permissive energy-rich states (Texas, Louisiana, North Dakota, Iowa, Nebraska), with 13-20% absorbed by five restrictive states (Virginia, California, New York, New Jersey, Oregon) and the majority of that restrictive share representing grandfathered pipeline working through interconnection rather than new permit activity. Federal preemption succeeds on AI model regulation but fails to override state environmental authority on data center infrastructure. The thesis fails only if permissive top-five share falls below 50% AND federal preemption successfully overrides state DC infrastructure authority through enacted statute upheld by the Supreme Court. All else is sensitivity, not structural failure.

Two events in 2026 will determine where US AI compute lands for the decade. FERC's June ruling on PJM data center co-location sets the regime for the largest constrained market. The November midterms determine whether federal preemption appetite hardens or stalls. Neither event is binary in current capital allocation models; both are.

Underneath, the structural force is older and deeper. Shale gas spent a decade waiting for a long-dated demand vector after coal-to-gas saturation. AI is it. The geographic fracture is the physical expression of that codependence: gas-rich permissive states absorb, climate-mandate restrictive states cap. Federal preemption succeeds on AI model regulation but fails on data center infrastructure, because state environmental police power is the strongest federalism shield in US constitutional law. The fracture is structural — not cyclical.

If you're evaluating AI infrastructure exposure and want this analysis applied to your specific context, reply to this email.

About Aroko: Aroko provides strategic advisory and capital allocation intelligence at the intersection of energy transition, technology infrastructure, and geopolitical risk. Our analytical process combines proprietary evidence infrastructure with human-directed thesis formation. Every keystone claim is verified against primary sources, and all editorial judgment and capital allocation framing is conducted by Aroko’s team. The Letter is published biweekly for institutional allocators.

Disclaimer: This article is for informational purposes only and does not constitute financial, investment, legal, or tax advice. The opinions expressed regarding macro trends and infrastructure investments are solely those of the authors. Past performance does not guarantee future results. Readers should consult with a qualified financial professional before making any investment decisions.