Somewhere in Northern Virginia right now, inside the most power-dense data center corridor in the Western Hemisphere, an Nvidia GPU cluster is drawing over 100 megawatts of electricity. The utilities serving that corridor upgraded their substations, expanded their transmission capacity, and absorbed years of infrastructure cost to make that draw possible. The monthly bill for those upgrades goes to the families in Loudoun County who heat their homes on the same grid. As of June 13, 2026, the Federal Energy Regulatory Commission is nearing a ruling that could end that arrangement.
What Actually Happened
FERC docket RM26-4-000 has been proceeding since October 2025, when Energy Secretary Wright invoked a rarely used authority under the Department of Energy Organization Act to direct FERC to consider new rules for how large electrical loads, defined as demand exceeding 20 megawatts, connect to the interstate transmission system. The directive was prompted by a crisis that had been building for two years: AI data centers were connecting to the US grid faster than the grid's interconnection and cost-allocation rules were designed to handle, and the cost of the resulting infrastructure upgrades was flowing almost entirely to existing ratepayers rather than to the data centers driving the demand. According to reporting from YourNews, FERC is expected to issue its ruling before the end of June 2026.
The docket presents two fundamentally different visions for how massive new loads should integrate with the US power grid. The first, called the integrated grid model, would continue current practice: data centers connect through the standard grid infrastructure, and the cost of required upgrades gets socialized across all ratepayers in the affected utility territory. This is the model that produced the 833 percent increase in PJM capacity prices between the 2024-25 and 2025-26 delivery years as AI data center demand overwhelmed grid operator projections. The second model, called direct connection or the large-load rider model, would require data centers to directly fund all transmission infrastructure necessary for their interconnection, paying the full cost of substations, transmission upgrades, and grid capacity investments rather than spreading those costs across household and commercial ratepayers. According to Newsmax coverage of the proceedings, FERC appears likely to adopt a hybrid approach that shifts a substantial portion of connection costs toward large loads while preserving some degree of grid socialization.
The numbers driving the urgency are stark. PJM Interconnection, the grid operator serving 65 million people across 13 states and Washington, D.C., projects summer peak demand growth from 160 gigawatts in 2025 to 253 gigawatts by 2046, a 58 percent increase driven primarily by data center construction. Data centers now represent approximately 40 percent of the $16.4 billion in costs allocated through PJM's most recent capacity auction. According to Power Magazine, households in the data center corridor regions of Virginia, Texas, and Georgia are already experiencing rate increases of 8 to 15 percent above the national average, with further increases expected as additional data center capacity comes online in 2026 and 2027. The June FERC ruling will determine whether those increases accelerate, moderate, or reverse.
Why This Matters More Than People Think
The FERC ruling is not primarily about fairness to ratepayers, even though that framing dominates the political debate. It is primarily about the cost structure of AI compute in the United States. Power is not a minor input in AI training and inference economics, it is increasingly the dominant variable. A 100-megawatt data center running continuously costs approximately $70 million per year in electricity at current commercial rates. That figure doubles if the facility runs AI training workloads at high utilization, because training loads are power-intensive in ways that standard data center benchmarks don't capture. If the FERC ruling shifts substantial infrastructure costs directly onto data centers, the amortized cost of electricity per unit of AI compute in the US rises, tightening margins for every AI company that doesn't own captive power generation. The ruling is, in practice, a decision about whether America's AI compute is going to remain cost-competitive with its international alternatives.
The largest hyperscalers, Amazon Web Services, Google Cloud, Microsoft Azure, Meta, and Oracle, anticipated this outcome. In March 2026, all five companies, along with OpenAI and xAI, signed what became known as the Ratepayer Protection Pledge: a White House-facilitated voluntary commitment to directly fund all necessary grid infrastructure improvements required by their data center expansions. The pledge was widely described at the time as a preemptive move to shape the terms of the FERC ruling by demonstrating industry willingness to pay, making a mandatory cost-allocation regime look redundant. The question the June ruling will answer is whether FERC considers that voluntary commitment sufficient or whether it codifies mandatory cost allocation in ways that extend beyond the seven companies that signed. The hundreds of smaller AI compute providers, cloud infrastructure companies, and colocation operators that did not sign the pledge are watching closely, because the ruling will determine their compliance obligations and infrastructure cost exposure regardless of their participation in voluntary commitments.
The consumer impact is already visible in the data. The 833 percent increase in PJM capacity prices between the 2024-25 and 2025-26 delivery years represents a cost that flows through utility rate structures to residential and commercial customers over an 18-to-24-month lag. The families and small businesses in Virginia's data center corridor who are currently experiencing 8-to-15 percent rate increases are experiencing the pass-through from capacity auction prices set before the most recent wave of AI data center construction was fully reflected in grid planning. The next round of capacity pricing, informed by the accelerating 2025-2026 data center pipeline, will be set in late 2026, and without a FERC ruling that allocates costs differently, the pass-through to ratepayers will be larger than what they're currently experiencing.
The Competitive Landscape
The Ratepayer Protection Pledge signatories, the largest seven hyperscalers, have structurally different interests in this ruling than the broader AI compute ecosystem. Amazon, Google, Microsoft, and Meta have the balance sheets to fund direct grid connection costs, and they are large enough to negotiate long-term power purchase agreements and preferred interconnection terms directly with utilities and grid operators. For them, mandatory cost allocation is an operational expense that compresses margins but does not threaten viability. For smaller AI compute companies, the GPU cloud providers, the specialized training clusters, the mid-market inference operators, mandatory direct cost allocation is potentially existential. A ruling that requires every new data center above 20 megawatts to fund its own transmission infrastructure would create a capital barrier to entry that effectively consolidates the US AI compute market around the seven companies large enough to absorb the cost. This is not a coincidental outcome. The hyperscalers' voluntary pledge positioned them as willing partners in the cost allocation discussion; the ruling will determine whether that positioning produces regulatory capture or genuine market-wide cost fairness.
The nuclear and small modular reactor deal wave that has swept through the tech industry since 2024 is partly motivated by a calculation about FERC authority that nobody discusses in press releases. When Microsoft restarts Three Mile Island, or when Meta signs a 6.6-gigawatt nuclear power purchase agreement, or when Google commits to Kairos Power's SMR fleet, these companies are not primarily solving a clean energy problem. They are solving a grid interconnection problem. A data center powered directly by an on-site nuclear plant or a contracted-but-collocated generation facility is not connecting to the interstate transmission grid in the same way as a data center that draws from the general utility distribution network. The specifics of FERC jurisdiction over these direct generation-to-load arrangements are genuinely complex, and the June ruling may clarify or complicate them. If the ruling extends cost allocation requirements to direct generation connections as well as grid connections, the nuclear deal economics change. If it doesn't, nuclear arrangements become even more attractive as a regulatory arbitrage strategy.
The international competitive dimension deserves attention. China is building AI data center capacity at a comparable pace to the United States, but within a power grid that is state-owned and where cost allocation to AI infrastructure is a policy decision made by central planners rather than a regulatory ruling subject to industry lobbying. Chinese AI compute does not face a Ratepayer Protection Pledge moment because the state makes infrastructure investment decisions without the market-pricing intermediation that produces the tension FERC is now trying to resolve. The EU's approach varies by member state, but several major data center markets, Ireland, the Netherlands, Germany, are seeing local electricity price increases from AI data center concentration that parallel the US pattern. The FERC ruling will be watched carefully by European energy regulators as a template for addressing the same challenge, and its terms will influence how AI compute cost structures evolve globally, not just in the United States.
Hidden Insight: Power Is the New Compute Moat
The AI industry's cost structure conversation focuses obsessively on chips: the price of H100s, the wait time for Blackwell clusters, the economics of Trainium versus GPU alternatives. This is a reasonable focus, chips are the most visible, most discussed, and most politically contentious input cost. But power is increasingly the binding constraint, and unlike chip availability, power pricing is determined by regulatory decisions that compound over time rather than by semiconductor market cycles that clear. A hyperscaler that locks in a 20-year nuclear power purchase agreement at 3.5 cents per kilowatt-hour is building an infrastructure cost advantage that no competitor can replicate through chip procurement optimization. The FERC ruling is a decision point that will determine whether that advantage grows or narrows over the next decade.
The relationship between AI compute costs and AI model pricing is direct and underappreciated. The dramatic decline in LLM API pricing over 2024-2025, a period when GPT-4-class inference costs fell by more than 90 percent, was driven by a combination of chip efficiency improvements and infrastructure scale. Power costs during that period were largely socialized through utility rate structures that hadn't yet fully absorbed the data center demand surge. If the FERC ruling shifts substantial power infrastructure costs onto data centers in 2026, the 2027 cost curve for AI inference will be shaped partly by that regulatory decision rather than purely by chip innovation. The companies that anticipated this, that locked in power purchase agreements, signed nuclear deals, and built on-site generation capacity, will have structural cost advantages that persist for years. The companies that didn't will face higher infrastructure costs at precisely the moment when AI inference markets are becoming more competitive and margin-compressive.
The broader hidden dynamic is what this ruling signals about the long-term geography of AI compute. If the US implements stringent cost allocation requirements that increase the expense of connecting large AI data centers to the US grid, the marginal incentive to build the next gigawatt of AI compute in the US weakens relative to jurisdictions with cheaper or more favorable grid connection terms. The data center industry is highly mobile over 3-to-5-year planning cycles, not every facility can move, but new construction decisions respond to regulatory and cost environments. Countries with favorable power pricing, grid connection terms, and renewable energy availability are already competing for AI compute investment. A FERC ruling that significantly raises US grid connection costs will not drive existing data centers out of the country, but it will affect where the next generation of capacity gets built, and that matters for US AI competitiveness over the next decade.
However, critics argue that the direct cost allocation model creates its own set of perverse incentives that could undermine the grid reliability benefits it's designed to achieve. The bear case for the ruling is straightforward: if large loads must fund all their transmission infrastructure independently, they have a strong financial incentive to underestimate their capacity needs, propose smaller initial interconnections, and then seek incremental expansions that exploit gaps in regulatory review processes. The grid operators who testified during the RM26-4-000 rulemaking process flagged exactly this pattern as a risk. A poorly designed cost allocation rule could produce a fragmented, poorly planned grid with data centers optimizing their individual connection costs rather than participating in the coordinated transmission planning that keeps the overall grid reliable. National Law Review coverage of the proceeding noted that grid operators and utilities were among the most vocal advocates for cost allocation frameworks that preserve integrated planning incentives rather than purely individual cost allocation.
What to Watch Next
The FERC ruling itself is the first milestone, expected before July 1. The specific cost allocation percentage, how much of transmission infrastructure cost shifts to large loads versus remaining socialized, will determine the magnitude of the impact on AI compute economics. Watch for the ruling's treatment of existing contracts and legacy connections: a ruling that applies only to new connections from a prospective effective date is dramatically less disruptive than one with retroactive implications for facilities already in service. The ruling's treatment of direct generation arrangements, whether nuclear and SMR-connected facilities count as "grid connections" subject to the new cost allocation rules, will determine the regulatory premium on the nuclear deals that the major hyperscalers have been announcing since 2024.
In the 90-day window following the ruling, watch for litigation. Any ruling that significantly shifts cost allocation toward large loads will be challenged in federal court by the data center industry. The legal question will center on whether FERC has exceeded its statutory authority under the Federal Power Act by effectively mandating a new pricing structure for large customers that is more appropriately a rate case question for individual utilities. If a court issues a stay pending appeal, the ruling's practical effect could be delayed by 12 to 18 months while the litigation proceeds. The hyperscaler signatories of the Ratepayer Protection Pledge face a strategic decision: a court challenge that succeeds preserves their voluntary-only commitment, but it also signals bad faith to the administration and the public in ways that invite more prescriptive legislation.
At the 180-day mark, watch the Q3 and Q4 2026 earnings calls from Amazon, Google, Microsoft, and Meta. Any FERC ruling with meaningful cost allocation requirements will appear in revised capex guidance as companies update their data center construction cost models. If hyperscalers increase their AI infrastructure spending estimates citing higher regulatory infrastructure costs, that's a direct signal that the ruling produced real cost impacts rather than the paper exercise that many in the industry hoped it would be. If capex guidance holds steady, it will likely mean either that the ruling was weaker than anticipated, or that the nuclear and direct generation deals already struck have effectively exempted the largest players from the ruling's direct cost exposure, leaving the burden concentrated on the mid-market operators who couldn't afford those hedges.
The question isn't who pays for the grid. It's who controls the cost structure of American AI compute for the next twenty years.
Key Takeaways
- FERC ruling expected by end of June, Docket RM26-4-000 will determine how costs for transmission infrastructure driven by AI data center growth are allocated, with options ranging from full socialization to mandatory direct cost assignment for loads above 20 megawatts.
- Data centers represent 40% of PJM capacity costs, AI data center demand now accounts for approximately 40 percent of the $16.4 billion in costs allocated through PJM's most recent capacity auction, with PJM projecting 58 percent total demand growth by 2046.
- Consumer bills already up 8-15%, Households in Northern Virginia, parts of Texas, and Georgia data center corridors are already experiencing electricity rate increases of 8 to 15 percent above the national average as grid capacity costs pass through utility rate structures.
- Nuclear deals are partly a regulatory hedge, The wave of nuclear and SMR power purchase agreements signed by hyperscalers since 2024 is motivated partly by avoiding grid interconnection costs and FERC jurisdiction over direct generation-to-load arrangements.
- Ratepayer Protection Pledge was preemptive positioning, The March 2026 voluntary pledge by Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI to fund grid infrastructure was designed to shape the terms of the FERC ruling by demonstrating willingness to pay before mandatory requirements were set.
Questions Worth Asking
- If the FERC ruling effectively raises the cost of connecting AI data centers to the US grid, will the next generation of AI compute infrastructure shift to jurisdictions with cheaper grid connection terms, and what does that mean for US AI competitiveness over the next decade?
- The Ratepayer Protection Pledge was voluntary and signed by only seven companies, does a FERC ruling that codifies mandatory cost allocation actually solve the fairness problem, or does it create a two-tier system where large hyperscalers can absorb the costs while smaller AI compute operators cannot?
- If nuclear and direct generation arrangements exempted from the ruling's cost allocation requirements create a regulatory moat for hyperscalers who already signed power deals, has the ruling produced a competitive market outcome or just shifted the barrier to entry from chip access to power infrastructure?
