FERC June Ruling Changes Who Pays for AI Grid Costs

The question has been building for three years: when an AI data center demands enough power to supply 750,000 homes, who pays to upgrade the transmission lines that carry it? By the end of June 2026, the Federal Energy Regulatory Commission will answer that question in a ruling that could reshape the economics of every AI lab, hyperscaler, and electricity customer in the United States. The decision affects more than power bills. It determines whether the next generation of AI infrastructure gets built on a level playing field or an incumbents-only one.

What Actually Happened

On April 16, 2026, FERC announced it would act on Docket No. RM26-4-000 by the end of June 2026, a landmark rulemaking on how large electrical loads, defined as any facility drawing more than 20 megawatts, connect to the interstate transmission system. The commission reviewed more than 3,500 pages of public comments submitted by utilities, technology companies, consumer advocates, and state regulators. As reporting from June 14 confirmed, the ruling is imminent and will directly address whether AI data centers must bear the full cost of grid upgrades they require, or whether those costs flow to existing ratepayers instead.

The rulemaking traces to October 2025, when Energy Secretary Chris Wright invoked a rarely used authority under the DOE Organization Act to direct FERC to accelerate new rules for large-load interconnection. The directive reflected a growing recognition that the standard generator interconnection queue, which was designed for power plants not power consumers, was failing under the scale of AI-era demand growth. FERC confirmed its June timeline in subsequent filings, noting the proceeding covers everything from study process design to cost allocation principles to co-location rules for data centers that connect directly behind a power plant's meter. Southwest Power Pool already launched its HILL (High Impact Large Load) program as a regional pilot, and PJM has 811 projects in its reformed queue cycle with 220 GW of capacity waiting for connection.

The numbers that frame this decision are stark. PJM Interconnection, the grid operator serving 65 million people across 13 states and Washington D.C., reported that data centers now account for 40 percent of the $16.4 billion in costs from its most recent capacity auction, according to Engineering News-Record. PJM's own demand forecast projects summer peak load rising from 160 gigawatts in 2025 to 253 gigawatts by 2046, a 58-percent increase driven almost entirely by the AI data center buildout. A single rack of next-generation AI chips now approaches one megawatt of power draw, enough to supply around 750 average American homes. The scale mismatch between AI compute demand and residential power infrastructure is no longer theoretical. It is already showing up in capacity auction prices.

Why This Matters More Than People Think

The immediate financial stakes divide cleanly along two scenarios. If FERC requires data centers to bear the full cost of grid upgrades their interconnection demands, hyperscale construction costs rise by an estimated 15 to 25 percent per project in high-demand regions. That cost increase translates to delayed deployment timelines of 12 to 18 months in congested grid zones, meaning the compute capacity currently on order from Nvidia, AMD, and custom silicon suppliers would sit without power infrastructure to run it. If FERC socializes interconnection upgrade costs across all ratepayers instead, every household in PJM territory faces a structural electricity bill increase, with consumer advocates estimating the impact at $8 to $14 per month per household in the highest-demand regions by 2030.

The long-term stakes are more consequential than the short-term bill math. PJM's forecast requires adding the equivalent of roughly 40 nuclear plants worth of peak generation and transmission capacity over the next two decades, in a grid where new transmission lines take an average of 10 years to permit and build. The AI industry, with its 18-month GPU refresh cycles and quarterly capacity commitments to cloud customers, is operating on a timeline the grid cannot match through organic development. FERC's ruling will determine how that mismatch is resolved: either by forcing data centers to pay the true cost of acceleration, or by socializing that acceleration cost across all grid users and effectively using ratepayer dollars to subsidize infrastructure that primarily benefits hyperscalers. Neither outcome is cost-free. They differ only in who bears the cost and who captures the benefit.

The ruling will also reshape competitive dynamics within the AI industry in ways most analysts have not yet modeled. A ruling that assigns full interconnection costs to large loads benefits entrenched hyperscalers like Google, Amazon, and Microsoft, which can absorb higher upfront costs that smaller AI labs and cloud competitors cannot. Google, Amazon, and Microsoft each filed extensive comments in the RM26-4-000 docket arguing for transparent co-location rules and accelerated study processes, according to Power Magazine. Their preferred outcome is precisely the regime that raises barriers to entry for the next tier of AI infrastructure builders. The ruling that looks like a win for consumers may also be the ruling that cements hyperscaler dominance over AI compute for the next decade.

The Competitive Landscape

The companies most directly affected by this ruling are the same ones spending the most to shape it. Google, Amazon, and Microsoft are all mid-cycle in multi-gigawatt data center buildouts that depend on grid interconnections currently sitting in PJM's 811-project queue. For Amazon, the decision intersects with its $25 billion Anthropic investment, which is predicated on securing up to 5 gigawatts of compute capacity. For Microsoft, a ruling that delays grid connection by 18 months would push back the timeline for MAI model training runs embedded in Microsoft Build 2026 roadmap commitments. Both companies have co-location agreements with nuclear plant operators that may need to be restructured depending on how FERC defines large-load cost responsibility.

The nuclear co-location deals already signed hang directly in the regulatory balance. Microsoft's agreement with Constellation Energy to restart Three Mile Island Unit 1, and Amazon's nuclear purchase agreement with Dominion Energy for the North Anna plant, were structured as behind-the-meter arrangements that bypass traditional transmission charge allocation. A FERC ruling that treats co-location as equivalent to grid-connected load would unravel the economic basis of those deals. Conversely, a ruling that affirms behind-the-meter co-location as exempt from traditional interconnection cost-sharing would accelerate the race to secure dedicated nuclear capacity, potentially pulling capital away from transmission-connected solar and wind projects that serve residential ratepayers.

The historical parallel is FERC Order 636, the 1992 deregulation of interstate natural gas pipelines that opened them to third-party shippers. That ruling triggered a decade of infrastructure buildout and competitive fragmentation. Critics argue the AI grid ruling risks repeating that pattern: separating AI-grade power delivery from residential service into two distinct markets creates efficiency gains in normal conditions but introduces cascading failure risks when demand spikes exceed either subsystem's capacity. The California energy crisis of 2000 to 2001, which followed deregulation of the electricity market, demonstrated how quickly theoretical market efficiency can become practical grid failure when physical constraints arrive faster than market signals can adjust. However, supporters of cost-causation pricing argue the alternative, subsidizing hyperscale compute expansion through residential electricity bills, is also a risk transfer, just less visible and less politically accountable.

Hidden Insight: Co-Location Is the Real Prize in This Ruling

The billing debate about who pays for grid upgrades is the visible contest. The deeper battle inside the FERC docket is about co-location: data centers connecting directly behind a power plant's meter rather than through the transmission grid. A co-located data center exits the traditional grid entirely. It does not pay transmission charges, does not compete with households in the capacity auction, and does not wait in PJM's interconnection queue. For a hyperscaler running a gigawatt of AI compute continuously, the long-run savings over 20 years at co-location rates versus grid rates are measured in billions of dollars. The FERC ruling will determine whether that exit ramp stays open, gets widened, or gets closed.

The implications for AI model pricing are direct and underappreciated. Training a frontier model at the scale of GPT-5 or Anthropic's Fable 5 consumes somewhere between 50 and 200 gigawatt-hours of electricity per training run, based on published estimates from infrastructure analysts. If the cost of that electricity rises by 15 percent for non-co-located compute because FERC assigns full interconnection costs to large loads, the economics of frontier model training shift by 10 to 15 percent in favor of co-located compute. Labs with co-location deals at nuclear or gas peaker plants will train models at lower effective cost than labs dependent on grid-connected cloud capacity. That cost gap, compounded over multiple training runs per year, creates a structural competitive advantage that has nothing to do with algorithmic innovation and everything to do with regulatory positioning secured years earlier.

The data center industry is already bifurcating in anticipation of this ruling. Tier-one hyperscalers are securing co-location agreements with power generators at a pace smaller competitors cannot match. A wave of nuclear purchase agreements, direct gas peaker interconnections, and on-site generation deals has been signed since 2024, creating a class of AI infrastructure that is essentially grid-independent. If the FERC ruling affirms co-location as the preferred model for large loads, the co-location race will accelerate. New entrants to AI infrastructure, lacking the balance sheet to secure dedicated generation capacity, will face a structurally higher power cost than incumbents, regardless of chip architecture or software efficiency gains.

The bear case deserves careful attention. FERC may attempt a compromise ruling that neither fully socializes costs nor fully assigns them to data centers, producing regulatory ambiguity that state regulators will fill with conflicting local rules. California's Public Utilities Commission, New York's PSC, and Illinois regulators all have active proceedings that would create a patchwork of interconnection standards across the largest AI markets. Texas, with its ERCOT grid entirely outside FERC's transmission jurisdiction, serves as an uncontrolled experiment: if AI investment concentrates in Texas after a restrictive federal ruling, it confirms that regulatory arbitrage drives compute geography more than fiber routes, water availability, or labor costs. A fragmented national market would impose the highest long-run cost on all parties, according to analysis from hdata research on the FERC rulemaking. The worst outcome is not a ruling that helps one side. It is a ruling too ambiguous to guide either.

What to Watch Next

The ruling itself is the first milestone, expected by June 30, 2026. Within 60 days, PJM and SPP must file compliance plans addressing how they will implement any new standards. Watch specifically whether the ruling treats co-location separately from traditional grid interconnection. That single distinction will determine whether the nuclear co-location deals already signed get regulatory green lights or face new cost-sharing requirements that alter their economics. The language around "behind-the-meter" versus "transmission-connected" load in the ruling text will be the most consequential 50 words in American energy infrastructure in a decade.

The PJM capacity auction scheduled for Q3 2026 will be the first major market signal of how participants are pricing the new interconnection framework. Capacity prices reflect generators' and load-serving entities' expectations about grid expansion costs. A sharp rise would confirm the ruling raised the cost of new large-load connections; stable prices suggest the market absorbed the ruling without major disruption. Watch also for new data center project announcements in the 90 days following the ruling. A surge of new permits in Texas and ERCOT territory would signal that developers are routing around the federal framework rather than working within it, which would confirm the patchwork outcome rather than the national standard the ruling is meant to create.

The 180-day indicator is whether Congress responds. Bipartisan bills in the 119th Congress would override or supplement FERC's authority on large-load interconnection, including proposals to create a federal data center grid designation and proposals to require hyperscalers to post performance bonds for interconnection upgrades. A ruling that forces electricity cost increases of $8 to $14 per month for residential ratepayers in PJM territory would create immediate political pressure from the 40-plus senators representing PJM states. The AI infrastructure lobby and the consumer utility lobby are equally well-funded and organized. If the ruling produces a clear winner, expect the loser to begin drafting legislation within 30 days and to have committee hearings scheduled within 90.

The grid is not a given. For three years, the AI industry treated electricity as an input cost. The FERC June ruling will determine whether it becomes a competitive moat.

---