The Grid Bill Comes Due

Position: A state-by-state patchwork of "who-pays" rules, not a federal standard, now prices where large AI loads can afford to land, and the gap between the most- and least-expensive grids widens through 2027 rather than closing. The cost sovereignty has moved from the permit to the bill.

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

For two years, a power generation deficit dictated where data centers are sited in the U.S. While that story remains relevant, especially in gas-rich states like Louisiana and Texas where fuel supply is abundant and land is relatively cheap, the regime is beginning to shift decisively towards the question of who ultimately foots the bill for the grid capacity that delivers the power to a data center's door. In the last sixty days, that question has been answered eleven different ways.

What started as a national imperative to lead the AI race has devolved into a state-by-state recalibration around the economic well-being of households, in response to mounting public opposition to rising electricity bills.

On June 18th, rather than supersede state authority with a national standard, FERC directed each of the six grid operators under its jurisdiction to justify or rewrite its own rules across five issue areas, cost-shift prevention among them. In essence, FERC delegated the rulemaking to six system operators that each start from divergent rules. And the grid that matters most to the spread sits beyond its reach entirely: ERCOT, in Texas, is not FERC-jurisdictional, so the cheapest pole in the country is one no federal standard could ever have compressed. The grid bill is coming due, and each operator is now pursuing a different answer.

Through 2025, the cost of upgrading the grid to serve a hyperscaler, including the new wires and substations and transformers between a power plant and a server hall, was mostly socialized. It went into the rate base and spread across every electricity bill in the territory. The data center got the power; the retiree down the road helped pay to deliver it. As recently as March, the federal answer to this was a voluntary pledge: seven hyperscalers promised to fund their own generation and delivery. The industry, by one account, greeted it with a shrug.

The politics had already turned. In the largest US power market, a capacity auction cleared 833 percent higher than the year before, and the cost landed on consumers as a bill roughly 76 percent higher. A federal watchdog's response was not to build faster. It was to tell the data centers to pay their own way.

Then, in about two months, the voluntary intimation hardened into law, state by state — eleven different ways in all. New York's legislature passed a statewide pause on any data center drawing 20 megawatts or more. Five states wrote ratepayer-protection laws, each through a different legal instrument, and more followed. Separately, at the federal layer, FERC reopened the narrower question of who pays when a data center co-locates with its own power plant. None of it was coordinated. Each measure moved the cost of delivering power off the shared bill and onto the load that caused it, by its own method, on its own schedule.

This shift matters more than any individual law. Because every jurisdiction is now architecting its own approach, the all-in price of plugging a large campus into the grid varies by location. That reality, more than streamlined permitting or available power, now decides where AI can afford to land.

In the initial AI data center ramp, Virginia and Texas arguably beat every other state on speed; both said "yes" while other states were still studying options. Texas remains highly attractive as far as data center siting goes compared to most states, but starting at the end of this year, a new large load in Texas pays $50,000 per megawatt just to get in the interconnection queue. This fee is non-refundable with an additional $100,000 study fee.

A cheap permit and a cheap grid are no longer the same thing. The state that is easiest to build in now hands you a bill at the door.

Across other jurisdictions, proposed laws travel under the same banner: protect the ratepayer. Below the banner, each instrument diverges significantly in approach.

Eleven jurisdictions have now answered the same question: who pays to connect a large load. The list below summarizes each approach, ordered from the states that put the most on the data center to the states that put the least. They run ten distinct instruments; only Florida and Minnesota land on the same one.

The spread inside a single phrase, "ratepayer protection," runs from Pennsylvania's full cause-and-pay to Louisiana's half-and-socialize to New York's you-cannot-build. That is genuine fragmentation, and it is hardening just as inference demand accelerates.

Regulators have put data centers on the highest-severity alert level and moved to register them as formal grid actors. The market operators are writing their own large-load rules; one of them, MISO, took care to write into its tariff that those rules do not supersede state authority. And when the federal opening came on June 18th to do so, FERC declined to take it. Rather than set one national who-pays standard, it ordered each of the six grids under its jurisdiction to justify or rewrite its own large-load tariff and left cost allocation with the operators. ERCOT, and the cheap Texas floor it runs, is not subject to the order.

How each state charges for grid capacity upgrades is a direct read on where AI infrastructure can afford to concentrate. Assume a developer proposes one 100 megawatt campus. At 2025 construction costs of $10.7 million per megawatt, that is about $1.07 billion worth of silicon and steel before it draws a watt.

At the relatively inexpensive end of the spectrum, Texas imposes a one-time $50,000 per megawatt fee, which is about $5 million for the campus (roughly half a percent of build cost) paid at the front-end and done. At the opposite end of the spectrum (relatively expensive), full-pierce states like Pennsylvania and Virginia impose an ongoing liability.

"But-for" upgrades, referring to grid infrastructure required for specific new customers, land on the load in full. Take-or-pay minimum commits new large load customers to pay for 80 to 85 percent of contracted demand for twelve to fourteen years whether it draws the power or not. In present-value terms, on a representative demand charge and discount rate, that commitment runs to something on the order of $850,000 per megawatt, or roughly eight percent of build cost, carried on the balance sheet for the life of the contract.

In other words, under current rules, the grid bill for an identical campus runs from a hard floor near 0.5% of build cost — Texas's $50,000-per-megawatt fee over the build, simple arithmetic — to a modeled ceiling near 8% in full-pierce states, a present-value figure that rides on a representative demand charge and discount rate. One end is arithmetic; the other is a careful estimate. And the inexpensive end is the one Washington cannot move: Texas sets the floor and sits outside FERC's June 18th order. A converging federal standard could still press more expensive grids down toward that floor — that is the live risk to this call — but it would be pressing against divergent starting tariffs. By contrast, it could never pull the floor up.

Also worth noting, some instruments pierce and some grandfather. Pennsylvania's "but-for" rule, Florida's and Minnesota's full cost-of-service assignment, and the Ohio, Virginia, and Kansas rate-class minimums all impose fees on the entire new large-load class. Meanwhile, Texas' fee applies only to loads connecting after December 31, 2026. Michigan explicitly protects existing customers. This concentrates the whole repricing on new siting decisions, which is exactly where the capital is being committed today. In other words, the past is not prologue for predicting where the AI infrastructure buildout will concentrate in the next five years.

We have run this experiment once already. In the late 1990s, the U.S. tried to restructure power markets, and it did so the same way it is pricing data centers now: state by state, with no two following the same playbook. Some states broke up their utilities and opened the market to retail competition; others kept the old vertically integrated model; a few started, panicked after California's crisis, and reversed. The patchwork that resulted, not the price of power, largely decided where merchant generators and heavy industrial loads were sited over the next decade.

The restructuring map of 1999 is more or less still the map today: states that deregulated then are still the predominant merchant markets now, and states that did not, remain more insulated. The lesson is not that fragmentation is good or bad. It is that fragmentation, once it sets, is sticky, and the stakeholders who priced the patchwork early were positioned for twenty years.

That is the primary takeaway from this letter: the grid bill fractures in similar ways and, more importantly, stays fractured long enough to matter. And this time the stickiness is not just analogy — it is written into the instruments themselves. The take-or-pay minimums run twelve to fifteen years; the grandfather lines protect installed bases for the same period. A state that prices its grid today locks that price into contracts that outlast the next several FERC chairs. The patchwork sets because the paperwork makes it expensive to reverse.

The counter argument to this thesis is worth consideration as well. The objection is not that the grid bill isn't real. It is that the bill might not impact the decision calculus for siting data centers.

Three credible voices make the case. S&P Global and Goldman argue that what actually drives where data centers go is power availability and time-to-power, not cost, and they expect only marginal tightening in investment. Carnegie argues the 2026-to-2030 power cliff will not be bridged by firm low-carbon supply, which means demand stays desperate enough to swallow almost any premium. And the industry itself, by way of Axios, disputes the premise that data centers raise consumer bills at all. Put those together and the grid bill is a tax the developer eats wherever it lands, not a force that sorts where they land. If that proves correct, then the jurisdictional patchwork remains an insignificant variable.

What is worth watching is whether the desperate marginal megawatt actually relocates jurisdiction — forum-shopping for the most favorable economics rather than simply chasing available power. This would cut against this letter: if loads forum-shop hard enough toward the inexpensive grids, demand concentrates at the floor and the effective spread compresses in practice, even as the statutes on paper keep diverging. A spread that no one acts on would still weaken the call.

There is a second objection, which the June 18th FERC decision sharpens. Six show-cause orders, which flips the burden to system operators and binds them together, is precisely how a de facto national standard gets built. One compliance filing at a time, even without a single national rule on the books. But held against the structure of what FERC actually did, the de facto standard argument weakens. The six system operators are all starting from divergent tariffs, each justifying their own rule rather than adopting a common one, and the cheapest grid in the country, ERCOT, is not subject to the order at all. A common rubric applied to divergent starting points, with the lowest denominator exempt, tends to produce divergent filings, not convergence. The kill-path is real, but it runs uphill.

Through December 31, 2027, Aroko expects more fragmentation, not a federal standard, to remain the variable that prices where large loads can afford to land in the U.S. The gap between the most- and least-expensive grids is expected to widen, not close. Each state continues using its own instrument (aligned with their respective system operator rules); each grandfathers its own installed base its own way; and Washington keeps standardizing process and reliability while leaving the "who-pays?" question to the states.

This call rests on two legs. On the first, if FERC's review of the six June 18th filings, or its still-open RM26-4 rulemaking, produces a binding national cost-allocation standard the major ISOs actually adopt in place of their own instruments, then the thesis expires. On the other, if the filings harmonize the jurisdictional grids' all-in cost toward one another by the end of 2027, the gap closes from the top even if no national standard is ever named. Texas sitting outside the order protects the floor, not the spread.

Given the opening on June 18th, FERC chose six separate show-cause orders over one national rule. FERC's first move on the live question was to hand it back to the operators, not claim it. A regulator bent on converging the grids had the cleaner path in front of it and passed. Washington usually does, eventually, standardize a market this politically hot; but with the pricing now formally delegated to operators who each start from a different position, and the cheapest grid in the country outside FERC's reach entirely, the spread between the least and most expensive state is more likely to widen before it closes.

What does the grid bill tell us about where we are in the Grid-Silicon Order, the coupling of power and compute that now governs the entire AI buildout? Plotted on two axes: 1) geopolitical architecture runs from Integrated to Fragmented (whether the rules and the capital cross borders or split into separate blocs); and 2) infrastructure velocity runs from Accelerated to Constrained (whether the buildout keeps moving or hits a hard limit). In the context of the grid bill, the fragmentation in question is not among countries, but among states. Same axis, finer grain.

The grid bill thesis is best aligned with Sovereign Stacks: the buildout is still racing, but the rules that price it have fractured state by state. Aroko's bet is essentially that we stay here through 2027 and do not snap back to Open Abundance. The more dangerous drift, over the next three to five years, is toward Fortress Era: if New York's moratorium spreads and velocity constrains, fragmentation plus scarcity is where this goes. That drift would not break the frame. It would deepen it, because exclusion is just the most extreme bill a state can send.

The permit told you where you could build. The bill tells you whether you can afford to. For the next two years, the map that prices U.S. compute is not the one with the cheapest power; it is the one with the cheapest grid. Price the bill, not the megawatt.

Then watch the layer underneath the bill. Everything here is denominated in dollars and state statutes. The next repricing is in atoms: the transformers, the high-voltage converters, the copper and steel, and the export licenses that decide whether the grid can be built at any price, in any state. That bill is partly written in Beijing.

The U.S. grid was built to socialize the cost of long-lived generation across a broad base of ratepayers. Hyperscaler load — concentrated, around-the-clock, arriving on a two-year clock — is the opposite of what that system was built to absorb, and a cost structure built to be shared cannot stay shared once a single class of customer doubles the bill. The next eighteen months settles most of this, but underneath the signals is a century-old arrangement coming apart. Once a 76 percent spike lands on voters, the politics finish what the economics started. And no institution can put the question back together: state commissions can assign cost but not past their own borders, FERC has the interstate reach but is standardizing process, not cost, and the White House had only a voluntary pledge. The who-pays question fragments precisely because it sits in the one place no single regulator's mandate reaches. That is not a temporary gap. It is the structure.

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.