Essay
The grid was the policy
The country wrote an industrial policy for compute. What it bought was a queue.
There is a question that has begun to settle, in the last eighteen months, over the people whose work is to deliver the country its electricity. It has not yet had its name. We expect it will be the central question of American energy planning for the next decade, and that it will outlast several of the louder questions currently in fashion.
Begin at a workstation in the planning offices of a regional transmission organization in Audubon, Pennsylvania, after seven on a weekday evening. The planner is at her desk. She has been at PJM, or at the utilities that operate inside PJM, for twenty-two years. Her colleagues know her caution; the engineers who model load against her forecasts have learned, over time, to bring their assumptions with them. There are two documents in front of her tonight. She has set them out the way she sets out anything she does not yet know how to read together — one on the left, one on the right, the keyboard pushed back to make room.
The document on the left is a memo from the Department of Commerce, eighteen months old now, asking PJM to “ensure timely interconnection” of CHIPS Act fab projects sited in its territory. The memo is, in its plain language, about a category of work. In the practical life of her office it is about one project — Intel’s Ohio One campus in New Albany, the marquee semiconductor fab the country announced inside PJM’s footprint, the one originally targeting production in 2025 and now, after a succession of public schedule slips, targeting 2030 or 2031.
The document on the right is the PJM Long-Term Load Forecast, paired — clipped to it, in fact, by a black binder clip she has reused for years — with Dominion Energy Virginia’s large-load delivery queue. The queue is a list of data centers waiting for service. It runs, at present, to approximately seventy gigawatts of pending load in a single utility’s territory. Dominion’s all-time system peak, set in January of the prior year, was 24.7 gigawatts.
She does not, at first, do what a planner does. She does not match a fab list against an interconnection queue. She has done that exercise; the exercise is in the binder behind her. What she finds herself doing tonight is something else. She is matching the policy the memo on the left describes against the load the queue on the right contains. The memo describes an industrial policy for semiconductors — a fab program, scored in wafer-starts and ribbon-cuttings, with a workforce annex for cleanroom technicians. The queue describes what that policy, executed across the eighteen months since the memo was written, has in practice become. It has become a data-center load program, in a single load-serving utility’s territory, roughly three times the size of the system serving it.
Her pen, an old felt-tip she keeps for marking up printouts, is on the queue. It has not moved. The discipline that has carried her through twenty years of forecasts — the discipline that says you redline the document in front of you, you do not redline the country — fails her tonight. She does not mark up the queue. She sets the pen down. She notices, with the small administrative clarity that comes after a long day, that the memo on the left does not appear to know the queue on the right exists.
She gets up. She makes coffee in the office kitchen, slowly, because moving slowly is what she does when she has registered something she does not yet know how to write down. She comes back to her desk. She does not sit. She stands behind her chair for a moment and looks at the two documents from above — the memo, the queue, the binder clip — the way one might look at two photographs that turn out to have been taken of the same room.
The question does not yet have a name. But the question that follows her home tonight is the question that will follow many of us, regardless of which side of the meter we sit on: what, exactly, did we just buy?
We were not buying compute. Or rather: we were buying it, and we were also buying the other thing, the one that almost nobody on the CHIPS scoring panel ever quite said out loud. The bill scored what could be scored. It counted wafer-starts and cleanroom technicians and the square footage of fabs. It counted ribbon-cuttings. It counted the headline dollar figure of the manufacturing incentive — $39 billion of direct funding, $13.7 billion of R&D and workforce — and it counted the press releases that followed each finalized award through the fall of 2024. What it did not count, because the panel had no column for it, was the institutional capacity the fabs would in fact require to operate: siting authority for the high-voltage lines that would feed them, firm transmission service to a specific substation by a specific date, the journeymen who would string the conductors at 500 kV, the regulatory horizon long enough to deliver any of it. None of these were absent from the awards because somebody on the panel had argued against them. They were absent because nothing in the program’s machinery was built to register them. The bill was, in this sense, a careful instrument pointed at the wrong unit.
That is the trick of the appropriation. It made it look as though compute was what we were buying. It was not. Compute was the unit, the thing announced, scored, photographed at ribbon-cutting, summed across the four big awards into a $25 billion line in a chart no congressional staffer ever had to defend in a hearing. The grid was the product — the thing the unit was, in fact, a proxy for. Firm transmission service to a specific bus by a specific date is what a fab requires to run. The fab was a proxy, and like every proxy, it worked until something arrived that broke the relationship between the proxy and the thing.
This is the distinction that the data-center queue is now making impossible to hide.
For most of the postwar history of American industrial policy, the capital expenditure and the institutional capacity to use it ran together. We could not have one without the other. The 1956 Federal-Aid Highway Act — Pub. L. 84-627, signed by Eisenhower on June 29 of that year — is the archetype. The bill authorized ~41,000 miles of what it called the National System of Interstate and Defense Highways. The number, on its own, would have been a number. What made it a program was what came welded to it. The Act created the Highway Trust Fund, capitalized by federal user taxes on gasoline, diesel, tires, and heavy-vehicle use, which meant the program had its own balance sheet rather than competing for general appropriations every year of the next quarter-century. It set a 90/10 federal-state cost-share that gave every state highway department a stable financial relationship with the federal road. And it carried the federal eminent-domain framework under which the right-of-way could in fact be assembled — a volume of land acquisition, in the end, greater than had been taken in the entire prior history of American road-building. The roads and the institutions shipped in the same box. The Eisenhower-era state did not just appropriate money for highways. It built the delivery vehicle that could spend the money. Trust fund. Cost share. Takings authority. Corridor map. The four pieces were drawn together on the same desk because nobody at the desk imagined they could be separated. The map could not be built without the takings; the takings could not be financed without the trust fund; the trust fund could not be defended in the next Congress without a cost share that gave fifty governors a reason to defend it. Each piece reinforced the others. The bill was, in this strict sense, an industrial policy: a corridor designed end to end, from the gas pump that paid for it to the right-of-way that received the pavement.
The 1936 Rural Electrification Act is the other archetype, and the more telling one for our purposes, because what the REA built was not a road but a load. The Act — Pub. L. 74-605 — authorized the Secretary of Agriculture, working through the Rural Electrification Administration, to issue low-interest federal loans, on the order of three percent and amortized over twenty-five years, to a category of borrower that had not, before the program, really existed at the scale required: the member-owned rural electric cooperative. The investor-owned utilities of the 1920s and early 1930s had declined to serve the countryside; the economics, by their lights, did not work. The REA’s move was not to subsidize the existing utilities into reluctant service. It was to design, capitalize, and pattern-stamp a new delivery vehicle. The agency wrote model bylaws for the cooperatives. It set construction standards and engineering practices for their distribution lines. It lent against a balance sheet it had, in effect, helped the borrower assemble. The capital and the institution were welded together at the point of issuance. The agency did not finance a build. It built a delivery vehicle that could be financed. The result, by 1950, was a country in which rural electrification had risen from roughly ten percent of farms to roughly ninety — not because the REA had built the lines, but because it had built the entity that could.
The 2024 CHIPS framework broke that welding. The capital expenditure shipped. The institutional capacity did not. What we shipped, in the announcements of November and December, was the appropriation. What we did not ship was the four pieces of statecraft an actual industrial policy for compute would have had to include.
The break can be checked. It is a matter of arithmetic, and the arithmetic can be done at a kitchen table in an afternoon.
Here is the test. Take the list of CHIPS Act fab awards finalized in 2024. Match each award against the public interconnection queue at the regional transmission organization in whose territory the fab is being built. Ask, for each project, whether the firm transmission service it needs to operate at its announced capacity exists today, is scheduled to exist by the fab’s announced production date, or is on a study list with no firm date.
The list is short. Four awards were finalized in the fall of 2024: TSMC Arizona at $6.6 billion (November 15); Intel at $7.86 billion (November 26); Micron at $6.165 billion (December 10); Samsung Taylor at $4.745 billion (December 20). The sum, ~$25.4 billion in direct funding, is the line the program’s defenders will quote when they say the year delivered. Of the four, exactly one sits inside the territory of the regional transmission organization most exposed to the data-center buildout the same period produced: Intel’s Ohio One campus in New Albany. TSMC Phoenix is WECC. Micron’s Clay fab is NYISO. Samsung Taylor is ERCOT. The single marquee CHIPS fab inside PJM is Intel Ohio.
And Intel Ohio has slipped. The 2022 announcement targeted production in 2025. In February 2024 the date moved to 2026. By late 2024 and early 2025 Intel had announced the first fab would not complete until 2030, with production in 2030 or 2031, and the second fab pushed to 2031 or 2032. The schedule has slipped into and through the very ten-year window a major high-voltage line takes to build. The marquee fab inside PJM is now on a delivery curve indistinguishable from the curve of the transmission service it would need.
The larger story sits next to it on the same page. Dominion Energy Virginia, the load-serving utility under which most of PJM’s data-center growth in fact gets delivered, reports a large-load delivery queue of roughly seventy gigawatts of pending interconnection requests. Dominion’s all-time system peak, set in January of the prior year, was 24.7 gigawatts. The 2024 PJM Long-Term Load Forecast projects roughly thirty-two gigawatts of peak load growth across the footprint from 2024 through 2030, of which roughly thirty is data-center driven. Wait times for new large-load delivery in Dominion’s territory have stretched to as long as seven years. The country wrote an industrial policy whose announced unit was a $25 billion fab program, and in the same window stood up, inside a single utility’s queue and with no statutory instrument touching it, a data-center load program multiples larger.
The exercise takes an afternoon. The answer is that the policy bought capital expenditure for a project the country has not yet decided to deliver power to.
The obvious objection is that the queue is structural. Every generator and every large load in PJM, and in every other RTO, waits years for interconnection; CHIPS fabs are not uniquely disadvantaged. This is correct, and it is also the point. The policy was written as though the queue did not exist. A program that buys capital expenditure on a class of project whose binding constraint is the very queue the program declined to address is not an industrial policy in the postwar sense. It is a procurement that has confused the announcement with the build.
And the queue is itself an institutional fact the industrial policy was supposed to address. FERC Order No. 2023, issued in July 2023, replaced the old serial first-come-first-served interconnection study with cluster studies on annual cycles. Cluster reform was a real improvement on the bookkeeping. It does not, in itself, build transmission. It only sequences who waits. The binding number is the build cycle behind the queue: a major high-voltage line in the United States routinely takes ten years from conception to energization, and the gold-standard long-haul projects take longer. SunZia: roughly seventeen years of planning before groundbreaking, then two to three of construction. TransWest Express: roughly fifteen years of federal review before notice to proceed, then four to seven of construction. These are the ordinary horizon of the institutional fact CHIPS does not touch.
It is worth being precise about what, in the postwar sense, an industrial policy for compute would have had to contain. The four awards of November and December 2024 are the appropriation. They are not, on their own, the policy. The policy — the corridor, in the language of the 1956 Act — is the set of institutional instruments without which the appropriation cannot in fact be spent on the thing it announced. The country had, at the moment the awards were finalized, none of them. Each is load-bearing in a way the bill’s drafters could in principle have known and in practice did not write. Four pieces of statecraft. Each load-bearing. None in the bill.
The first is siting and eminent-domain authority for transmission. The Federal Power Act framework places transmission siting with state public utility commissions, not with FERC. The federal government had, before 2005, no general backstop authority over transmission siting at all; the lines that crossed three states were permitted, route by route, by three separate state commissions, each answerable to its own legislature and its own ratepayers. The Energy Policy Act of 2005 added Section 216 to the FPA in an attempt to introduce a narrow federal backstop: the Department of Energy was directed to designate National Interest Electric Transmission Corridors, and inside those corridors FERC was to receive backstop permitting authority, including federal eminent domain, in cases where a state had “withheld approval for more than 1 year” on a siting application. FERC moved to implement the new authority in Order 689 (2006), interpreting “withheld approval” to include outright denials. The Fourth Circuit, in Piedmont Environmental Council v. FERC, 558 F.3d 304 (4th Cir. 2009), reversed. The court held that the plain meaning of the word “withheld” does not include the word “denied” — a state that denies a permit has not withheld it; it has acted on it. After Piedmont, FERC never again issued a §216 permit. Two years later the Ninth Circuit, in California Wilderness Coalition v. DOE, 631 F.3d 1072 (9th Cir. 2011), vacated the underlying DOE NIETC corridor designations themselves on consultation and NEPA grounds; after 2011, DOE designated no further corridors. IIJA §40105, in 2021, attempted to patch the Piedmont hole — explicitly authorizing FERC backstop when a state denies an application, fails to rule within a year, or imposes conditions rendering a project uneconomic — and FERC finalized its revised procedures in 2024. As of mid-2026, zero §216 backstop permits have been issued under the new authority. The practical implication, for the policy at hand, is the one the planner in Audubon already knows: a 500 kV line that crosses three states is not a federal project even when the buildout it serves is a federal program. The CHIPS framework owns the fab. It does not own the line the fab requires. Nothing in the bill changes that, and nothing in the bill could have changed it without an instrument the bill did not contain.
The second is a workforce pipeline for the trades that build substations and string the conductors. The IBEW transmission-line journeyman pipeline — the pipeline that supplies the linemen who in fact assemble, energize, and maintain high-voltage transmission — is in a severe shortage period. Industry estimates suggest roughly forty-five percent of experienced linemen are projected to retire within the next decade. Training programs run applicant-to-seat ratios on the order of ten to one at top schools; roughly two hundred lineworker training programs operate nationwide; five or six new programs launch in a typical year. Union apprenticeship and community-college enrollment has grown on the order of twenty percent over the last five years — real growth, and still well behind both the retirements and the new demand the compute buildout has imposed on the trade. A 500 kV journeyman certification typically requires on the order of seven thousand hours of supervised work, or roughly three and a half to four years of apprenticeship; this is a number the schedule cannot be rushed past without consequences the work itself will deliver. The CHIPS Act’s workforce funding scope is, by design, almost entirely outside this pipeline. The bill funds semiconductor-industry technicians and operators — cleanroom technicians, process engineers, advanced packaging specialists — the workforce a finished fab needs to run. It does not fund the construction trades that build the fab, and it does not fund the line trades that build the substations and the conductors that feed it. CHIPS-funded fab construction is, separately, subject to Davis-Bacon prevailing wage on the federally-supported portions of the build; this raises the cost of trades labor in the affected projects, but it does not, in itself, deepen the pipeline by a single apprentice. The implication is the one the schedule already shows: even where the transmission line is sited, approved, and financed, the trade depth to build it on the announced calendar may not exist. The Act funds the fab-operator pipeline. The buildout’s binding constraint depends on a different one.
The third is a clearing instrument between the Department of Commerce and the regional transmission organizations. There is, at present, no institution at which Commerce’s award decisions and an RTO’s queue can be reconciled. Commerce does not have the authority to demand interconnection on behalf of a CHIPS awardee; the RTO does not have the authority to refuse, on policy grounds, capacity to a project carrying federal backing. Neither has the institutional capacity, in the ordinary course, to acknowledge that the other exists. The closest historical analogue would be a Defense Production Act–style coordinating body operating in the electricity sector — a body authorized to translate a federal industrial-policy priority into a binding instruction inside the queue. The United States does not have one. The DPA, when invoked, applies to specific items — most famously ventilators in 2020, and pharmaceutical precursors in the years since — and it has not been written into a regular instrument of grid coordination. The structural reason coordination cannot simply be “ordered” is one most readers outside the energy sector do not appreciate but a practitioner recognizes at once: the RTO is itself a private, FERC-regulated entity, not a federal agency. There is no chain of command between the Department of Commerce and PJM. Commerce sends an award letter; PJM sends back a queue number; nothing institutional sits between them to adjudicate the gap. The fab announcement and the interconnection decision happen, in a strict sense, in different governments. The memo on the planner’s desk in Audubon was an attempt — one of the few public ones — to bridge that gap with a request. A request was the strongest instrument the writer had.
The fourth is a regulatory horizon longer than the next administration. Major U.S. high-voltage projects take, on average, about ten years from conception to operation; the gold-standard cross-state HVDC builds take longer. SunZia Southwest Transmission — 550 miles, 550 kV HVDC — ran from a 2006 conception through a 2023 BLM Record of Decision and groundbreaking to commercial operation expected in 2025 or 2026: roughly seventeen years of planning before a shovel turned. TransWest Express — 732 miles, 500 kV HVDC — ran from a 2009 plan of development through a 2016 BLM Record of Decision and a 2017 WAPA Record of Decision to a 2023 groundbreaking, with operation expected between 2027 and 2030: roughly fifteen years of federal review before notice to proceed. The political horizon on any given industrial policy is two to six years — one administration, or one and a half. The mismatch is structural and pre-dates AI. The compute buildout is the genre’s most legible recent demonstration of it: a federal program scored on a four-year political horizon is being asked to deliver capital projects whose delivery vehicle takes three to five times that horizon. DOE’s CITAP rule (2024) aims to compress the federal permitting component from roughly four years to a binding two-year window — a real improvement, on the federal portion. It does not touch state siting; it does not touch easements; it does not touch construction itself. No instrument inside CHIPS lengthens the horizon. No instrument inside CHIPS shortens the vehicle. The Act is asked to deliver inside a horizon shorter than its own delivery vehicle’s, and it is not given the tools to close the gap from either side.
What the four pieces share is the thing the bill could most easily have noticed and least easily have written. They are all institutional, not capital. The country could have written any of them. It wrote none. AI Affairs is a publication of diagnosis: whether any of the four pieces can be written, in this political horizon, by these institutions, is a separate question from whether their absence is the actual story of the 2024 industrial policy.
There is a reading of the last twenty-four months in which all of this is going well. It is the reading that has shaped most of the trade press coverage and most of the investor calls. The announcements have continued. The ribbon-cuttings have continued. The press cycle that surrounds each finalized award has, by its own metrics, performed: the photographs were taken, the governors were quoted, the figures were summed into a national line, and the line was, at the margin, defensible. A class of professional advocates for the buildout has, in good faith, treated this output as evidence that an industrial policy is in motion. They are not lying. They are reading what is in front of them. What is in front of them is the procurement file — the awards, the cost shares, the negotiated milestones, the public-facing schedule. The procurement file is a real document. It is not the policy. We are reading a procurement transcript and calling it Hamilton. The country that wrote the 1956 Highway Act wrote a corridor; the country that wrote the 1936 REA wrote a delivery vehicle; the country that wrote the 2022 CHIPS Act wrote, instead, a procurement, and the people who are now describing the procurement as a policy are performing, in real time, the substitution the bill performed at the level of statute. They believe they are auditing an industrial policy. They are reading aloud from a contract.
There is no compute industrial policy, properly understood. There is a procurement office, a press cycle, and an interconnection queue. The first two are ours. The third is the one writing back. The state PUC siting docket, the IBEW local’s apprentice intake list, the RTO’s cluster study, the load-serving utility’s seven-year wait — these are the parties holding the pen on what the country is in fact building. None of them was invited to the announcement. None of them was scored. They are nonetheless, between them, writing the industrial policy of the compute decade. They did not know they were the authors of it when they began. Several of them have only recently realized it. The policy is being written on a calendar none of them controls, in a register that none of the policy’s announced authors yet reads.
The planner, by now, has gone home. Tomorrow she will come back to the same desk; the memo will still be on the left, the queue will still be on the right, the binder clip will still be holding the forecast to the Dominion intake. She will, at some point in the morning, send a short email to a colleague at another RTO who is reading a different queue against a different memo and arriving at a version of the same arithmetic. She will not mark up either document. She has decided, without quite phrasing it to herself, that the document she would in fact be marking up is not on her desk. It is not on anybody’s desk. It is the queue itself, and the queue is being marked up, line by line, by the parties that submitted to it — by the data-center developers and the load-serving utility and the transmission planner herself, each entering, each accepting, each scheduling, none of them aware that the cumulative effect of their entries is the actual industrial policy of the United States in the compute decade.
She turns off the lamp. The queue software is still open on her monitor, the cursor still blinking on the row she did not edit. She closes the laptop. The hallway lights are on a timer; they will go off at eleven. She takes the elevator down to the parking lot in Audubon and drives home on the same road she has driven for twenty-two years. The memo and the queue will be where she left them. So will the country.
What we wrote was a memo. What the country built was a queue. The industrial policy of the compute decade was, in the end, the queue itself — written line by line by parties who did not know they were the authors of it, on a calendar none of its announced authors controls, in a register the announcement was not equipped to read.
§