In the 1880s an English landowner could read two documents about his own life and reach opposite conclusions. One was the rent roll. The other was the invitation list.

The rent roll was failing. American wheat had been arriving in volume since the previous decade, carried by railroads that turned the interior of a continent into arable land and by steamships that made the crossing cheap. British grain prices fell and did not recover. Arable rents followed them down. By the time a Royal Commission was appointed to study agricultural depression in 1893, the depression it was appointed to study was twenty years old [1].

The other functions had already gone, quietly and earlier. The purchase of army commissions, which had made the officer corps something close to landed property, was abolished in 1871 [2]. Competitive examination for the civil service, recommended in 1854 and extended by Order in Council in 1870, replaced patronage with a test that the sons of clergymen and merchants could pass [3]. Estate duty arrived in 1894 [4]. The Lords lost their absolute veto in 1911 [5].

And yet the invitation list held. His name still carried a county, still moved a marriage market, still made an opinion worth soliciting, still made a subscription list look serious. Country houses filled with guests through the entire decline. Historians who have counted the land transfers put the real break after 1918, roughly four decades after the economics had gone [6].

That interval is the thing worth staring at. The function ended in the 1870s. The position ended in the 1920s. The actual history was messier than those two sentences, because the class kept real assets that the grain price never touched: London ground rents, mineral royalties under the coalfields, City directorships, and marriages into American money. Some of the standing was hollow and some of it was still bought and paid for, and from inside there was no reliable way to tell which was which. That is the condition worth naming, and it is not confined to landowners.

Almost none of them experienced it as a decline. The income kept arriving, smaller than before and later than it should have, but on schedule. The invitations kept coming. A man is badly placed to notice that he is being invited out of habit.

The same interval has opened under the credentialed intellectual

What a competent language model produces today is not a parlor trick and is not a first draft. It is a high-fluency average of the published conversation on almost any subject: the settled position and the standard objections to it, cross-disciplinary synthesis that would have taken a person a week in a library, a policy memo that reads like a policy memo, an imitation of almost any established style, and a serviceable guess at how a field will react to an event before the field reacts.

Before this, possessing that fluency was what marked a person out. It is now close to free.

What models reveal is more awkward than their own performance. Their arrival exposes what was actually being paid for. Much of what protected the ordinary intellectual was never exceptional thought. It was three things wearing the costume of one: information scarcity, distribution, and credentialing. The professor had read the literature. The columnist could assemble an argument fast. The think tank researcher could produce something that looked credible to a person who could not evaluate it. All three of those advantages have been compressed, and the compression happened in public, on everyone’s own screen, in about three years.

I have argued elsewhere that the credential is a futures contract on trainability, still clearing the gate while the ladder behind it is withdrawn. That was the entrant’s problem. This is the incumbent’s, and it is the stranger of the two, because the incumbent’s numbers still look fine.

The standing rent

What survives the compression is a standing rent: the share of an intellectual’s income and influence that comes not from the difficulty of the thought but from the institutional right to have that thought counted. It is a rent because it is charged for position rather than production, and because it can be collected by someone who has stopped producing.

The standing rent is not new and is not illegitimate on its face. It is the thing that makes a signature mean something. A journal’s acceptance, a chair’s endorsement, a named professorship attached to a report, a former official’s byline on an op-ed: each of these does work that the argument alone cannot do. It renders a claim admissible to institutions that have no independent way to evaluate it. Someone with standing can certify a claim, introduce its author, recommend a grant, hire a researcher, or make a proposal acceptable to a ministry. The identical paragraph produced without standing does none of that and carries no liability for anyone.

One property of the standing rent is what makes the aristocratic comparison worth more than a flourish. As the underlying thought commoditizes, the rent does not fall. It rises as a share of the total, because it is the part that is left. And to the person collecting it, a rising share of a shrinking base feels exactly like security. The invitations are still coming. The opinion is still solicited. The pay has not moved much. Only the composition of the pay has changed, and composition is invisible from inside.

The aristocrat’s last decades were his most ceremonial and he mistook the ceremony for continued standing. That mistake was correct for one generation and fatal for the next.

What is actually scarce

Outside expertise still has a market. What no longer pays on its own is the ability to produce intelligent commentary, and the capacities that do remain scarce are mostly not the ones the profession selects for.

Producing new evidence is scarce. Running the experiment, building the dataset, getting the interview, reading the primary document that nobody has opened. Models are already leaking into parts of this, into instrument operation, dataset construction, and document discovery, so the scarcity is not a capability ceiling. It is permission, capital, physical presence, and liability for being wrong in the world, and none of those gets cheaper when text does.

Deciding which objection survives is scarce. A model will generate twenty objections to any framework on request, and this is close to worthless, because the difficulty was never generating objections. It was knowing which one is real. A critic may contribute no original theory of his own and still perform the most valuable act available, which is establishing that a beautiful theory is false.

Bearing responsibility is scarce in a way the other two are not, because it is not a cognitive capacity at all. Making a call, attaching a name to it, and accepting the consequence when it is wrong is a relationship rather than an output, and nothing about model capability touches it.

Then there is constituency work. Describing a group is commodity labor now. Convening one, holding its trust, and speaking for it in a room where something gets decided is not, and it never was an intellectual skill in the first place, which is why so few people trained for it.

The scarcest thing of all is a calibrated public record: not a list of predictions but an auditable one, with the misses still visible.

Not one of those is what the profession selects for. Most of the credentialed intellectual class was trained, hired, promoted, and rewarded for work of a high and genuinely difficult kind that a machine can now approximate, and the pipeline producing more of them has not noticed.

Why the lag will be long

Certification is the slowest thing in any institution to break, because it is the last thing that is actually load-bearing. Everything else an institution does can be outsourced, digitized, or quietly abandoned. The authority to say that a claim counts cannot be, without the institution ceasing to be one.

Peer review, tenure, editorial gatekeeping, expert witness qualification, advisory committee membership, and the ministerial reading list all run on positions rather than arguments. Peer review is designed to evaluate a claim partly independently of who made it, and sometimes it does. What none of these can do is rely on that method exclusively, at the volume and speed at which institutions have to decide, because attribution is what allows a body to defer without being blamed for deferring. That is not corruption. It is how a body that cannot evaluate a technical claim manages to act on one anyway.

So the lag is structural, and it will run for a long time. There are two ways out of it. One is to be bought: frontier labs are hiring economists, lawyers, and institutional scholars directly, which raises the value of standing for the few who are hired and quietly lowers it for everyone else, because the certification now happens inside a firm rather than through a field. The other is to move into the scarce categories on your own, which is harder, slower, and far less open than it appears. Most people in the lag will do neither.

Long enough to be a career. Not long enough to be a life.

The objection at full strength

The strongest reply is that the standing rent is not a rent at all. It is payment for a real service that becomes more valuable, not less, as generated text becomes abundant.

The argument runs like this. Institutions need accountable human judgment. A model has no license to revoke, no reputation to lose, no career to end, no liability to attach. When a hospital, a regulator, or a court acts on a claim, it needs a person who can be questioned about it, and the supply of people who can credibly be questioned does not expand when text does. Abundance of argument makes the certifying function more scarce relative to demand, not less. On this reading, the intellectual is not the aristocrat at all. He is the auditor, and audit gets more valuable as the books get longer.

This is largely right, and it is why the lag is long rather than short. But it does not rescue the position, for one reason. The service being described is liability, not thought. If the compensation is for standing behind a claim rather than for producing it, then the training, the selection, the credential, and the self-understanding of the entire profession are all aimed at the wrong thing. A field can survive discovering that its product has changed. It survives much less well when the discovery arrives one cohort at a time, in the form of individual careers that quietly fail to start.

The place to watch is entry rather than pay. Pay is the last thing to move, and the aristocracy shows why: the original economic foundation had been failing for decades while the houses stayed full. If the professions that trade on standing keep letting new people reach certifying positions at anything like historical rates, the comparison collapses, because the defining feature of the aristocratic pattern is a position that stops being enterable before it stops being valuable. I expect the narrowing instead, and I expect it on three different clocks. A chair is a stock, a byline is a flow, and an advisory seat is a political appointment. New bylines and junior appointments should contract first, chairs can stay occupied for decades after the field beneath them has changed, and advisory positions will move discontinuously, whenever the politics that fills them changes.

The ending was not confiscation

The aristocracy did not end in a seizure. There was no morning when the titles were abolished and the houses taken. What happened instead was that the invitations continued while the reason for them expired, and then one generation discovered that the door which had always opened for its father opened for it only as a courtesy, and then not at all.

Nobody was expropriated. The last of them were treated well, consulted respectfully, and quietly not consulted about anything that was going to be decided.

What is arriving for the credentialed intellectual is not unemployment, and treating it as unemployment is why so much of the conversation misses it. It is ceremony: an increasingly honored position, increasingly detached from the production of anything that could not be produced otherwise, still paid, still solicited, still seated near the front. The title outlived the function by two generations. Nobody inside the lag ever calls it a lag.

Notes

[1] Royal Commission on Agricultural Depression, appointed 1893, Final Report 1897.

[2] Purchase of commissions in the British Army was abolished in 1871 under the Cardwell reforms, by royal warrant after the Lords declined to pass the enabling bill.

[3] Report on the Organisation of the Permanent Civil Service (Northcote-Trevelyan), 1854; open competition extended by Order in Council, 1870.

[4] Finance Act 1894, introducing estate duty on the principal value of property passing at death.

[5] Parliament Act 1911, replacing the House of Lords’ absolute veto with a delaying power.

[6] F. M. L. Thompson, English Landed Society in the Nineteenth Century (London: Routledge & Kegan Paul, 1963), on the scale and timing of land transfers after 1918. See also David Cannadine, The Decline and Fall of the British Aristocracy (New Haven: Yale University Press, 1990).

Set aside the horse race. As a document, that memo is a party telling an industry that the machinery of electoral persuasion cannot produce public consent for the industry’s core physical asset, and that the industry is on its own. It is an admission of failure filed upward, from the institution that exists to win arguments to the institution that needs one won. The interesting thing about it is not the panic. It is that the panic is entirely about a Senate seat.

Something has come apart between the fight and the decision. The public objection to data centers is loud, sincere, cross-partisan and electorally effective. In the venue where the terms of the buildout are actually set, that objection does not appear. Something resembling one piece of it appears instead, argued by different people on different grounds, and it has been winning.

This essay is about that gap, and about a second thing that is harder to see and worse. The win in the second room was not produced by the fight in the first one. It has no parent among the people it protects. That is why it will be difficult to keep.

The One Grievance This Is About

The objection to data centers is at least six objections: cost, water withdrawal, noise, land conversion, tax abatement, and a general distrust of the people building them. Only the first is at issue here. And within cost, only one component of it: the direct shifting of a specific project’s infrastructure cost onto other customer classes.

Not generation scarcity. Not capacity prices. Not fuel markets, where new gas plants built for data center load buy on the same market as every existing plant. Not a transmission expansion whose forecast turns out wrong. All of those move a household’s bill and none is addressed below.

That one component is the part of the objection a rate proceeding is built to hear, which is exactly why it is the part that got answered. Nothing here claims more.

The Room Doesn’t Need Your Anger

Every governing decision has a room, and every room has rules about what can be said in it that are prior to, and more determinative than, any argument made inside it.

A zoning board will hear that a project is too loud, too large, too close to a school, out of character with the county. It will not hear that the applicant’s industry is overcapitalized. A legislature will hear almost anything, which is part of why it decides so little. A rate proceeding before a state utility commission will hear that a cost has been misallocated between customer classes, that a forecast is unsupported, that a capital expenditure is not used and useful. It will not hear that the applicant should not exist. Standing, evidence rules, statutory mandate and the professional habits of the participants combine to decide what an objection has to become before it can be answered at all.

Call the filter venue translation. If a grievance is going to be heard as a grievance, it must be restated in the language of whichever institution holds authority over the decision, and the restatement is not a formality. It is the price of being answered, and it is charged before anyone weighs the merits.

Two consequences follow, and this essay needs both.

The first is the obvious one. A grievance that cannot make the translation is not heard at all, however loud it gets. That is what happened to the demand that the buildout stop.

The second is stranger and is the harder half of the story. A result favorable to a class of people can be produced in a room that none of them entered, by a participant pursuing an interest that happens to point the same way. Nothing was translated, because nothing was submitted. The outcome arrives regardless. That is what happened to the cost objection.

Neither observation is new on its own. Schattschneider argued sixty years ago that the definition of the alternatives is the supreme instrument of power, and that who wins depends on which conflict gets to be the conflict. Political scientists since have described venue shopping, in which an organized interest picks the forum most likely to give it what it wants; the lawyers’ version of the same filter is standing doctrine. Derrick Bell’s interest convergence thesis described the second pattern, in which a subordinated group’s gains arrive when they happen to serve a dominant group’s interests.

What follows differs from all of them in one respect. Those accounts have a claimant: someone choosing a room, or a group whose advancement is at stake and who knows it. Here there is neither. The people protected did not select the venue, did not appear in it, do not know it exists, and cannot tell whether anyone inside was arguing on their behalf or merely arriving at a result they happen to like.

Three properties make that possible.

The room is not chosen by the person with the grievance. It is fixed by the structure of the decision, long before anyone objects. Anger about a data center is one thing; whether it is heard as a land use question, a rate question or a political question is four different proceedings with four different sets of people in them, and only some of them decide anything.

Volume and effect come apart. A grievance can be extremely loud in a venue that decides nothing and completely silent in the venue that decides everything, and there is no mechanism that corrects this. Loudness does not migrate. A hundred people at a county meeting do not become a filing.

And the room’s occupants are whoever the room’s rules admit, which need not include anyone who shares the grievance. When a favorable result comes out of such a room, the household has not won an argument. It has received a byproduct.

The Fight You Can See Is Losing

Look at what the visible fight has produced, on its own terms.

Its loudest demand was that the buildout stop, and measured against that demand it has mostly lost. Fourteen state legislatures introduced bills restricting new data center construction, and as of July 2026 none had been signed into law. [2] Maine came closest: LD 307 would have paused projects above twenty megawatts until late 2027, passed both chambers, and was vetoed by Governor Janet Mills in April 2026 over its effect on a single redevelopment project at a shuttered paper mill in Jay. [3] New York’s legislature passed a twenty megawatt moratorium the governor has still not signed; instead, on July 14, Kathy Hochul issued Executive Order 62, pausing state environmental permits for facilities of fifty megawatts or more for up to a year while agencies write standards. [4] That is the first statewide moratorium in the country, and it arrived as an executive workaround around the legislature’s own bill.

Local action has done better. Seattle adopted a one-year moratorium on projects above twenty megawatts in June, the largest American city to do so. Monterey Park, California banned them outright by ballot measure with eighty-eight percent of the vote. Texas paused new approvals in August pending audits by its utility commission and grid operator. [5] But a local moratorium more often relocates a project than prevents one, and developers have generally been willing to relocate.

Stopping the buildout was never the only demand. Plenty of the same organizing asked for cost protection, water limits, noise standards or a community benefits agreement, and some of that succeeded locally. But the demand that carried nationally, the one the polling measures and the memo is frightened of, was the categorical one.

Measured as electoral pressure, the fight has done considerably better. On November 4, 2025, Democrats Alicia Johnson and Peter Hubbard unseated Republican incumbents Tim Echols and Fitz Johnson on the Georgia Public Service Commission, each by roughly fifty-nine to forty-one, after the commission approved six Georgia Power rate increases since 2023 that added an estimated five hundred dollars a year to the average household bill. They were the first Democrats to win a non-federal statewide race in Georgia in nearly two decades. [6]

The polling is not ambiguous about direction, though the instruments disagree about magnitude in a way worth noticing. A Fox News survey of registered voters conducted July 17–20, 2026 found seventy percent opposed to a data center being built in their area and thirty percent in favor, with nearly eight in ten preferring slower construction to speed. Gallup found seventy-one percent opposed in March, forty-eight percent strongly. A POLITICO poll conducted by Public First the same month as the Fox survey found forty-one percent would oppose a data center within three miles of their home, up from twenty-eight percent in January. [7] [8] [9] Two of those instruments ask nearly the same question in the same month and come back thirty points apart, which should make anyone careful about the exact number and confident about the trend.

So the anger is real and seats are really changing hands. What has not happened is the thing the anger asked for. Construction continues. Announced capacity keeps arriving. The visible fight has generated an enormous quantity of political consequence and a very small quantity of stopped concrete.

The Fight You Can’t See Is Winning

Now look at the venue nobody is watching.

A tariff, in American utility practice, is the filed schedule of rates and conditions under which a utility serves a class of customer. Your residential rate is a tariff. It is approved by a state commission in a proceeding with a docket number, a hearing examiner, utility counsel, a consumer advocate and whatever intervenors bother to appear. A large load tariff creates a new class for very big customers, usually defined by a megawatt threshold, and attaches conditions to membership in it.

As of July 2026, twenty-four states had approved at least one, with six more pending. [10] A separate tracker counting individual utility tariffs rather than states put it at fifty-one approved and twenty-six proposed across thirty-six states as of March. [11] The conditions are unglamorous and substantial. Minimum take commitments, so a customer pays for contracted capacity whether or not it draws power. Upfront capital contributions, so the transmission a project requires is funded by that project rather than folded into a rate base and recovered from everyone. Collateral against abandonment. Exit fees. Curtailment obligations that put the data center’s load ahead of the household’s in a grid emergency rather than behind it.

New Jersey’s version, signed in July 2026, is the clearest to state. It directs the Board of Public Utilities to define a large data center customer for each utility at a threshold no greater than fifty megawatts, aggregating commonly owned or contiguous facilities so a campus cannot be subdivided below the line, and requires financial guarantees that the customer pay for at least eighty-five percent of the service it requests for at least ten years, plus upfront deposits toward new transmission. [12] Virginia’s GS-5 tariff requires data centers above twenty-five megawatts to sign fourteen-year contracts and post collateral of $1.5 million per megawatt. [11] Virginia also kept its sales tax exemption and added a consumption tax of $0.011 per kilowatt-hour on data center electricity, capped at $600 million a year, effective July 1, 2026. [13]

None of this stops a data center. All of it bears on who carries the cost of one.

Three qualifications, all of which matter later. The designs vary widely across twenty-four states and some will not do much. An approved tariff is a rule, not a result: New Jersey’s statute sets standards and then gives utilities a hundred and eighty days to apply them, so the state that got the most coverage has enacted an instruction to build the thing rather than the thing. And several of these are temporary by construction. Virginia’s consumption tax expires June 30, 2028 unless the legislature extends it. [13]

Still: in about half the country, the household holding a utility bill has been answered on the one part of its objection a commission is competent to hear. It did not ask in that language. It mostly does not know the answer was given.

Nobody Did This for You

The natural reading of the last section is that public anger got translated. That it entered the rate proceedings as cost causation and won there in a form it would not recognize. That reading is attractive, and it is wrong.

The template is specific enough to trace. A minimum billing obligation around eighty-five percent, a long contract term, collateral, an exit fee, a megawatt threshold for membership in the class. It did not originate in 2026 and it did not originate with anybody’s constituents.

In March 2023, AEP Ohio stopped connecting new data centers in central Ohio, a moratorium it imposed on its own authority without asking the commission first. In May 2024 it applied to the Public Utilities Commission of Ohio for a dedicated data center tariff covering customers above twenty-five megawatts. The proceeding ran fourteen months and was contested; more than a dozen parties filed testimony and the hyperscalers opposed it. On July 9, 2025, the commission approved a stipulation joined by AEP Ohio, commission staff, the Ohio Consumers’ Counsel, the Ohio Energy Group, the Ohio Manufacturers’ Association Energy Group and Industrial Energy Users-Ohio. It required qualifying data centers to pay for at least eighty-five percent of subscribed load for up to twelve years, with creditworthiness requirements, collateral and exit fees. In the same order, the commission directed AEP Ohio to file the tariffs and to lift its moratorium and connect the load as soon as possible. [14]

That last clause tells you what the utility was after. Not to stop anything. To be paid for infrastructure it would otherwise be left holding if the forecast proved optimistic, and then to build. The commission’s stated ground was cost causation: costs should be borne by the customers who cause them, a principle it would have applied in an empty room, because applying it is the job.

Now compare dates. Ohio moratorium, March 2023. Application, May 2024. Order, July 2025. The Georgia commissioners lost in November 2025. The moratorium bills, the local bans, the polling collapse, the party memo and the New Jersey statute are all 2026. New Jersey’s eighty-five percent guarantee for ten years closely resembles the Ohio settlement, arriving two years downstream. That is similarity and sequence rather than proven descent, the same design could be arrived at independently by any utility facing the same stranded-asset problem, which is itself the point but the recurrence of the same package makes the resemblance worth noticing.

So the mechanism was not translation. Nobody restated the public’s objection in the commission’s language, and the commission did not hear a muffled version of the county meeting. A utility protecting its own balance sheet drafted a rule whose incidental effect was to protect households. A consumer counsel and some industrial users improved it at the margins, which is the closest thing to representation in the story and is real, though the Ohio Consumers’ Counsel is a statutory office, not a movement. A commission approved it on grounds internal to ratemaking. Then it spread, because a design that has survived a contested proceeding somewhere is the cheapest thing for a commission elsewhere to adopt, and because every utility faces the identical problem.

Be precise about what the chronology does and does not establish. It does not show that no public concern touched the Ohio proceeding; Ohio ratepayers were already unhappy, and the Consumers’ Counsel was in the room. What it shows is that the design preceded the national political wave by two years, and that the wave is now producing legislation which codifies rules the utilities had already written. The politics is downstream of the design.

This is a better outcome than most public fights get, and it should be said plainly: the protection is real, it is spreading, and households are better off for it. But it is not a victory in the sense the word usually carries. It is an alignment. For as long as the utility’s interest in avoiding stranded assets points the same direction as the household’s interest in not paying for them, the household is protected by a rule it did not ask for, drafted by a party that does not represent it, on a rationale that has nothing to do with fairness.

Alignments end. The one that produced this depends on a specific set of facts: uncertain load forecasts, expensive transmission, and a real risk that announced capacity never arrives. Change any of those and the incentive inverts. A utility confident the load is coming has every reason to build first and allocate later. The exemptions are already visible in the original, AEP Ohio’s tariff grandfathered existing data centers and expansions below the threshold and every exemption is a place where the alignment has already stopped holding.

Someone Tried to Move the Room

Which brings us to the part of this year that received almost no coverage at all.

In October 2025 the Secretary of Energy invoked Section 403 of the Department of Energy Organization Act, a rarely used authority, to direct the Federal Energy Regulatory Commission to open a rulemaking on the interconnection of large loads to the interstate transmission system, generally those above twenty megawatts, and to take final action by April 30, 2026. [15] Load interconnection had been a state and local matter for as long as there has been a division of authority over the grid.

Read as a venue question rather than an engineering one, that is an attempt to redraw the boundary around a decision. Be precise about what it was not: the proposal explicitly disclaimed reach over retail sales, local distribution, siting, behind-the-meter arrangements and loads under twenty megawatts. [15]

The state commissioners did not treat the disclaimer as dispositive. On November 11, 2025 their association, NARUC, passed a resolution urging the Commission to preserve state retail authority. Ten days later it filed comments arguing that FERC had never asserted jurisdiction over end-user load interconnections, that doing so falls outside the Federal Power Act’s boundaries, and that states set rates across customer classes. The filing also said something that will matter at the end of this essay. The venue for a retail customer affected by the service, NARUC wrote, is a state commission, and those customers get recourse by participating in state proceedings and through electing or influencing the appointment of state regulators. [16] Roughly a hundred and fifty parties filed initial comments. [17]

On June 18, 2026, the Commission did not decline federal action. It declined the single nationwide rule the directive had sketched. Instead it issued six orders to show cause under Section 206 of the Federal Power Act, one to each jurisdictional grid operator and its transmission owners, each preliminarily finding the region’s tariff unjust and unreasonable for want of clear provisions on large load integration, and each giving sixty days to justify the status quo or propose revisions. [18]

The federal venue expanded. The boundary was restated. The Commission took up transmission service to large loads, the study processes behind it, and the network upgrade costs that enter wholesale rates, and it said the orders were not intended to intrude on state authority, framing its action as reaching Commission-jurisdictional transmission service and transmission cost shifting while leaving retail customer protection to state regulators. [19]

Be careful how much that settles. These are preliminary findings opening proceedings, not a final rule and not an adjudication of the jurisdictional line. The accurate description is narrower than “the room held” and more interesting: the first federal attempt to redraw the boundary ended by formally restating the boundary it had put under pressure. That restatement is now on the record, in an order, available to be cited by the next state commission that needs it. It is also only that.

Notice how it was produced. By an association of state regulators filing comments in a federal docket, against a cabinet secretary, on a record most of the affected population will never know existed. Members of the public could have filed; the Commission maintains an office whose purpose is to help them do it. There was no rally, no referendum, no public event around which a constituency could form. The defense was conducted by the professionals already in the room, which was fortunate, because nobody else was coming.

The venue was contested within eleven months of producing the result, at the level of jurisdiction rather than merits, and it survived on a legal boundary and the people paid to notice it. Next time the drafting may be better, or the Commission differently composed.

$200 Million in the Wrong Room

It would be tidy to say the money is aimed at this. It is not.

The two largest AI super PAC networks have raised more than two hundred million dollars between them this cycle. Leading the Future, funded by OpenAI president Greg Brockman, Andreessen Horowitz and others, has raised about $140 million. Public First Action, launched in February with a $20 million donation from Anthropic that its spokesman says is restricted to public education rather than political spending, had raised $80 million by the end of June. Together the two put at least $44 million into forty House and Senate candidates in the first half of the year. [20] [21] Almost none of the advertising, on either side, mentions artificial intelligence at all. [22]

Nor do they want the same statute. Leading the Future argues for a national standard against a state patchwork; Public First Action was created to oppose federal efforts to freeze state progress on AI oversight. [21] The disagreement is real, not staged.

And AI governance is a different question from the one that decided your exposure to the buildout. The December 2025 executive order that launched the preemption campaign excludes from its proposed preemption, in terms, AI compute and data center infrastructure, other than generally applicable permitting reforms alongside child safety and state procurement. [23] The White House framework built on it, released March 20, 2026, preserves state police powers, and on ratepayers runs the other way entirely: it advises Congress to ensure that residential ratepayers do not face higher electricity costs as a result of new AI data center construction and operation. [24] The rate class survived the year not because the preemption campaign spared it. The campaign was never pointed at it, and on this question is nominally on the same side.

So the mismatch worth naming is not one of money. Campaign spending and docket intervention are not substitutable goods, and no quantity of the first buys a better outcome in the second. The mismatch is one of attention. Two hundred million dollars is contesting the governance of a technology in the one venue where a constituency can be assembled and rewarded. The question of who pays for the physical plant that technology runs on was divided between federal and state authority in a docket where no constituency existed at all.

The Win With No Owner

The problem with an unowned victory is not that it is unfair. It is that nobody will fight for it.

A movement can defend what it built. It cannot defend what it did not build and does not know it has. Test that operationally rather than rhetorically. No recognizable electoral constituency has formed around large load cost allocation. Few candidates run on it and no incumbent has lost a seat for weakening it. And no mass-membership organization exists to make one answer for it, consumer counsels and regulators’ associations do this work, but they are statutory offices and professional bodies, not a base that can be mobilized or disappointed. Three absences, each checkable.

The visible fight, if it goes well, makes this worse rather than better. A candidate wins in November on data center anger. Georgia’s commissioners are already gone. The sensation is that something was accomplished, and something was: seats changed hands over a real grievance, which is how the system is supposed to work. But the part of the grievance that got answered was answered elsewhere, by other people, for other reasons, and it can be unwound the same way. A public that believes it has already won is the least likely public to notice.

Four things follow. They are shaped like the venue rather than like a movement, because a movement is what the venue cannot produce.

Put the retail boundary in statute. The line held this June on a commission’s characterization of its own intent, in preliminary orders, which is the weakest durable form a legal boundary takes. It survives the current commission’s composition and the current drafting. A statutory reservation of retail cost allocation to the states is a specific ask, addressed to a specific committee, and it is the only version of this year’s outcome that does not have to be re-won.

Fund a standing intervenor whose mandate is the boundary. Consumer advocates already appear in these dockets and did real work in Ohio. Their mandate is rates. Nobody’s mandate is the jurisdictional line as such, which is why its defense in June depended on an association of regulators having the institutional reflex to file. A reflex is not a plan.

Send the anger to the appointment. This is the one place where volume genuinely migrates, and it is not my suggestion. It is NARUC’s, on the record in the federal docket: retail customers get recourse through state proceedings and through electing or influencing the appointment of state regulators. [16] Roughly ten states elect their commissioners. The rest appoint them, usually by a governor, usually with almost no public attention on the choice. Anger cannot file a brief. It can decide who sits on the commission, and Georgia has already demonstrated that it will.

Watch three specific things, since no public can monitor a docket. A special contract regime, in which individual projects negotiate around the tariff class rather than joining it, which turns a rule into a default. Erosion of thresholds and grandfathering, which was in the original Ohio design and is where every subsequent exemption will be argued from. And a federal definition of large load broad enough that retail cost allocation gets settled at the wholesale level before any state gets to argue it, which is the drafting error the next attempt will not make.

The party memo saw the shape of this correctly, which is why it is worth reading twice. It is frightened about Ohio. It is not frightened about anything upstream of Ohio, and it does not spend a sentence pretending to be.

You are winning a fight you do not know you are having, and you are winning it because someone else’s interest currently points your way. It does not appear on any ballot, and it was defended this year by an association of the people who do this for a living. Nobody defends a victory that was never theirs.

Notes

[1] Alex Isenstadt, “Exclusive: GOP warns AI companies that data centers are politically radioactive,” Axios, August 19, 2026 (Maria Curi contributing). The memo, headlined “Ohio Data Center Risk,” was obtained by Axios and posted in full by the reporter. Quoted phrases are the memo’s own language as reported. The comparison of data centers to spent nuclear waste is attributed in the same piece to internal party and industry polling described to Axios.

[2] NBC News, “New York to impose the country’s first statewide moratorium on data centers,” July 14, 2026; CNBC, “New York becomes first U.S. state to impose AI data center ban,” July 14, 2026, reporting that fourteen state legislatures had introduced bills restricting new data center construction and that none had been signed into law.

[3] Maine LD 307, passed by both chambers and vetoed by Governor Janet Mills in April 2026. The bill would have paused projects requiring at least 20 MW until late 2027. The veto cited the effect on a redevelopment project at the former International Paper mill in Jay.

[4] New York Executive Order No. 62 (July 14, 2026), imposing a one-year moratorium on discretionary state environmental permits for data centers of 50 MW or more and directing the Department of Public Service to examine large load interconnection in Case 26-E-0045. The legislature’s Responsible Data Center Development Act, containing a 20 MW moratorium, had not been signed as of the order.

[5] Seattle one-year moratorium on projects of 20 MW or more, adopted June 9, 2026; Monterey Park, California ballot measure prohibiting data centers citywide, approved by 88 percent of voters June 2, 2026; Texas pause on new approvals pending Public Utility Commission of Texas and ERCOT audits, announced August 2026.

[6] Georgia Public Service Commission special elections, November 4, 2025. Alicia Johnson (D) defeated incumbent Tim Echols (R) in District 2 and Peter Hubbard (D) defeated incumbent Fitz Johnson (R) in District 3, each with about 59 percent. Reported by the Atlanta Journal-Constitution, WABE and Axios Atlanta. The six Georgia Power rate increases since 2023 and the roughly $500 annual household increase are from AJC reporting.

[7] Fox News poll conducted July 17–20, 2026 by Beacon Research (D) and Shaw & Company Research (R); 1,003 registered voters; margin of error ±3 points. Asked whether respondents would favor or oppose the building of a data center in their area to support AI technology: 30 percent favor, 70 percent oppose. Nearly eight in ten preferred slower construction to rapid buildout.

[8] Gallup, “Americans Oppose AI Data Centers in Their Area,” May 2026, reporting a March survey in which 71 percent opposed construction in their area and 48 percent were strongly opposed.

[9] POLITICO poll conducted by Public First, July 2026: 41 percent would oppose a data center built within three miles of their home, up from 28 percent in January.

[10] Edison Electric Institute, “Large Load Projects and Tariffs,” updated July 2026: as of July 2026, 24 states have approved at least one large load tariff and another 6 have pending large load tariffs.

[11] Smart Electric Power Alliance, Database of Emerging Large Load Tariffs (DELTa), March 31, 2026 update: 51 approved and 26 proposed tariffs and service rules across 36 states and 60 utilities. Virginia GS-5 terms (25 MW threshold, 14-year contracts, $1.5 million per megawatt collateral) as reported in contemporaneous trade analysis of the Virginia State Corporation Commission’s approval.

[12] New Jersey A796/S731, signed July 2026. The enacted text directs the Board of Public Utilities to identify the defining characteristics of a “large data center” and “large data center customer” for each electric public utility, including a minimum megawatt size designation “which shall not be greater than 50 megawatts,” and to aggregate the peak monthly demand of data centers under common ownership or control, on the same or contiguous sites, or sharing substantial physical, operational or interconnection infrastructure, treating them as a single large data center. Utilities apply the rules within 180 days of the board’s order, to new and existing facilities at or above the threshold. The statute requires financial guarantees that a large data center customer will pay for at least 85 percent of the service it requests for not less than 10 years, and upfront deposits toward new transmission.

[13] Virginia HB 30, the 2026–2028 biennial budget, passed June 22 and signed by Governor Abigail Spanberger on June 30, 2026. It imposes a data center electricity consumption tax of $0.011 per kilowatt-hour effective July 1, 2026 through June 30, 2028, capped at $600 million annually with pro rata refunds above the cap, and preserves the existing data center retail sales and use tax exemption under Va. Code § 58.1-609.3(18).

[14] Public Utilities Commission of Ohio, Case No. 24-0508-EL-ATA. AEP Ohio applied on May 14, 2024. On July 9, 2025 the commission adopted a stipulation joined by AEP Ohio, commission staff, the Ohio Consumers’ Counsel, the Ohio Energy Group, the Ohio Manufacturers’ Association Energy Group and Industrial Energy Users-Ohio, requiring qualifying customers above 25 MW to pay for a minimum of 85 percent of subscribed load for up to 12 years with creditworthiness, collateral and exit-fee provisions, and directing AEP Ohio to file the tariffs and lift its connection moratorium. The moratorium had been imposed unilaterally in March 2023. Existing data centers and expansions below the threshold were grandfathered.

[15] Interconnection of Large Loads to the Interstate Transmission System, Advance Notice of Proposed Rulemaking, FERC Docket No. RM26-4-000, transmitted October 23, 2025 with the letter of Secretary of Energy Chris Wright under Section 403 of the Department of Energy Organization Act, 42 U.S.C. § 7173, requesting final action by April 30, 2026. The ANOPR addressed loads generally above 20 MW and disclaimed federal authority over retail sales, local distribution, siting, behind-the-meter and intrastate arrangements, and loads under 20 MW.

[16] National Association of Regulatory Utility Commissioners, Resolution Urging the Federal Energy Regulatory Commission to Preserve and Affirm State Retail Regulatory Jurisdiction in Its Large Load Interconnection Proceeding (passed by the Committee on Electricity Nov. 10, 2025; adopted by the NARUC Board of Directors Nov. 11, 2025). See also Initial Comments of the National Association of Regulatory Utility Commissioners, FERC Docket No. RM26-4-000, at 3–6 (filed Nov. 21, 2025), Accession No. 20251121-5132. The comments state that FERC has never asserted jurisdiction over end-user load interconnections and that doing so would exceed the boundaries imposed by the Federal Power Act. They further explain that “the venue for a retail end-use customer who is directly affected by the services provided by the electric supplier is a state commission,” with recourse available “by directly participating in state proceedings and through the process of electing or influencing the appointment of state regulators.”

[17]More than 150 initial comments were filed in Docket No. RM26-4-000, as reported in contemporaneous trade coverage.

[18] Six orders to show cause under Section 206 of the Federal Power Act, all issued June 18, 2026 by unanimous vote: PJM Interconnection, L.L.C., 195 FERC ¶ 61,211 (2026) (Docket No. EL26-67); Midcontinent Indep. Sys. Operator, Inc., 195 FERC ¶ 61,212 (2026) (Docket No. EL26-70); Sw. Power Pool, Inc., 195 FERC ¶ 61,213 (2026) (Docket No. EL26-68); Cal. Indep. Sys. Operator Corp., 195 FERC ¶ 61,214 (2026) (Docket No. EL26-71); ISO New England Inc., 195 FERC ¶ 61,215 (2026) (Docket No. EL26-72); N.Y. Indep. Sys. Operator, Inc., 195 FERC ¶ 61,216 (2026) (Docket No. EL26-69). Each makes preliminary findings that the tariff appears unjust and unreasonable with respect to large load integration and sets a 60-day response deadline. The intervention deadline was July 9, 2026; replies to the show cause responses were due September 16, 2026. See also Interconnection of Large Loads to the Interstate Transmission Sys., 195 FERC ¶ 61,045 (2026) (order regarding intent to act).

[19] The Commission stated that the orders were not intended to intrude on state authority, framing its action as directed to Commission-jurisdictional transmission service and transmission cost shifting while leaving retail customer protections to state regulators, and indicating that the findings were not intended to preempt state large load tariffs. See PJM Interconnection, L.L.C., 195 FERC ¶ 61,211 (2026).

[20] CNBC, “What AI companies want for the millions they’re spending on elections,” July 9, 2026: Leading the Future raised $125 million by the end of 2025 and spent more than $24 million on primaries through the end of June; Public First Action raised $80 million through the end of June and spent $20 million, including a $20 million donation from Anthropic that a PAC spokesman describes as restricted to educating the public on AI policy rather than political purposes; the two PACs together put at least $44 million into 40 House and Senate candidates through the end of June. Reuters reporting in August 2026 put Leading the Future’s total raised at about $140 million, with funders including Greg Brockman, Andreessen Horowitz, Joe Lonsdale, Ron Conway and Perplexity.

[21] NPR, “Groups tied to OpenAI and Anthropic are spending big on the midterms,” June 22, 2026, describing Leading the Future’s opposition to stricter regulation and Public First Action’s stated purpose of opposing federal efforts to freeze state progress. Anthropic’s $20 million contribution was announced in February 2026.

[22] Ben Kamisar, “Ads funded by AI industry are flooding the 2026 election. They’re about everything except AI.,” NBC News, February 27, 2026.

[23] Executive Order 14365, “Ensuring a National Policy Framework for Artificial Intelligence,” December 11, 2025. The order directs that the recommended federal framework not propose preemption of state laws concerning child safety protections, AI compute and data center infrastructure other than generally applicable permitting reforms, state government procurement and use of AI, and other topics to be determined.

[24] The White House, National Policy Framework for Artificial Intelligence, released March 20, 2026 pursuant to Executive Order 14365. It calls on Congress to preempt state AI laws that impose undue burdens while preserving state police powers, and advises Congress to ensure that residential ratepayers do not experience increased electricity costs as a result of new AI data center construction and operation.

The modern university sold admission. Knowledge came with it, and the knowledge was real, but access was the product. For most of a century the degree functioned as a certified claim on the professional entry layer: the place where young people became economically useful, formed judgment through practice, and began the process the labor market called a career. The signal was genuine. What the degree said was: you have been selected, therefore you deserve a beginning.

That beginning depended on something the university did not control. It depended on employers maintaining the entry-level work through which a beginning could happen.

AI is compressing that work. Not eliminating it, not everywhere, not yet, not uniformly. But compressing it at the layer where the credential was always redeemed: the first drafts, the research summaries, the document reviews, the first-pass models, the cleaned datasets, the formatted decks, the support calls, the code cleanup, the market scans. These were never the point of the work. They were the mechanism through which inexperienced people became experienced, through which institutions reproduced themselves, through which professions renewed the judgment they needed to survive.

This series opened with that fracture. What follows is about the institution that sold tickets to it.

The university is still selling the credential, employers are still requiring it, and the entry layer where it was supposed to be redeemed is being compressed. The student signs the note. Nobody updates the terms.

The Degree Was Never a Certificate of Knowledge

Universities have always done many things: generated research, housed inquiry, transmitted culture, produced social networks, conferred status, and occasionally changed what people understand about the world. But for most students, in most institutions, across most of the twentieth century, the core transaction was simpler than any of those. The degree certified selection. It told employers that the holder had been evaluated, admitted, and survived a multi-year sorting process administered by an institution with a reputation to protect. The content varied. The signal was consistent. A degree from a recognizable institution meant something before anyone read the transcript, saw the portfolio, or conducted the interview. It meant: this person has been screened.

Licensed fields are the genuine exception. Where a bar examination or a medical board gates practice, the credential certifies knowledge directly and is designed to. The claim here is narrower than higher education as a whole, and concerns the professional-class bargain: the tracks where a degree bought entry rather than a license.

The employer understood what the screen was for. Not genius, not mastery, trainability. Nobody expected the entry-level hire to arrive knowing how to do the job. They expected someone who could be taught it.

The degree was a futures contract on trainability.

The student paid tuition, time, debt, and opportunity cost now in exchange for a claim on future professional standing. The employer paid a salary and training investment now in exchange for a claim on future productivity. The university sat in the middle, collecting on both sides: tuition from the student, reputation from the employer who returned to recruit, and prestige from the graduates who succeeded.

The arrangement worked because it rested on a stable underlying asset: the junior role. Someone had to do the lower-stakes work while developing the judgment to do the higher-stakes work. Employers maintained that tier not because they preferred expensive inefficiency but because professional reproduction required it. You cannot have a firm of only senior partners. You can only have senior partners if you once had junior associates. The ladder was not charity. It was infrastructure.

The university was also a sorting interface, making young people legible to the professional class before they arrived at it. Admissions sorted, grades ranked, internships routed, career offices translated ambition into employer-readable categories, and rankings told employers which pools were worth fishing in. All of it condensed into a portable object that could move across firms, cities, and professions, presenting screened and ranked candidates to employers too busy to evaluate them from scratch.

That interface worked as long as there was still a beginning on the other side to hand them to. That is what is changing.

The Entry Layer Was Always Where the Degree Was Redeemed

Professions do not reproduce themselves through instruction alone. A law school can teach doctrine. It cannot teach a student how to read a room before a judge, how to write a brief that a partner will not rewrite, how to know when a client is lying, or how to recognize that the statute does not mean what it says on first reading. None of it survives transmission through lecture. It forms through practice inside a hierarchy that permits error at low stakes and corrects it at close range.

Medicine understood this first. The residency exists because clinical judgment cannot be produced any other way. You cannot simulate the responsibility of being the person who decides; you develop it by being that person, under supervision, with real patients, real consequences, and real correction from people who have done it before.

What medicine formalized through residency, other professions accomplished informally through the junior role. The consulting analyst spent a year summarizing before she could synthesize; her counterparts in law, engineering, and banking ran the same apprenticeship. These were the transfer mechanism for tacit professional knowledge, and the mechanism required the junior work to exist.

AI is compressing the junior work. The early-career decline that opened this series has since widened. In the August 2026 revision of that research, employment among workers aged 22 to 25 in the most AI-exposed occupations stands roughly nineteen percent below where it would be had it kept pace with their less-exposed peers, with no comparable gap among experienced workers, and the divergence operating primarily through reduced hiring rather than increased separations. The finding survives controls for interest-rate exposure, remote work, and technology-sector concentration. It is weaker in other respects. The pattern attenuates once education is controlled for, some of the divergence predates generative AI, and it runs stronger in the payroll sample than in national survey benchmarks. The authors describe it as an early descriptive indicator rather than a causal estimate, and that caution belongs in the argument rather than only in a note. What the data establishes is the shape of the gap. What connects that gap to the claim made here, that firms are hiring fewer juniors because AI has absorbed the work which used to pay for their formation, is a mechanism rather than a measurement, and no one has yet demonstrated it at the level of the firm. What the numbers show is not mass unemployment but compression at the intake.

The compression is sharpest in cognitive professional work, in the roles where the first years involved structured, documented, repeatable tasks. Fields where physical presence and hands-on judgment remain central, trades, clinical nursing, laboratory work, are less exposed at the entry layer for now. But the professional tracks the university historically sold most confidently, law, finance, consulting, engineering, accounting, are precisely the tracks where the junior role is being reorganized fastest.

Senior professionals become more productive. Fewer juniors are hired to do the work that made seniors. Companies still want experienced workers, but the pipeline producing them is narrowing at the intake. The profession still looks healthy from the top while the intake thins from below.

The credential was a ticket to the entry layer. The entry layer is the asset the ticket was written against.

The Student as Residual Risk Holder

The university collects tuition at enrollment. The employer decides, years later, whether to maintain the entry layer.

The student carries the gap.

This is the structural position that no one announced and everyone is now discovering. Universities have no contractual obligation to ensure that the professional pathways their degrees imply will remain open by the time the student graduates. Employers have no obligation to maintain junior roles that their AI infrastructure has made less necessary. Accreditation bodies certify curriculum, not career outcomes. Rankings measure inputs and research outputs, not the integrity of the claim the degree is implicitly making. The student signs a note whose underlying asset is controlled by a third party who was never party to the transaction, and that third party is quietly revising the terms.

Law firms are using AI to compress document review, the entry task for associates. Consulting firms use it for first-pass research and slide construction, the apprenticeship layer for analysts. Investment banks use it for financial modeling, the training ground for junior bankers. Software companies use it for code completion, testing, and documentation, the work through which junior engineers once learned the codebase. Accounting firms use it for initial audit procedures, the entry work through which new accountants developed judgment. [1]

None of these firms announced that the ladder was changing. In many cases they still recruit at universities, still attend career fairs, still describe junior roles in language that implies the old bargain. The form is preserved while the substance is compressed.

The student graduates into a market where the credential is still required and the thing it was supposed to purchase has become scarcer. She still owes the note. Outstanding student loan balances in the United States stood at roughly $1.65 trillion in mid-2026, and the share of those balances ninety or more days delinquent has risen above ten percent, from well under one percent before the pandemic-era repayment moratorium ended. [2] These are commitments made against expected labor-market outcomes, and the expectations were formed in an environment where the entry layer was assumed to persist. The debt is still there. The assumption is under revision.

Credential Inflation Is Panic Buying

When a gate starts closing, people buy more tickets. Credential inflation is the behavioral signature of a population that suspects the first credential no longer clears the gate and is responding by acquiring more. The graduate degree after the bachelor’s. The professional certification after the graduate degree. The bootcamp certificate after the certification. The AI badge after the bootcamp. The portfolio. The unpaid internship. The side project that demonstrates initiative. The conference talk.

Each additional credential is rational for the individual acquiring it. If the first degree no longer reliably distinguishes candidates, add a second that might. If the diploma no longer signals trainability, demonstrate it directly. If the credential no longer certifies selection, produce more evidence of being selected. The aggregate result is an escalating floor on entry that serves no one systematically. Students spend more on signaling. Employers face higher signal noise. The credential bar rises without the underlying access expanding. The cost of attempting to enter the professional class increases while the probability of succeeding diminishes.

There is a complication here worth stating plainly. Research on job postings has documented a degree reset: employers who had raised formal degree requirements later removed them from a wide range of roles, with most of the change beginning before the pandemic and appearing structural rather than cyclical. [3] If employers are dropping the credential requirement, the gate is not closing at all.

But the reset does not open the gate so much as change what the ticket is. When employers can evaluate demonstrated capability directly, they stop paying for the credential as a proxy and screen on the capability instead. The requirement was never a fixed condition of entry, only an option the employer held, exercised while credential screening was the cheapest available filter and abandoned once something cheaper appeared. The student who financed the credential financed a screening convenience the counterparty is free to stop using, and competition for the underlying positions does not fall when the formal requirement is dropped.

Credential inflation is the sound of people buying more tickets to a gate that is closing. The university profits from each ticket sold. The university does not control the gate.

The Structural Fraud Nobody Committed

Fraud, in law, requires a misrepresentation of material fact, knowledge that the representation is false, an intention that someone rely on it, actual reliance, and resulting loss. Higher education fails that test early. University administrators are not deceiving anyone about what their institutions do. Employers are not making promises they intend to break. Legislators who fund student loan programs are not designing a trap. Individual faculty, admissions officers, career counselors, and university presidents are mostly acting in good faith inside systems they did not design and cannot individually change. There is no misrepresentation and no scienter.

Fraud is therefore the wrong legal category and the right structural metaphor. The aggregate structure produces outcomes with the shape of fraud. An institution sells a claim on a future asset. The future asset is controlled by a third party under no obligation to maintain it. The institution keeps selling the claim after the third party has begun withdrawing the asset. The student bears the loss.

This is structural fraud without a fraudster.

The university’s interest is in maintaining enrollment. Enrollment depends on the perception that the credential is worth acquiring. The credential appears worth acquiring because employers still require it. Employers still require it even as they compress the tier where it was historically redeemed, partly because the requirement is embedded in HR systems and organizational habit, partly because the credential still screens for something useful, and partly because alternatives to credential screening are still being built. Each of the three parties, university, employer, and student, behaves rationally inside its own constraints. The system-level result is a transfer of risk onto the least powerful party for a product whose underlying asset is deteriorating.

William Deresiewicz described the credential system as producing excellent sheep: students skilled at clearing the next gate without understanding what the gates were for. [4] That diagnosis was sociological. AI makes it structural. The gates keep requiring the tickets. The territory beyond them is being reorganized by a party that issues neither.

The Reproduction Crisis

A society reproduces its professional class through the entry layer. Medicine reproduces through residency, law through associate programs, finance through analyst classes, journalism through beats and copy desks and editors who corrected in real time. These were knowledge transfer systems as much as employment arrangements. They moved tacit professional knowledge from one generation to the next through supervised practice at low stakes, and they allowed professions to exist in the future by continuously training the people who would inhabit them.

AI is entering that transfer system at its most vulnerable point. The junior work that made the transfer function is precisely the work most susceptible to compression: structured, documented, repeatable, and lower-stakes by design, which is exactly the profile of task that current generative systems handle most effectively. A controlled trial of AI coding assistance found a large productivity gain on a bounded, structured programming problem, the kind of work through which junior developers traditionally build fluency. [5] The productivity gain and the hiring compression are the same signal, because a tool that raises output at the junior level also reduces the urgency of maintaining that level at all.

Nothing about this is hidden. Thomson Reuters, surveying professionals across law, tax, audit, and compliance in 2026, counts among the costs of the current transition a generation of professionals slower to develop the independent judgment their work requires. That records what practitioners expect rather than skill already measured as lost. Deloitte, surveying 1,874 workers across four countries, puts the mechanism plainly: AI is being built to automate the tasks early-career workers handle, which may reduce entry-level openings and the on-the-job learning that career growth depends on, leaving executives with a pipeline that struggles to produce future leaders. [6] The firms compressing the transfer layer are also the ones documenting what compression does to it.

If the transfer layer thins, professions eventually exhaust their supply of experienced practitioners. Not immediately, not in one cycle, but structurally, across a generation. A profession that stops training its next cohort does not notice for years, and then it cannot find the people it needs. The reproduction crisis is less a prediction than a name for what happens when a transfer system is disrupted faster than anyone recognizes it was the transfer system.

Firms Need Seniors, So Firms Will Train Juniors

That is the strongest objection to everything above, and it deserves to be stated at full strength. Professional reproduction is a private necessity before it is anything else. A law firm without associates has no partners in fifteen years. An investment bank without analysts has no managing directors. If the junior layer is genuinely load-bearing, self-interest maintains it, and what looks like collapse is a transition the market resolves as it has resolved others. New categories of work have repeatedly appeared where old ones were automated away, and the professional hierarchies of 2045 may rest on junior roles that do not yet have names. The same 2026 survey describes something like this already underway: associates whose work shifts toward interpreting AI-assisted output, workflow design, and technology oversight. [6]

The objection has real force. It fails, if it fails, on three points.

The first is that firms never paid for training at all; they paid for work, and formation came attached to it. The junior associate’s document review was billable and the judgment it produced was a byproduct. Training was free to the firm because the trainee generated value while being trained. AI removes the byproduct. Once the first draft no longer requires a human, the junior’s output stops financing the junior’s formation, and training converts from a byproduct into a line item with a payback period longer than the tenure of the executive who approves it.

The second is that formation is portable and the firm paying for it cannot capture it. The trained associate can leave. That was tolerable while training was a byproduct and becomes intolerable once it is a cost, because every firm would prefer that the industry keep training juniors and that some other firm carry the expense. This is the standard structure of an underprovided input, and good intentions do not resolve it.

The third is timing. The correction arrives through shortage, and shortage becomes visible only after the cohort that would have prevented it was not trained. By the time a profession discovers it cannot staff its senior tier, the decade in which it could have acted has already passed.

None of this establishes that new junior work will not appear. It may. But formation requires more than tasks. It requires tasks embedded in a hierarchy that permits error at low stakes and corrects it at close range, with someone senior near enough to see the error and invested enough to explain it. Whether supervising machine output constitutes such a hierarchy is an open question, and it is the question on which the objection finally rests.

The credential still clears the gate. It still requires debt. It still signals selection. What it can no longer reliably do is redeem itself in the layer for which it was always a ticket.

The university sold the ladder for a hundred years. Nobody told it the ladder was load-bearing.

Notes

[1] On AI absorption of entry-level professional work across sectors: McKinsey Global Institute, “The Economic Potential of Generative AI: The Next Productivity Frontier,” June 2023, identifying knowledge work at the analyst tier among the highest-exposure categories; Thomson Reuters Institute, “AI in Professional Services Report 2026,” recording generative AI use among law firm legal teams at 41 percent against 28 percent a year earlier and among corporate legal departments at 47 percent against 23 percent, with document review, legal research, and contract analysis among the primary use cases; Deloitte, “Artificial Intelligence Insights for Internal Audit,” on generative AI producing the initial draft of workpapers and conducting the first round of review and quality checks. https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier

[2] Federal Reserve Bank of New York, Center for Microeconomic Data, Quarterly Report on Household Debt and Credit, second quarter 2026. Student loan balances fell by $7 billion to $1.65 trillion; the share of student loan balances ninety or more days delinquent stood at 10.6 percent, against a fraction of one percent before the pandemic-era repayment moratorium ended. The New York Fed cautions that re-reporting of previously defaulted student debt continues to distort the student loan series, so the delinquency level should be read as a range rather than a point. https://www.newyorkfed.org/microeconomics/hhdc

[3] Joseph B. Fuller, Christina Langer, Julia Nitschke, Layla O’Kane, Matthew Sigelman, and Bledi Taska, “The Emerging Degree Reset: How the Shift to Skills-Based Hiring Holds the Keys to Growing the U.S. Workforce at a Time of Talent Shortage,” Burning Glass Institute, February 2022, produced with Harvard Business School’s Project on Managing the Future of Work. The study found material degree resets in 46 percent of middle-skill and 31 percent of high-skill occupations between 2017 and 2019, with 63 percent of the changed occupations showing structural rather than cyclical resets. It reverses the degree inflation documented in Joseph B. Fuller and Manjari Raman, “Dismissed by Degrees,” Harvard Business School, 2017. https://www.burningglassinstitute.org/research/the-emerging-degree-reset

[4] William Deresiewicz, Excellent Sheep: The Miseducation of the American Elite and the Way to a Meaningful Life (Free Press, 2014). Deresiewicz wrote before the labor-market conditions described here, and the argument was addressed to what elite education does to students rather than to what happens on the far side of the gate.

[5] Sida Peng, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer, “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot,” 2023, arXiv:2302.06590. Developers assigned to implement an HTTP server completed it substantially faster with AI assistance. The design is a single bounded task, so it measures the size of the gain on work of that kind and does not compare gains across task types; it is evidence about bounded work, not evidence that bounded work is where AI helps most. Nor does it show that junior developers become unnecessary. It shows why firms may be tempted to reorganize the junior layer around fewer people and more tool-mediated supervision, which is the compression claim rather than the elimination claim. https://arxiv.org/abs/2302.06590

[6] Thomson Reuters Institute, “Future of Professionals Report 2026,” June 2026, the fourth annual edition, drawn from 1,816 responses gathered in March and April 2026 across 62 countries. It counts among the costs of the current transition a generation of professionals slower to develop independent judgment, and describes senior associate roles shifting toward interpretation of AI-assisted output, workflow design, and technology oversight. These are practitioner expectations reported in a survey, not observed measures of skill formation. Elizabeth Lascaze et al., “AI is likely to impact careers. How can organizations help build a resilient early career workforce?” Deloitte Insights, December 6, 2024, surveying 1,874 workers in the United States, Canada, India, and Australia, of whom 65 percent were early career: AI technologies are being developed to automate the tasks these workers typically handle, which could reduce entry-level openings and the on-the-job learning that matters for career growth, with consequences for talent pipelines and the sourcing of future leaders. https://www.thomsonreuters.com/en/institute/future-of-professionals-2026/report and https://www.deloitte.com/us/en/insights/topics/talent/ai-in-the-workplace.html

The interesting case is that majority. Not the near-peer that makes itself hard to abandon by becoming indispensable to others, but the ordinary medium state that holds no frontier layer at all and knows it. It is dispensable. It will remain dependent on foreign infrastructure for the parts of governance that now run through models. The interesting question is not how such a state escapes dependence. It cannot. The question is whether it can stay sovereign inside dependence it will never eliminate.

Renting a mind is not like renting a machine

Industrial dependence has always been survivable. A state that buys its aircraft, its steel, its turbines from abroad does not thereby change what its finance ministry can understand about the economy or what its health ministry can know about a disease. The dependence sits downstream of the state’s own judgment. It constrains what the state can build, not what it can see.

Renting cognition is not wholly new, but it is a sharp intensification. States have long taken in things that sit upstream of their own judgment: credit ratings, foreign intelligence feeds, the settlement rails their payments clear through. What changes with models is scope and mutability. When a ministry’s analysis, a court’s evidence handling, a central bank’s fraud detection, and a hospital’s diagnostics all run through models the state neither trained nor controls, the dependence moves upstream of judgment across many domains of government at once, and the provider can revise its defaults remotely, after the contract is signed. The provider’s defaults become the state’s priors. Nothing is conquered. The flag stays, parliament sits, the courts convene. But increasingly the state reasons through infrastructure whose governing conditions it does not set, and a provider that changes its terms is changing, quietly, what the state is able to decide.

That is why the standard sovereign-AI slogan, build your own so you depend on no one, is the wrong instruction for the country that cannot build its own. It sets an unreachable bar and calls everything short of it failure. The reachable bar is different. The stronger recent treatments already sense this: they abandon autarky and redefine the goal as resilience, switch providers, refuse lock-in, hold selective domestic capacity, keep contingency in reserve.¹ They are right, and they stop at the level of capability. Dependence that has moved upstream of judgment is not only a capability problem. It is a constitutional one. Once foreign models sit beneath what a state can perceive and decide, the power to substitute them is no longer procurement prudence; it is part of the state’s capacity to govern itself. That is the bar worth naming precisely, and it is why two very different strategies both answer to the word “sovereign.”

Two ways to hold a dependence

The first way buys entry by binding yourself to a single stack’s home jurisdiction. The clearest live specimen is the arrangement around G42 in Abu Dhabi. Microsoft’s 2024 investment came wrapped in a first-of-its-kind Intergovernmental Assurance Agreement, a binding private framework developed with the American and Emirati governments that commits the parties to meet or exceed United States standards on cybersecurity, physical security, export controls and technology transfer, data protection, and customer vetting.² It is a serious instrument, and it does convert raw dependence into something governed and inspectable. But look at where the conditions point. The Emirati side retains a golden-share style veto over national-security decisions on paper, while the operative constitution of the arrangement, the part that decides what may be built and who may touch it, runs toward Washington’s export-control regime and licensing discretion. This is dependence made conditional. It is not dependence made reversible. The state has bought access to the frontier by deepening its alignment to one pole.

That trade has its theorist. argues that the middle power’s real choice is between a permanent periphery and binding itself to American frontier AI, and that the binding is worth its price: offering sites, energy, capital, and data-center capacity in exchange for access, and accepting deeply entrenched technological dependence and even geopolitical submission as the cost of staying inside the AI economy.³ He is right that the periphery is real and right that access cannot simply be refused. But his account ends where the harder problem begins. It answers the question of entry and treats the price as paid once, at the door. The dependence that decides sovereignty is the one that keeps rewriting its own terms after the state is already inside, and against that, alignment to a single pole is not protection but exposure.

The second way keeps the capability foreign but engineers the dependence so it can be inspected, wrapped, substituted, and left. France offers the cleanest working model. Its “cloud de confiance” doctrine sits on the ANSSI SecNumCloud qualification, whose 3.2 revision adds explicit immunity criteria against extraterritorial law, and it is delivered through French-law joint ventures, S3NS built on Google technology, Bleu on Microsoft’s, that legally isolate the operation from the foreign parent.⁴ The technology is American. The custody, the legal exposure, and the operational control are domestic. The design goal is stated plainly by the regulator: no foreign kill switch, no extra-European access to the data, guaranteed localization.

The European Union has turned the same instinct into hard law. Since September 2025 the Data Act has given every cloud customer in the Union a statutory right to switch providers and receive technical cooperation to port their data, and from January 2027 it forbids switching and egress charges entirely, removing the single most effective lock-in mechanism a provider has, the exit toll.⁵ Portability stops being a favor and becomes an entitlement.

Singapore shows the capability side of the same posture. Its state-backed SEA-LION models are not an attempt to replace the frontier labs. They are, in the words of the program’s own director, an effort to remain a “smart consumer” of global systems and to “engage on equal footing,” with open, forkable licensing as the deliberate hedge, backed by a national multimodal programme and a multi-year public compute and research commitment.⁶ The point is not a sovereign model that beats OpenAI. The point is enough domestic capacity that the state understands what it is renting and could keep essential functions running if the terms changed.

These are not three national choices among which a state picks one. They are layers of a single posture: legal custody, statutory portability, and a domestic capability floor, each covering an exit the others leave open.

What none of this buys

Here is the honesty these prescriptions usually skip. None of it is autarky, and pretending otherwise sets the reader up to be disappointed by the first hard fact.

The head of France’s own cyber agency said it directly when the trusted-cloud qualification drew fire: the badge does not mean the absence of dependence. By the agency chief’s own public estimate, the two flagship hybrid offerings could sustain operations for perhaps six to twelve months if they lost access to their American parents’ updates.⁷ The Data Act kills the egress fee, but proprietary formats, bespoke interfaces, and identity binding survive the fee ban, so a right to leave on paper is still not an exit you can execute unless you have kept your data and workloads portable in fact. Singapore’s flagship regional model is itself adapted from a Chinese-origin foundation, which means its hedge against one dependence runs partly through another. But even together these protections reach only some layers of the dependence. Beneath legal custody, statutory portability, and model forkability sits the hardware, the accelerators and the fabrication that decide whether there is any compute to run the fallback on at all, and no exit clause reaches it. Software can be forked; the machine that runs it still has to be sold to you, and that sale is another government’s decision. That the deepest layer stays out of reach does not refute the posture; it defines its limits. The claim is not that every dependence can be reversed. It is that reversibility at the layers a state can reach, custody, portability, forkable weights, a domestic inference floor, is worth more than submission across all of them. A hedge that fails at the hardware base can still hold at every layer above it. And even a clean exit is not a return to the same state. Swap the model and the function survives, but the judgment shifts, since two systems will classify the same evidence, price the same risk, and triage the same patients differently. Portability preserves what the state can do, not how it decides, which is the thing the dependence touched in the first place.

So the reachable condition is not independence. It is reversibility, maintained continuously, layer by layer, contract by contract. What that reversibility races against is speed: dependence becomes command at the point where a provider can impose a change faster than the state can substitute away from it. France’s six-to-twelve-month figure is not trivia. It is that race made numerical, the runway a trusted-cloud arrangement buys before a withdrawal of access turns into control over what depends on it. And that runway is not fixed. It shortens as the frontier pulls away, because the fallback a state would exit into falls further behind the system it is exiting with every month the gap widens, so the exit is cheapest to architect now and dearer every quarter it is deferred. Reversibility is not a state you arrive at and hold. It is a property that decays from two directions at once: from the moment procurement stops enforcing it, and from the speed of the very divergence it exists to hedge. That is a weaker claim than the sovereign-AI literature likes to make. It is also the only one a country that cannot build the stack can actually keep.

Synthetic nonalignment

Call the resulting posture synthetic nonalignment. The Cold War’s nonaligned movement was diplomatic. It was made of communiqués, summits, and public refusals to join a bloc. A state could sign every declaration and still be, in practice, wholly inside one patron’s orbit, because alignment then was a matter of treaties and basing rights that a leader could announce or renounce.

Alignment now is built lower down, and it cannot be announced away. A state can declare neutrality between Washington and Beijing while its ministries, its banks, its military logistics, its universities, and its public services all run on a single foreign stack. That state is not nonaligned in any sense that survives a change of terms. Its neutrality is a press release sitting on top of a dependency it has never architected.

It is not the newer answer either. The most recent proposals revive nonalignment as a coalition, an open and collaborative bloc smaller states can join for shared capability.⁸ That is a movement, and a movement is something you join and can be eased out of. Synthetic nonalignment asks no one’s permission and needs no coalition to hold. It is what a single state does on its own procurement authority, and open weights are what make the unilateral move possible: a model you can hold, run, and adapt without asking is one whose provider cannot quietly become your sovereign. Forkability here is not a preference for openness as a good. It is the mechanical precondition of an exit you can take alone.

Synthetic nonalignment is therefore not a diplomatic stance but an engineering and procurement discipline. It lives in things that sound too boring to be sovereignty: data-custody clauses, interoperability requirements, model portability, audit rights, fallback models, domestic inference capacity, compute reserves, sovereign-cloud joint ventures, forkable licenses, and the tested ability to switch providers. The constitution of the dependent state is being written, in part, in its supplier contracts. A provider that revises a model’s behavior, retires a version, narrows permitted use, or changes what its system will and will not answer has altered what an agency can perceive and execute, and it has done so without a single line of law being amended. Whoever drafts the exit clauses is doing constitutional work whether they know it or not.

This is the precise inverse of the binding logic modern integration runs on: states become integrated not when they agree but when exit becomes prohibitively expensive, and legitimacy follows the binding. Synthetic nonalignment is the deliberate refusal of that condition. It is the work of keeping exit affordable, provider by provider, so that no single dependence ever hardens into the thing you cannot walk away from. Where binding makes leaving unthinkable, this keeps leaving costed, drilled, and possible.

The instruments

None of the instruments is new. Multi-sourcing, interoperability mandates, exit clauses, contingency reserves: the policy literature has recommended them for a year, as prudence.¹ The shift is in why they matter. Read as procurement hygiene they are optional, the first line cut when a budget tightens. Read as constitutional maintenance they are what keeps a dependent state governing itself, and they decay the moment enforcement stops.

So, for a state that will rent the frontier and still intends to stay sovereign inside that dependence: write portability and a costed exit price into every critical contract, so leaving is a known number and not a discovered catastrophe. Fund a domestic inference floor, enough sovereign compute and at least one forkable model to keep essential functions running through a change of terms, sized as insurance rather than autonomy. Wrap sensitive workloads on the French pattern, foreign capability under domestic legal custody, with no foreign kill switch as a stated requirement rather than a hope. Make multi-sourcing a rule, so no provider becomes indispensable to a critical function, because indispensability once granted is what gets priced against you later. Buy audit rights, so a dependence you cannot remove is at least one you can inspect. And drill the substitution, because the exit you never test is the exit you do not have.

None of this produces a sovereign AI. It produces a state that can say no, or at least can leave, which in an order organized around infrastructure is most of what saying no has ever meant.

The third position

The coming order is usually drawn as two blocs, an American stack and a Chinese one, with every other state sorted into one column or the other, and the counsel that follows is to choose a column early and make oneself valuable inside it. That map has room for a third position it tends to omit: states whose principal strategic capability is neither building the frontier nor choosing a patron, but managing dependence across providers, and, where possible, across rival stacks, so that none can convert access into command. That fuller balancing is still more direction than destination. Singapore, hedging a Chinese-origin model base against American frontier systems, is nearer to it than most, and even it holds only fragments. Such a state is not powerful in the old sense. It owns almost nothing at the frontier. What it holds is the harder, quieter competence of never letting a supplier become a sovereign.

That competence will not be visible in summits or communiqués. It will be visible, if at all, in the fine print of procurement, in the boredom of a well-run exit clause, in a ministry that has rehearsed leaving and a provider that knows it. The next nonaligned movement, if there is one, will be built from those clauses rather than from declarations.

Sovereignty is no longer the stack you own. It is the exit you can still afford.

Notes

1. Chatham House, Digital Society Programme, “How middle powers can weather US and Chinese AI dominance: the case for ‘sovereign AI’ strategies” (February 2026): frames full-stack sovereignty as unattainable for middle powers and recommends provider switching, multi-vendor interoperability standards in government procurement, selective sovereign capacity in critical domains, and pre-negotiated contingency arrangements for rapid provider substitution.

2. Microsoft and G42, Intergovernmental Assurance Agreement announced alongside Microsoft’s 1.5 billion dollar investment in G42 (April 2024), and Microsoft’s subsequent UAE investment framework (November 2025): binding commitments on cybersecurity, physical security, export controls and technology transfer, data protection, responsible AI, and customer vetting, developed in consultation with the US and UAE governments; UAE golden-share control over national-security decisions. Microsoft corporate statements; Core42 remarks on the physical-diversion and access-control elements (July 2026).

3. Anton Leicht, “What happens to countries that don’t build frontier AI?” Asterisk, Issue 15 (July 2026): argues that states without a domestic frontier developer face a “permanent periphery,” that sovereign-AI autarky is a trap, and that middle powers should move toward American frontier AI, offering sites, energy, capital, and data-center capacity for access and accepting entrenched technological dependence and geopolitical submission as the price of avoiding the periphery.

4. ANSSI SecNumCloud qualification, version 3.2 (2022), including immunity criteria against extraterritorial legislation. S3NS (Thales / Google Cloud) SecNumCloud 3.2 qualification for its PREMI3NS trusted-cloud offer (late 2025); Bleu (Microsoft / Capgemini / Orange) in qualification. Both structured as French-law joint ventures isolating data governance from the foreign technology parent.

5. Regulation (EU) 2023/2854 (Data Act), Chapter VI cloud-switching provisions in force since 12 September 2025; Article 29 phase-out of switching charges, with a full prohibition on switching and egress charges from 12 January 2027. Extraterritorial in scope: applies to any provider serving Union customers.

6. SEA-LION, AI Singapore, under the National Multimodal LLM Programme; companion MERaLiON model (A*STAR). Program director’s framing of “smart consumers” of global systems and engaging “on equal footing,” with open, forkable licensing. National AI research and talent commitment exceeding one billion Singapore dollars, 2025 to 2030.

7. ANSSI director general, clarification on the scope of SecNumCloud qualification (January 2026): qualification does not signify the absence of technological dependence; its guarantees are the prevention of a foreign kill switch, protection against extra-European authority access, and EU localization. Estimate that the qualified hybrid trusted-cloud offers could maintain operations for six to twelve months if cut off from their US technology parents’ updates.

8. For the coalition framing, see Martin Tisné, “Towards an Open, Resilient, Non-Aligned AI,” Le Grand Continent (February 2026), which proposes a new non-aligned movement organized around open and collaborative AI in the run-up to the AI Impact Summit.

That project was destroyed in 1945. The strategic geography beneath it did not disappear.

Japan, Taiwan, South Korea, Singapore, the Philippines, and the wider maritime region have since become more functionally integrated than the imperial planners of the 1930s could have achieved through occupation. Their governments do not answer to one capital. Their populations do not imagine themselves as members of one political civilization. Their industries nevertheless operate as different organs of the same production machine. The old sphere tried to make Asia serve Japan. The emerging one makes the world depend on capabilities concentrated across East Asia.

The resemblance ends there. Imperial Japan began with sovereignty and tried to manufacture integration underneath it. Postwar East Asia began with production, trade, specialization, investment, logistics, technical standards, and accumulated operational trust. Political unity never arrived. It never needed to.

The region became coherent without becoming one.

The Two Hierarchies

The American alliance system in Asia now contains two hierarchies. Security dependence runs toward Washington. Industrial dependence runs back through Tokyo, Seoul, Taipei, and the wider maritime corridor.

The first hierarchy is visible. The United States maintains alliances with Japan, South Korea, the Philippines, and Australia. It supplies extended deterrence, long-range military power, intelligence, logistics, nuclear guarantees, and the naval presence that prevents any regional power from easily converting economic scale into territorial command. Taiwan’s survival calculations remain inseparable from American capability and political will. Singapore is not a treaty ally, but its security environment also rests partly on the wider balance sustained by the United States.

The second hierarchy sits beneath the first. American technology and military power depend on fabrication, memory, materials, industrial machinery, batteries, shipbuilding, repair capacity, ports, cables, and logistical routines concentrated across the same states that Washington formally protects. The United States remains upstream in military force, software, finance, design, and strategic permission. East Asia has become upstream in execution.

This creates a relationship that conventional alliance language describes poorly. A traditional protectorate runs on asymmetric dependence: the large power guarantees the security of the smaller territory, and the smaller territory accepts limits on its strategic autonomy. Protection moves outward from the center. Dependence moves inward from the periphery. In maritime East Asia, those directions increasingly cross. The protected states remain militarily dependent while the protector becomes operationally dependent on their industrial ecosystems.

This is reversed protection. It does not make the allies equal to the United States, and it does not erase Washington’s ability to impose export controls, condition technology access, pressure firms, or alter alliance terms. It introduces a second hierarchy running in the opposite direction. A power can dominate the rules of a system while depending on others to keep the system physically alive.

Command and execution separate. The commanding power controls permission. The supposedly subordinate states control whether permission can become reality.

Reversed protection exists only under demanding conditions. The capability located in the protected state must be difficult to reproduce within the strategic timeframe. Its loss must degrade the protector’s military, technological, or economic power. That dependence must alter the protector’s behavior toward the jurisdiction that contains it. Industrial importance alone is insufficient. Indispensability becomes protection when the cost of losing a capability exceeds the cost of preserving the political space in which it operates. The first two conditions are already visible. The third, a measurable shift in how Washington protects, exempts, subsidizes, or defers to the jurisdictions it depends on, is the one still forming. Small signs exist: the Navy has begun bending its own rules to route ship overhauls through allied commercial yards. Where that behavior generalizes, reversed protection becomes real; where it does not, indispensability is only leverage waiting to be used. For now the concept names a tendency the system is entering, not a settled feature of the alliance.

Even then, protection is not guaranteed. Concentrated capability can attract coercion as easily as deterrence. It can persuade a protector to defend the node, or persuade the protector to move the capability somewhere safer. Indispensability buys exposure in the same motion that it buys leverage.

The Machine and the Corridor

The sphere named here is not the entire East Asian economy. It is also not a closed bloc. Three terms carry separate weight in what follows: the machine is the global production system; the corridor is the allied portion of it engineered to stay available under fracture; the sphere is the political order that mutual dependence begins to produce as the corridor’s connections harden into arrangements.

The global production machine includes China, Europe, the United States, Southeast Asia, and firms distributed across dozens of jurisdictions. Chinese processing, assembly, batteries, machinery, shipping, and intermediate goods remain embedded throughout it. Dutch lithography sits beneath advanced semiconductor fabrication. American design software, capital, cloud platforms, and chip architecture remain wired into the same system. No politically honest map can draw a clean line around the machine.

Inside that global machine, however, an allied continuity corridor is beginning to emerge. The corridor consists of jurisdictions expected to keep critical layers operating under coercion, blockade, industrial disruption, or strategic conflict. Economic participation does not establish membership. Three tests do: the node’s removal must damage several linked systems; coalition partners must be unable to replace it quickly; and its capacity must remain politically available during the crisis the corridor is designed to survive.

This separates ordinary supply-chain participation from strategic indispensability. A large exporter can remain replaceable. A smaller economy can become systemically central when several operating layers have accumulated around capabilities that cannot be moved on command.

The corridor has an East Asian core but extends beyond East Asian geography. Japan, Taiwan, and South Korea are execution nodes. Singapore is a coordination node. Australia provides resources, energy, land, and strategic depth. The Philippines provides territorial access across the maritime routes linking Northeast Asia, the South China Sea, and the Pacific. The United States supplies design, capital, software, military protection, and permission architecture. The Netherlands remains an external technological anchor because advanced fabrication still passes through one country’s lithography ecosystem.

Malaysia, Vietnam, Thailand, Indonesia, and other regional states also occupy important positions, and the boundary can move as their capabilities deepen. The corridor is not a diplomatic club with permanent seats. It is a functional topology. States enter its core when their removal would degrade several layers of allied continuity at once.

The distinction between the machine and the corridor also explains China’s position. China remains materially inside the global production machine. It sits outside the allied continuity corridor because its capacity cannot be assumed to remain available during the conflict against which that corridor is being organized. The boundary is strategic availability, not civilization, ideology, or latitude.

The same standard cuts against the corridor’s most important member. Taiwan is its execution center, yet Taiwanese capacity is the most likely to become unavailable in the exact conflict the corridor exists to survive. Leading-edge fabrication concentrated on an island under blockade or bombardment fails the availability condition in the scenario that matters most. This does not remove Taiwan from the corridor. It marks Taiwan as the place where indispensability and exposure peak together, the node the system can least afford to lose and least reliably keep. The corridor’s central asset is also its central vulnerability, and alliance language does not dissolve that. It can be hedged only partly, and only before a crisis rather than during one, by building redundancy off the island while the ecosystem still holds.

The machine is global. The corridor is the part being prepared to survive geopolitical fracture.

The Anatomy of Execution

Japan’s position is concentrated in the layers that disappear from public view precisely because they work. Japanese firms remain deeply embedded in semiconductor equipment, silicon wafers, photoresists, specialty chemicals, packaging materials, precision machinery, sensors, industrial components, and advanced manufacturing systems. Japan no longer dominates the finished-chip market as it once did. Its leverage migrated into the inputs and tools without which fabrication elsewhere becomes harder.

Taiwan became the execution center of advanced semiconductor manufacturing. Its importance cannot be reduced to fabrication plants. The Taiwanese ecosystem includes yield knowledge, supplier synchronization, advanced packaging, engineering discipline, failure recovery, and trusted relationships with the world’s largest technology firms. TSMC’s scale makes the point visible: in 2025 it reported manufacturing 12,682 products for 534 customers across hundreds of process technologies.^[2] The number matters because it reveals an operating environment, not simply a collection of factories.

South Korea occupies several adjacent layers: memory, high-bandwidth memory, electronics, batteries, displays, automobiles, steel, and high-value shipbuilding. As AI systems consume greater quantities of advanced memory alongside logic, Korean industrial capacity becomes part of the execution path from model design to functioning data center. Its shipyards also connect commercial manufacturing depth to the maritime capacity on which American power depends.

Singapore performs a different function. Its value is distributed across maritime services, aviation, finance, arbitration, corporate headquarters, data connectivity, and the routines through which regional actors route capital, ships, contracts, information, and trust. No single Singaporean object occupies the position of a leading-edge foundry. The node matters because too many systems use it to coordinate.

Australia and the Philippines show that industrial systems also require space. Australia’s minerals, energy, land, political alignment, and distance from the most exposed sections of the first island chain provide depth. The Philippines contributes ports, bases, cables, islands, and access across strategic sea lanes. One supplies resources and rear geography. The other supplies continuity across contested geography. Neither plays Taiwan’s role, but the machine cannot operate through fabrication alone.

These functions are different, and treating them as interchangeable would empty the concept of meaning. Execution nodes create or transform critical goods. Coordination nodes connect actors and transactions. Resource nodes supply upstream physical inputs. Access nodes preserve movement, repair, and territorial continuity. External anchors provide technologies or permissions the core cannot reproduce. The corridor becomes strategically significant because these functions link together.

This is not an empire. It is a machine without an emperor.

Three Tests of Reversed Protection

Semiconductors provide the clearest case. A processor designed in California may require fabrication in Taiwan, Japanese equipment and materials, Korean memory, advanced packaging, servers assembled across a wider electronics ecosystem, and logistics routed through regional ports and financial systems. American design leadership does not eliminate this dependence. It makes the dependence more consequential because the value of American intellectual property cannot become physical capability without the manufacturing environment beneath it.

TSMC’s expansion into Arizona demonstrates both sides of the mechanism. The company describes geographic expansion as a response to customer demand for flexibility and government support, and in 2025 announced additional American investment.^[3] New capacity reduces concentration risk and brings advanced production closer to American customers. It also reveals how difficult genuine relocation is. A fab can be financed and constructed faster than the supplier density, engineering routines, managerial knowledge, packaging capacity, and failure-recovery environment that make the fab productive can be recreated.

Arizona therefore does not simply replace Taiwan. It extends a Taiwanese capability into the United States. Over time, extension can become replication, and replication can reduce Taiwan’s leverage. The strategic unit is not the factory. It is the coordination environment around the factory. Reshoring changes the hierarchy only when the ecosystem moves with the asset.

Shipbuilding provides a second test. The United States retains extraordinary naval technology and military reach, but its commercial shipbuilding base and repair capacity have thinned while South Korea and Japan preserved deep maritime industries. This is no longer an abstract argument about comparative advantage. In March 2025, the USNS Wally Schirra completed a seven-month overhaul at Hanwha Ocean in South Korea. The work included dry docking, hull-corrosion repairs, a full rudder replacement, and more than 300 separate work items. The U.S. Navy described in-theater maintenance as a way to reduce downtime and enhance readiness.^[4] Japan crossed the same line weeks later. In April 2025, the USS Miguel Keith, an expeditionary mobile base, completed a five-month regular overhaul at Mitsubishi Heavy Industries in Yokohama, the first time a Japanese commercial yard won an overhaul contract of that scale for an American vessel.^[5] A U.S. statute normally bars major overhauls in foreign shipyards; the work qualified only because it ran under six months and the ship was not due back in the United States within fifteen.

These cases do not prove that America has outsourced its navy to East Asia. They show the direction of dependence. Two allied commercial yards, one Korean and one Japanese, absorbed major overhauls that American yards could not clear without lengthening their own backlog, and the Navy used an existing statutory opening to make that division of labor possible and freed its own workforce to let that happen. This is industrial capacity the United States no longer holds at scale at home. Korean and Japanese yards can shorten repair cycles and supply endurance that procurement orders alone cannot summon. The protector’s operational readiness begins to pass through the protected states’ yards.

The 2019 Japan-South Korea materials dispute provides the negative case. Japan tightened export procedures for fluorinated polyimide, photoresists, and high-purity hydrogen fluoride, inputs important to Korean semiconductor and display production. South Korea brought the dispute to the World Trade Organization.^[6] Two American allies occupying adjacent layers of the same industrial stack used their respective dependencies against one another.

The dispute exposed the limit of spontaneous integration. Distributed specialization had created mutual dependence, but no authority owned continuity across the system. The alliance map did not automatically govern the industrial machine beneath it. Political coordination failed at precisely the point where production had become strategically entangled.

The lesson is larger than the dispute. Indispensable nodes do not naturally form a resilient order. They can protect one another, coerce one another, or become points through which an external power fractures the system. Capability becomes continuity only after institutions decide how it will be allocated, protected, repaired, and restored under pressure.

China and Sovereign Closure

China offers a competing structure of industrial power. Its scale, domestic market, manufacturing depth, infrastructure financing, and state capacity allow more functions to be organized inside one political jurisdiction. Neighboring states trade with China, host Chinese investment, use Chinese equipment, and remain exposed to decisions made in Beijing. Yet China is not a single factory directed from a control room. It contains provincial competition, redundant industrial clusters, private tacit knowledge, and an enormous internal division of labor.

The allied corridor also contains centers of control. American export jurisdiction, design software, dollar clearing, cloud platforms, and military command can turn interdependence into permission. Japan and South Korea use industrial policy, subsidies, and export controls. Every major actor now combines markets with state power.

The decisive difference is sovereign closure. China’s sprawling industrial stack ultimately sits beneath one sovereign veto. The corridor’s critical layers remain distributed among governments that cannot command one another. Beijing can order. The corridor has to coordinate.

That difference creates opposing strengths. China can mobilize resources, establish priorities, and absorb duplication within a single authority. The corridor can preserve political variation, prevent one capital from owning the whole stack, and generate competing centers of capability. China gains coherence from command. The corridor gains survival value from vetoes that do not collapse into one.

Those vetoes are also a liability. Japan and South Korea can obstruct one another. Washington can use security pressure to reorganize commercial decisions. Taiwan cannot participate normally in many formal institutions. Southeast Asian states maintain different relationships with China. Domestic elections can reverse industrial policy or weaken alliance commitments. The absence of an emperor prevents domination, but it does not guarantee coordination.

No capital can rule the new sphere. That is its architecture and its central problem.

Distribution Is Not Resilience

The industrial corridor already possesses distributed capability. It does not yet possess distributed survivability.

Three structures are often collapsed into one. Concentration places several capabilities inside one jurisdiction. Distributed specialization places different indispensable functions in different jurisdictions. Distributed redundancy allows more than one jurisdiction to perform the same critical function. Only the third reliably converts geography into resilience.

The present system is dominated by distributed specialization. Taiwan fabricates leading-edge logic. South Korea produces critical memory. Japan supplies equipment and materials. The Netherlands provides advanced lithography. The United States controls major design tools, chip architectures, capital, and cloud demand. This arrangement prevents any one state from owning the entire system. It also means that failure at one irreplaceable layer can propagate through all the others.

Dispersion does not eliminate chokepoints. It can arrange them in sequence.

The political value of the corridor therefore does not come from pretending every country can reproduce every layer. That would be ruinously expensive and technically implausible. It comes from choosing which capabilities require duplication, which require protected inventories, which can be restored through allied surge capacity, and which must remain concentrated because the knowledge environment cannot yet be moved.

Redundancy should be selective. More fabrication in the United States, Japan, and Europe can reduce geographic concentration without dismantling Taiwan’s ecosystem. Additional packaging and memory capacity can limit cascades. Shared inventories of critical chemicals and components can buy time. Cross-certified repair facilities can turn Japanese and Korean yards into a common maritime reserve. Alternative ports, cables, data centers, and energy connections can prevent one territorial disruption from disabling coordination across the region.

The goal is not independence. It is distributed continuity.

That distinction also clarifies the danger of indiscriminate reshoring. If American subsidies move a factory from an ally into the United States without increasing total capacity, the system has changed ownership without gaining redundancy. If subsidies trigger parallel capacity while retaining the allied ecosystem, the corridor becomes harder to break. A fab transferred from Taiwan to Arizona can reduce allied leverage and leave overall resilience unchanged. A fab added in Arizona while Taiwan’s ecosystem remains dense can extend the system.

Industrial policy must be judged by the continuity it creates, not the flags placed above individual facilities.

The Historical Name Cannot Be Made Innocent

The Greater East Asia Co-Prosperity Sphere cannot be separated from occupation, forced labor, extraction, racial hierarchy, puppet governments, and mass violence. The phrase was designed to convert domination into solidarity at the level of language. Its use now is defensible only if the inversion remains explicit.

Imperial Japan recognized that the Western Pacific and maritime Southeast Asia formed a connected strategic space in which industry, shipping, raw materials, naval access, energy, and political control could not be separated. It interpreted that connection through the grammar of empire. Coordination meant command. Interdependence meant subordination. Security meant placing one capital above all others.

The method was part of the idea. Occupation could produce obedience but not trusted coordination. It could extract resources but not create the voluntary synchronization, distributed knowledge, and failure recovery on which complex production depends. The imperial sphere began with sovereignty and attempted to force cooperation underneath it. The system that emerged after 1945 developed in reverse.

Japan became more useful after it stopped trying to rule. Taiwan and South Korea became indispensable by developing capabilities no regional center could fully own. Singapore became powerful by coordinating systems it did not conquer. The alliance architecture endured because no single Asian state could convert its economic position into political mastery.

The historical sphere failed because it placed empire above interdependence. The emerging corridor works because interdependence prevents empire from completing itself.

From Supply-Chain Accident to Continuity System

The sphere does not yet exist as a conscious political order. It exists as overlapping dependence, partial coordination, and a growing recognition that economic security belongs inside alliance strategy. American and Japanese leaders now describe economic-security cooperation as an indispensable part of their alliance, and bilateral and trilateral initiatives increasingly address critical minerals, technology, supply chains, and industrial resilience.^[7] The language has moved ahead of the institutions.

A genuine continuity system would require a small number of concrete instruments.

First, the corridor needs standing allocation agreements for critical inputs. Shared inventories are insufficient if governments begin bidding against one another when a shortage arrives. Pre-agreed priorities, emergency release rules, and visibility into stockpiles would limit cascades across fabrication, memory, energy, and transport.

Second, trusted production must remain trusted across jurisdictions. Mutual recognition of industrial-security standards, export-control compliance, cybersecurity requirements, and supplier certification would allow capacity in one allied state to serve the others without a new political negotiation during every crisis.

Third, Japan and South Korea should be treated as parts of a common allied maritime-industrial base. Repair access, cross-certification, workforce exchanges, shared component inventories, and pre-negotiated emergency capacity would connect their commercial depth to American naval endurance. This does not require transferring command of the fleet. It requires recognizing that readiness already depends on an industrial geography wider than the United States.

Fourth, subsidy policy must stop treating allied capacity as foreign capacity. American, Japanese, Korean, Taiwanese, Australian, and European programs can create redundancy or cannibalize one another. Coordination should measure whether incentives add capacity, move it, or destroy a viable ecosystem somewhere else. National announcements count facilities. A continuity strategy counts the system that remains after the subsidies are spent.

Finally, the corridor needs an owner during stress. No single capital can command it, but designated institutions can map dependencies, run disruption exercises, assign restoration responsibilities, and maintain a common picture of the stack. The relevant model is not a supranational government. It is an operating compact among sovereigns that expect to disagree but refuse to discover their interdependence for the first time during a blockade or industrial shock.

These arrangements would produce visible consequences. American strategy would increasingly treat allied production as part of the defense base. Washington would protect indispensable ecosystems while simultaneously attempting to reduce dependence on them. Smaller allies controlling several linked capabilities would gain more influence than states holding one isolated commodity. Reshoring would alter political leverage only when coordination environments moved with factories. Crisis allocation and repair, rather than summit declarations, would reveal whether the corridor had become real.

The sphere will not be created by a treaty announcing its existence. It will appear when protecting the region and preserving the machinery of advanced civilization become the same operational task.

The Sphere That Protects Its Protectors

The twentieth-century map placed power in capitals. Washington protected allies. Tokyo commanded colonies. Beijing sought continental depth. Smaller states appeared as buffers, clients, bases, markets, or territory to be controlled.

The industrial map places power inside systems: fabrication, memory, machinery, energy, logistics, cables, shipyards, data centers, standards, finance, and repair. These systems remain governed by states, but they no longer fit cleanly inside the hierarchy of states. A militarily subordinate ally can become operationally indispensable. A commanding power can retain strategic authority while losing the capacity to execute alone.

East Asia occupies the center of this crossed hierarchy. Its states remain divided. Many remain militarily dependent on the United States. None can secure the regional order alone. Together, however, they hold enough of the industrial machinery that the formal protector increasingly depends on the protected.

The client becomes infrastructure.

Imperial Japan tried to build a sphere in which Asia would preserve Japanese power. It understood regional coherence through conquest, hierarchy, and command. The empire disappeared. A different coherence formed across the same strategic geography: production without political unity, integration without a sovereign center, and dependence running in both directions.

This order remains incomplete. The global machine includes actors the corridor may one day have to operate without. Its most valuable capabilities remain exposed. Its distributed specialization creates sequential chokepoints. Its governments can still turn dependence against one another. The system exists economically before it exists strategically.

That is precisely why its structure matters. The United States still protects the geography in which the East Asian machine operates. The East Asian machine increasingly preserves the operational capacity through which American power operates. The shield and the machine now protect each other.

The old sphere tried to escape the world by subordinating a region. The new corridor matters because the world cannot easily escape it.

America protects the map. East Asia protects the machine.

Notes

1. On the Co-Prosperity Sphere’s Pan-Asian rhetoric, extractive political economy, occupation systems, and wartime mobilization, see Gregg Huff, World War II and Southeast Asia: Economy and Society under Japanese Occupation (Cambridge University Press, 2020), and the established historical literature distinguishing the sphere’s public language from its governing reality.

2. Taiwan Semiconductor Manufacturing Company, 2025 Annual Report and 2025 Form 20-F, including the company’s reported process technologies, products, customers, advanced packaging, and manufacturing footprint.

3. TSMC, 2025 Annual Report, discussion of geographic flexibility and the March 2025 announcement of additional investment in the United States. The distinction between object scarcity and embedded execution is developed in Taiwan Must Build the Synthetic Shield.

4. U.S. Pacific Fleet, “USNS Wally Schirra Completes Major Maintenance at South Korean Shipyard,” March 13, 2025.

5. U.S. Ship Repair Facility and Japan Regional Maintenance Center (SRF-JRMC), “USS Miguel Keith Completes ROH at MHI,” April 2025, on the completion of the vessel’s five-month Regular Overhaul at Mitsubishi Heavy Industries, Yokohama, on April 15, 2025, described as the first overhaul contract of its scale awarded to a Japanese commercial shipyard; the restriction did not apply because the availability lasted under six months and the ship was not due to return to the United States within fifteen months.

6. World Trade Organization, Japan - Measures Related to the Exportation of Products and Technology to Korea, dispute DS590, initiated in 2019.

7. United States-Japan Joint Leaders’ Statement, February 2025, describing economic-security cooperation as an indispensable part of alliance cooperation; see also subsequent U.S.-Japan and U.S.-Korea arrangements on critical minerals, technology, and resilient supply chains.

The Missing Owner of the Decision

A human expert has traditionally done two things at once. The expert establishes that a method is reliable and adopts the resulting conclusion as their own opinion. Method and judgment arrive together, in one person who stands behind both.

Machine-generated evidence pulls those functions apart. An engineer can establish that a system is reliable, properly validated, and correctly operated, and still decline to say that its conclusion in the particular case is correct. The engineer vouches for the system. The machine produces the judgment. No one necessarily owns it.

My paper argued that when machine-generated evidence expresses a case-specific judgment, the party offering it should have to produce either a qualified person who genuinely adopts that judgment or a record detailed enough to reconstruct how it was produced. That is an evidentiary proposal. But the same gap will open across the economy, wherever production becomes synthetic while responsibility stays human.

The Answerability Economy

Most arguments about AI and employment ask whether a human is still required to perform a task, and increasingly the honest answer is no. AI may read the scan, assess the claim, screen the applicant, price the credit risk, flag the suspect, draft the audit, or determine eligibility for a benefit, and do it faster and more consistently than the professionals who used to.

But productive necessity is only one reason institutions employ people. They also employ people because decisions have to be authorized, certified, explained, appealed, attributed, and defended. None of those functions improves the machine’s output; they make the output usable inside legal and administrative systems that still assign rights and responsibilities through human beings and the organizations they control. The AI does not need those workers, but the institution does.

This is the answerability economy: a layer of white-collar work organized around converting machine output into decisions an institution can stand behind. It is not the same thing as checking whether the machine is right.[1] Verification asks whether an output is correct. Answerability asks who is entitled to adopt it, who must defend it, how it can be contested, and where its consequences land. An output can be perfectly verifiable and still have no authorized owner, and a person can be required to stand behind a judgment whose correctness no one can fully confirm. What the institution needs is not another check on accuracy but a recognized point of attribution.

Its roles are already taking shape. There are people who review machine judgments and assume responsibility for them; auditors who test whether a system performs and catch it failing systematically; staff who preserve the inputs, model versions, and prompts behind an output; specialists who reconstruct how a particular result was produced; officers who hear appeals of automated decisions; investigators who assign responsibility after a failure; professionals who certify systems for specified uses; and expert witnesses who explain and defend machine conclusions in court. These workers occupy the space between machine execution and institutional consequence. The system produces the answer; the human clears it for use.

White-Collar Work Moves Toward Permission

The pattern repeats across nearly every consequential sector. In medicine, AI reads the images and recommends the treatment while a physician reviews and adopts the conclusions that carry the hard cases. In insurance, automated systems resolve routine claims while human reviewers certify the high-consequence denials and handle the contested classifications. In employment, models rank applicants and flag dismissals while someone decides whether those conclusions can be legally defended. In banking, AI underwrites while compliance staff reconstruct adverse decisions and confirm that prohibited factors did not drive the outcome. In government, automated systems calculate benefits, taxes, and enforcement priorities while administrative reviewers keep a channel open for appeal and attribution. In litigation, the machine produces the forensic conclusion while one expert establishes the system and another decides whether to stand behind the result.

The profession does not survive as a single role. Its function splits: execution moves into the machine, authorization stays attached to the institution, and human labor gathers along the boundary between them. White-collar work shifts from production toward permission.

Accuracy Does Not Produce Authority

A system more accurate than any available human can still generate this work, because accuracy is not the same thing as authority. Institutions have to answer the individual case in which the system may have failed. Fairness is the clearest reason, though not the only one; due process, professional licensing, insurance, administrative finality, and the plain need to place liability all generate the same demand. A person denied a job, their liberty, medical coverage, or a public benefit does not experience the model’s overall accuracy rate; they experience one decision, and the law gives them a way to challenge and reconstruct it. That is why automation can produce jobs that look economically unnecessary. The capability to reach the answer already exists. The extra worker exists because consequential authority cannot yet travel directly from a machine output to a legal effect. The job is produced by an institutional constraint, not a productive one.

Modern economies already run large versions of this layer. Compliance officers, auditors, inspectors, claims reviewers, licensing authorities, and administrative judges do not manufacture the underlying goods; they make complex systems governable, legible, and contestable. AI expands this function even as it thins the labor that performs it. Every automated system that approaches consequential authority creates a surrounding demand for certification, provenance, appeal, and the allocation of liability. The more execution moves into machines, the more an institution has to specify what turns an output into an authorized decision.

Legitimacy is labor-intensive.

Real Review, or Review in Name Only

Not every human placed inside an automated process will actually exercise judgment. Organizations will be tempted to build oversight that exists only on paper: an employee receives the machine’s recommendation, clicks approve, and becomes the official decision-maker the institution can point to as proof that a human stayed in control.

Substantive adoption takes more than a signature. It requires understanding the material inputs, evaluating the conclusion, weighing plausible alternatives, and being willing to present the judgment as your own. A worker without the authority, competence, time, or access to the underlying record to reject the machine cannot supply that; they are an interface placed in front of a decision formed elsewhere. The title stays human while the judgment has already moved upstream.

This distinction decides whether answerability work becomes a genuine profession or just a mechanism for pushing liability downward. Some workers will hold real authority to overrule a machine conclusion. Others will absorb responsibility for systems they do not control, which is not oversight at all, only a place to send the blame.

More Job Titles Is Not More Jobs

The answerability economy can multiply categories of white-collar work without producing enough employment to replace what automation removes. One adopting professional can supervise thousands of machine-generated decisions. One audit team can govern systems that replaced whole departments. Appeal officers handle only the small fraction of automated decisions that are ever formally contested. The layer can grow in importance, spending, and authority while employing fewer and fewer people. The titles proliferate while total employment contracts.

This is not the comfortable claim that AI will create as many jobs as it destroys. It is a narrower point about where some of the new work comes from and why institutions keep employing people after machines can perform the underlying tasks. Production can keep growing while the income it used to distribute shrinks, and labor loses its position as the main channel through which income, status, and a recognized place get handed out. Answerability work can slow that displacement, but it will not reverse it.

The Jobs of the Transition

These jobs belong to the interval between two orders. In the old order, humans perform the work, form the judgment, and carry responsibility for the result. In the emerging order, machines perform the work and increasingly form the operative judgments, while responsibility stays attached to human beings and the legal entities they run. A more distant order might recognize machines themselves as holders of authority, assets, duties, and liability. If that arrives, some of this layer could thin. But legal personhood is not a solvent: corporations have held it for centuries and still need officers, auditors, and witnesses to answer for them. Machine personhood would more likely move the answering around than end it.

During the transition, courts will keep demanding witnesses, regulators will keep demanding accountable officers, citizens will keep demanding a way to appeal, insurers will keep demanding someone who can assume liability, and organizations will keep demanding the signature that converts machine output into authorized action. AI will remove the worker who made the decision. Law, administration, and fairness will create another worker whose job is to make that decision answerable. That worker will not operate the machine. That worker will stand between the machine and its consequences.

The machine makes the decision. A person is kept to answer for it.

Notes

[1] The nearest economic account is Christian Catalini, Xiang Hui, and Jane Wu, Some Simple Economics of AGI (Feb. 24, 2026), which argues that as machine execution becomes cheap, the binding constraint shifts to human verification, with rents migrating to validation, provenance, and liability underwriting. Answerability, as used here, is the adjacent but distinct function: not confirming that an output is correct, but owning and defending it once it acquires institutional effect.

Microsoft and OpenAI are a different species of power.

They do not enter through the battlefield. They enter through the document, the inbox, the spreadsheet, the hospital note, the cloud contract, the procurement system, and the classified deployment environment. Palantir asks the republic to accept the stack as hard power. Microsoft and OpenAI teach institutions to experience the stack as ordinary work.

The distinction is one of emphasis, not a clean line. Palantir also runs through civilian and administrative workflows, and Microsoft and OpenAI are now deep inside defense and intelligence. But the center of gravity differs. Palantir’s terrain is action. Microsoft and OpenAI work one layer earlier, on the cognition that runs before action: what intelligence is available before coordination, targeting, and execution begin.

That earlier layer is the whole game. The Microsoft-OpenAI model is not simply another case of sovereignty migrating into infrastructure. It is the emergence of something quieter and broader, an access regime for machine intelligence.

The question is no longer only who owns the model, who controls the cloud, or who profits from enterprise AI. The more important question is who decides how intelligence itself is distributed: who receives it, at what tier, under what permissions, with what audit requirements, through which interface, and under whose terms.

The state keeps the authorization. The partnership increasingly runs the execution layer. And the mission language explains why this is not capture but progress.

The Access Regime Is Not the Security State

The access regime is the emerging order in which access to machine intelligence becomes tiered, permissioned, priced, audited, and differentially distributed across society.

It is not the welfare state, which allocates benefits. It is not the administrative state, which allocates rules. It is not the security state, which allocates suspicion and force. The access regime allocates cognition.

It decides which actors reach machine reasoning, summarization, prediction, and decision support, and which institutions receive consumer, enterprise, audited, classified, or sovereign deployment.

A free-tier user receives one version of intelligence. A paid subscriber receives another. An enterprise customer receives a different set of capabilities, privacy guarantees, and data-retention terms. A government agency receives another configuration again, and a classified user inside a defense or intelligence environment receives yet another, deployed through restricted infrastructure and surrounded by authorization, auditing, and compliance.

This is usually described as product segmentation. That framing is too small. When intelligence becomes a general-purpose input into institutional action, product segmentation becomes governance. It decides who can think with machines, at what level, and under what constraints.

The regime does not announce itself as a new order. It appears as pricing plans, compliance tiers, cloud regions, model-access rules, procurement certifications, and enterprise dashboards. Power no longer needs to appear as command. It can appear as access.

The Partnership Is an Access Architecture

The Microsoft-OpenAI partnership is usually read as a business story: the multibillion-dollar commitment, the cloud arrangement, the integration of OpenAI models into Microsoft products. That story is accurate and insufficient.

The more important fact is structural. OpenAI supplies frontier model capability. Microsoft supplies cloud infrastructure, enterprise distribution, government authorization pathways, security certifications, and the productivity interface through which much of the institutional world already works. Together they do not simply sell AI. They build the architecture through which access to AI is granted, restricted, monitored, and normalized.

When Microsoft announced its third investment phase in OpenAI in January 2023, the formal language was partnership. The structural reality was more specific: Azure became the exclusive cloud for OpenAI workloads, OpenAI’s compute ran on Microsoft’s infrastructure, and Microsoft’s products ran on OpenAI’s models. The dependency was mutual and deliberate. [1]

The relationship has since hardened into something closer to constitutional structure, then loosened again without losing its core. The October 2025 restructuring converted OpenAI’s commercial arm into a public benefit corporation and left Microsoft holding roughly 27 percent of OpenAI Group PBC, an investment the company valued at about 135 billion dollars. The April 2026 amendment then relaxed the exclusivity: OpenAI can now serve its products across any cloud, while Microsoft remains its primary cloud partner, OpenAI products ship first on Azure under defined conditions, and Microsoft keeps a non-exclusive license to OpenAI model and product IP through 2032. [2]

The structural point survives every amendment. Microsoft is a major shareholder, primary infrastructure partner, distribution channel, and enterprise interface for OpenAI-derived intelligence. The deeper return was position: Microsoft placed itself between institutional demand for intelligence and the infrastructure required to deliver it, and OpenAI gained one of the most important enterprise and government distribution channels in the world.

The result is a hybrid that does not fit existing regulatory categories. OpenAI keeps its mission governance and nonprofit parent; Microsoft holds commercial deployment rights, cloud priority, and the productivity layer. Neither fully controls the other, and neither is directly accountable to the institutions that increasingly depend on their combined system. This is not a software company and its cloud vendor. It is a distributed access architecture with no sovereign above it.

Azure Is the Control Plane

It helps to stop thinking of cloud as storage or compute. In the AI age, cloud is the control plane for permissioned intelligence. It determines where models run, what data they touch, which institutions can deploy them, what audit trails exist, and which users reach which capabilities.

This is most visible in government. In May 2024, Microsoft deployed GPT-4 to an isolated, air-gapped Azure Government Top Secret cloud for the Department of Defense. [3] By September 2024, Azure OpenAI had received FedRAMP High authorization and Defense Information Systems Agency approval for DoD Impact Levels 4 and 5. [4] In January 2025, GPT-4o was authorized for use at the intelligence community’s top-secret level under Intelligence Community Directive 503. [5] By April 2025, Azure OpenAI had been authorized across all U.S. government data classification levels, from unclassified through the Secret-tier Impact Level 6. [6]

Within roughly a year, OpenAI models moved from enterprise productivity tools to authorized infrastructure for classified defense and intelligence work at every classification level the government recognizes.

This did not happen only because Microsoft is powerful, though it is. It happened because the institutions adopting Azure OpenAI had a coordination problem existing systems could not solve. Agencies, military commands, and intelligence bodies produce more information than they can integrate, and they need systems that surface signals, summarize fragmented data, and deliver outputs into operational workflows at machine speed. Azure offered the path of least resistance: secure cloud, model access, compliance history, procurement familiarity, and a vendor already embedded across the enterprise state.

The certifications are not theater. FedRAMP High and the DISA authorizations are serious security processes, and Microsoft earned them. They do not prove dependency. They create the conditions under which dependency becomes institutional. Once an agency deploys Azure OpenAI at high classification levels, its workflows, training cycles, procurement habits, and institutional routines begin reorganizing around that infrastructure. The question stops being whether the system is safe enough and becomes whether the institution can function without it. The state certifies the infrastructure, and the infrastructure becomes part of the state’s capacity.

Copilot Is the Civilian Interface

Azure is the substrate; Copilot is the interface. This is where the model departs from harder forms of state-adjacent AI. Palantir is visible as power because it enters domains where power already appears nakedly. Microsoft enters through routine.

The meeting needs a summary. The doctor needs a note. The analyst needs a draft, the manager a report, the inbox a triage. Nothing feels constitutional. Everything feels useful. That is the genius of the model.

By late 2024, Microsoft reported that nearly 70 percent of the Fortune 500 was using Microsoft 365 Copilot. [7] A UK government trial of 20,000 civil servants reported average savings of 26 minutes a day, though that number was self-reported, and other government trials, one with a control group, found smaller gains or none at all. [8] The exact figures are contested. The direction is not: the interface is arriving inside institutional work at scale.

The productivity framing is true. It is also a retail explanation. The structural reality is that Copilot changes the surface through which institutions think. It does not only save time. It changes what gets surfaced before human judgment arrives. It drafts the document that gets revised rather than written, summarizes the meeting that will be remembered instead of the meeting that occurred, and runs the compliance check that decides whether an action proceeds. Each is a small relocation of cognition.

Before Copilot, a human decided what to look at, what to include, what to omit, and how to summarize. After Copilot, those choices are increasingly made first by a system whose parameters were set in Redmond and San Francisco, whose behavior was fixed by training and alignment processes no ordinary institution reviewed, and whose outputs arrive already formatted as helpfulness. The institution still decides. But the field of decision has already been shaped.

This is why Copilot matters beyond productivity. It is one of the most widely deployed civilian interfaces between private AI infrastructure and institutional cognition. The danger is not that Microsoft owns the state. It is that institutional cognition begins to pass through Microsoft’s interface before anyone experiences that as governance.

Tiered Access Is Governance

OpenAI’s deployment architecture is explicitly stratified, from the free consumer tier up through enterprise, government, and classified deployments, each with its own controls, data protections, and audit rules. [9]

The scale is easy to miss because the floor is so familiar. Roughly 900 million people now use ChatGPT every week. [10] They occupy the consumer level of an architecture whose upper tiers reach into air-gapped classified environments. One structure. Nine hundred million people at the bottom of it, a few thousand cleared analysts at the top.

That image is a gradient of permission and governance, not a single chain of command, and not a simple ladder of raw capability. The cleared analyst’s system is not necessarily smarter than the consumer’s; authorization is slow, and the classified tier may run older or more restricted models. What differs across the tiers is what each is allowed to do with machine intelligence, under what oversight, and against what data. The student and the cleared analyst do not answer to the same officer. They draw on related model families inside one connected commercial architecture, permissioned differently at each level.

That pricing tiers are themselves a political technology, that formal access can be broad while effective access stays narrow, is the argument of a companion essay. The point here is narrower and architectural: the same model family, governed differently at each tier, produces categorically different institutional objects. The Pentagon analyst running GPT-4o in a classified environment is not using the same thing as the student on a consumer chatbot. The model may be related. The governance context is not.

The architecture decides who gets what mind.

No single actor allocates all of this. The distribution is the product of many hands: customer demand, Microsoft’s pricing, OpenAI’s policy, government procurement, security rules, and rival vendors. But the aggregate behaves like an allocation. The sum of these choices, none of them a public decision, determines who reaches which capability. That is how access regimes form: not through a constitutional convention, but through product architecture. By the time the political system recognizes that access to machine intelligence has become a public question, the distribution may already be embedded.

Every such regime also produces a residual: actors who cannot afford the tier, cannot satisfy the compliance requirements, or receive only a degraded version of machine cognition. In the access regime, intelligence is not equally available. It is permissioned.

The Therapeutic Shield

Every transfer of power into private infrastructure needs a legitimacy argument. Palantir’s is martial: Western survival, democratic hard power, civilizational defense. Microsoft and OpenAI use a different one, therapeutic and humanitarian. AI will empower workers, save time, reduce burnout, return clinicians to patients, and benefit all of humanity.

This language is not false, which is why it works. Copilot does save time. Azure does let institutions reach capabilities they could not build. The benefits are real. But a shield does not have to be false to be structural. It takes a complex transfer of power into infrastructure and makes it narratable: dependency reframed as empowerment, automation as augmentation, unequal access as innovation rather than allocation.

OpenAI’s mission language does the same work at the civilizational level. The company was founded to ensure that artificial general intelligence benefits all of humanity, and every subsequent governance turn, the capped profit, the restructuring fights, the public benefit corporation, has been narrated as preservation of that mission. The point is not that the mission is fake. It is that mission language helps stabilize a structure in which accountability is hard to locate. The mission says the partnership is accountable to humanity. The structure says no specific public institution governs the access regime. That is the contradiction.

Dependency Arrives as Convenience

A hospital that deploys clinical documentation AI does not experience itself as surrendering capacity. It experiences less administrative burden. An agency that deploys Azure OpenAI does not experience outsourced sovereign cognition. It experiences faster analysis. Dependency does not announce itself. It accumulates as convenience.

An institution can change its CRM, migrate email with enough pain, replace a project tool. An institution that has reorganized its analytical workflows around a specific AI interface faces a different problem. The interface has not merely stored its information. It has shaped how the institution sees. That is not only vendor lock-in. The deeper risk is that the institution forgets how to work without the interface.

And the interface is not neutral. A meeting summary decides what mattered. A document draft sets the first structure of an argument. A compliance assistant encodes a model of risk. Each output can become the first version of reality the institution sees, and institutions are path-dependent: the initial summary becomes the shared memory, the first draft becomes the basis for revision. Human review remains present, but review is not authorship. A person editing a machine-generated summary is not in the position of a person who decided from scratch what the meeting meant. The human stays in the loop. The loop has changed.

The access regime does not need to replace human judgment. It only needs to supply the environment in which judgment operates.

Where Accountability Goes When Allocation Fails

In October 2023, New York City launched a chatbot called MyCity, built on Microsoft’s Azure AI cloud and trained on the city’s own regulations, to help small businesses navigate the rules. Within months, an investigation by The Markup found it telling business owners they could take a cut of workers’ tips, that landlords could turn away tenants with housing vouchers, and that a business could refuse to accept cash, each of them illegal under New York law. [11] Asked identical questions, it returned different answers to different users.

What happened next is the part that matters. The mayor acknowledged the bot was wrong in places and declined to take it down. [12] The city added a disclaimer telling users not to treat its answers as legal advice. Microsoft said it was working on accuracy but would not say what was causing the errors. The bot stayed live for nearly two more years, until a new administration shut it down as unusable. [13]

Notice where the accountability went. The judgment ran through a private system the city did not build and could not fully explain. The wrong answers could have entered at the model, the retrieval design, the system prompt, the source material, or the contractor’s implementation, and the city could not publicly say which. When it failed, the vendor pointed at ongoing fixes, the city pointed at a disclaimer, and the wrong answers stayed online. No one owned the output. The disclaimer, which told citizens to check the government’s own tool against the government’s own website, is the legitimacy wrapper in its purest form: the human is told to stay in the loop precisely so the institution does not have to answer for the machine.

The reflex has a cleaner statement in a Canadian tribunal case: an airline argued it could not be liable for its chatbot’s wrong fare advice because the bot was, in its telling, a separate legal entity responsible for its own actions. [14] The tribunal rejected it, but the argument is the tell. When an automated system produces a consequential error, the deploying institution’s first move is to place the authorship somewhere else.

This is what an earlier essay called the allocation state: the public institution held accountable for judgments produced in systems it does not govern. The access regime supplies the cognition; the allocation state inherits the liability.

The same shape recurs higher up, where it is harder to see. Consider the chain behind a government analyst using Azure OpenAI in a classified environment. OpenAI trained the model. Microsoft deployed and secured it. DISA and other bodies authorized the environment. A contracting officer approved procurement, an integrator built the workflow, an analyst used the tool, and a decision-maker acted on the output. If that system surfaces incomplete intelligence or shapes a decision badly, each link points to the next, and none is answerable for the whole.

A classified analytical system differs from MyCity in almost every way: its users, its data, its oversight, its stakes. It adds one more difference that matters here. No outsider can run the test. A journalist could type questions into a public chatbot and publish the wrong answers; no one can do that to a classified system. The conditions for the same diffusion of accountability exist there, in the tiers where the stakes are highest, precisely where it cannot be seen. Security review can confirm that a system meets technical controls. It cannot determine how machine-generated cognition reshapes institutional judgment over time. The institution that outsources part of its perception may also outsource its ability to know what it is missing. That is the failure no one is ready to govern.

The State Certifies What the Stack Has Already Made Necessary

Regulators are not absent. The EU has scrutinized Microsoft’s bundling. Antitrust authorities have examined the partnership. State attorneys general have watched OpenAI’s restructuring. Defense and intelligence authorizations required serious review. The state is present. But it arrives downstream. It reviews contracts after dependency has begun, certifies security after architecture is designed, and investigates market power after the product is embedded.

Deployment moves at the speed of enterprise procurement. Regulation moves at the speed of law. The stack moves first, and the state arrives later to certify, regulate, or legitimize a system that has already become useful. The access tiers were designed before the political theory caught up. The state does not lose sovereignty in name. It loses the operational independence that sovereignty needs in order to mean anything.

Does This Survive Fragmentation

The strongest objection is that this regime is already breaking apart. Open-weight models are catching up. Rival clouds compete for the same workloads. Sovereign-cloud and data-localization rules push deployment onto national infrastructure. And the April 2026 amendment freed OpenAI to serve its products on any cloud, ending Azure’s exclusivity.

These are real, and some cut the right way: competition and open weights can lower prices and weaken any single vendor’s grip. What they do not do is make the allocation equal, transparent, or publicly governed. The April 2026 amendment is the clearest case: OpenAI’s intelligence can now be served across Azure, AWS, and others, which does not flatten access so much as add a layer of tiering between clouds. A world with three frontier labs, four hyperscalers, and a dozen sovereign stacks is not a world in which machine intelligence is equally available. Fragmentation changes who holds the toll, and how many tolls there are. It does not remove them.

The Access Regime Enters the IPO Pipeline

In early June 2026, the access regime entered the IPO pipeline. Anthropic confirmed on June 1 that it had confidentially submitted a draft S-1 to the Securities and Exchange Commission. Within the week, OpenAI announced its own confidential filing, noting that timing remained undecided. [15]

The financial press asked the financial questions: whether revenue supports the valuations, whether the margins exist, when the burn ends. The constitutional question is different. If these companies proceed to public listings, the firms helping allocate machine cognition to states, enterprises, hospitals, schools, and ordinary users acquire a new governing constituency: public shareholders.

A listing changes who a company answers to. A public benefit corporation’s directors owe their duties to the corporation and are bound to weigh its mission alongside profit, so the mission language can survive on paper. But the daily gravity of a listed company is disclosure, growth, guidance, and the share price, and that gravity pulls in one direction. The access regime may go public in precisely the wrong sense, owned by the public as investors before it is governed by the public as citizens.

The infrastructure math explains the pull. In late 2025, Sam Altman described roughly 1.4 trillion dollars in infrastructure commitments over eight years. By February 2026, OpenAI was reportedly telling investors that it expected around 600 billion dollars in compute spending through 2030. The number fell; the dependency did not. A company operating at either scale has to keep raising, and to keep raising it has to satisfy the markets that supply the capital. [16] The allocator of intelligence becomes, structurally, an applicant to capital. The listing does not corrupt the access regime. It completes it: every tier now has a price, including the company’s own.

The Constitutional Question

Microsoft and OpenAI are the clearest present example of a structure that will not stay unique to them. Wherever frontier AI requires massive compute, safety evaluation, and government authorization, a small number of private actors will sit between institutional demand and machine intelligence. Their power will not come only from owning models. It will come from controlling access.

The future is not one universal intelligence system available equally to everyone. It is a layered regime shaped by capability, price, data protection, security clearance, and regulatory permission. That makes the distribution of machine cognition one of the central political-economic questions of the age. The politics of AI will not only be about whether models are aligned. It will be about who receives which aligned model, under what conditions, and through whose infrastructure. A society that distributes machine intelligence unequally is not simply adopting a technology. It is building a cognitive class structure.

The constitutional question of the AI age may not begin with weapons, borders, or surveillance. It may begin with something quieter: who allocates intelligence, and who governs the interface before the decision is made.

By now the pieces have names: the substrate, the interface, tiered access, the therapeutic shield, the dissolving of accountability, and the discipline of the capital markets. Together they form the soft infrastructure through which institutions think.

This is not capture in the old sense, a company seizing public authority from outside. It is subtler: a private-public intelligence regime becoming necessary to how institutions function. The state still speaks. The institution still decides. The human stays in the loop. But the intelligence available before the decision is increasingly allocated through private infrastructure, and the access regime does not need to own the decision. It only needs to govern what can be seen, summarized, ranked, and drafted before the decision is made.

The question was never who owns the model. It was who allocates the mind.

Notes

[1] Microsoft’s January 23, 2023 announcement of its third investment phase in OpenAI named Azure as the exclusive cloud for OpenAI workloads across research, products, and API services. Microsoft, “Microsoft and OpenAI extend partnership,” Official Microsoft Blog, January 23, 2023: https://blogs.microsoft.com/blog/2023/01/23/microsoftandopenaiextendpartnership/

[2] Microsoft’s investment in OpenAI has totaled roughly 13 billion dollars since 2019. Under the October 28, 2025 restructuring, OpenAI’s commercial operations became a public benefit corporation and Microsoft’s stake in OpenAI Group PBC was valued at about 135 billion dollars, roughly 27 percent on an as-converted basis. On April 27, 2026 the companies amended the partnership: OpenAI may serve products on any cloud, Microsoft remains primary cloud partner with OpenAI products shipping first on Azure under defined conditions, and Microsoft retains a non-exclusive license to OpenAI model and product IP through 2032. Microsoft, “The next chapter of the Microsoft-OpenAI partnership,” October 28, 2025: https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/ and “The next phase of the Microsoft-OpenAI partnership,” April 27, 2026: https://blogs.microsoft.com/blog/2026/04/27/the-next-phase-of-the-microsoft-openai-partnership/

[3] Microsoft deployed GPT-4 to an isolated, air-gapped Azure Government Top Secret cloud for Department of Defense use, announced May 7, 2024. DefenseScoop, “Microsoft deploys GPT-4 large language model for Pentagon use in top secret cloud,” May 7, 2024: https://defensescoop.com/2024/05/07/gpt-4-pentagon-azure-top-secret-cloud-microsoft/

[4] In September 2024, Azure OpenAI Service was approved within the FedRAMP High authorization for Azure Government and within the DISA DoD Impact Level 4 and 5 provisional authorization. Microsoft Azure Government blog, “Azure OpenAI, including GPT-4o, approved as a service within the FedRAMP High Authorization,” updated September 3, 2024: https://devblogs.microsoft.com/azuregov/azure-openai-fedramp-high-for-government/

[5] GPT-4o was authorized for use in Azure Government Top Secret under Intelligence Community Directive 503, announced January 16, 2025. DefenseScoop, “OpenAI’s GPT-4o gets green light for top secret use in Microsoft’s Azure cloud,” January 16, 2025: https://defensescoop.com/2025/01/16/openais-gpt-4o-gets-green-light-for-top-secret-use-in-microsofts-azure-cloud/

[6] Microsoft Azure Government blog, “Azure OpenAI Service now authorized for all U.S. Government data classification levels,” April 16, 2025: https://devblogs.microsoft.com/azuregov/azure-openai-authorization/

[7] Microsoft, “Ignite 2024: Why nearly 70% of the Fortune 500 now use Microsoft 365 Copilot,” Official Microsoft Blog, November 19, 2024: https://blogs.microsoft.com/blog/2024/11/19/ignite-2024-why-nearly-70-of-the-fortune-500-now-use-microsoft-365-copilot/

[8] UK Government Digital Service, “Microsoft 365 Copilot Experiment: Cross-Government Findings Report,” trial of about 20,000 civil servants across 12 organizations, September to December 2024, published June 2025; findings summarized in a UK parliamentary written statement, June 2, 2025: https://questions-statements.parliament.uk/written-statements/detail/2025-06-02/hlws667. The 26-minute figure is self-reported; a later Department for Work and Pensions study using a control group found a 19-minute average, and a Department for Business and Trade trial found no clear productivity gain.

[9] OpenAI’s standard API tier retains inputs and outputs for up to 30 days by default; eligible enterprise customers and endpoints can receive Zero Data Retention; classified government deployments on Azure operate under separate access controls and authorization frameworks.

[10] OpenAI and subsequent reporting placed ChatGPT at roughly 900 million weekly active users in 2026.

[11] Colin Lecher, “NYC’s AI Chatbot Tells Businesses to Break the Law,” The Markup, co-published with Documented and THE CITY, March 29, 2024: https://themarkup.org/artificial-intelligence/2024/03/29/nycs-ai-chatbot-tells-businesses-to-break-the-law. The MyCity chatbot was built on Microsoft’s Azure AI cloud.

[12] Colin Lecher, Katie Honan, and Maria Puertas, “Malfunctioning NYC AI Chatbot Still Active Despite Widespread Evidence It’s Encouraging Illegal Behavior,” The Markup and THE CITY, April 2, 2024: https://themarkup.org/artificial-intelligence/2024/04/02/malfunctioning-nyc-ai-chatbot-still-active-despite-widespread-evidence-its-encouraging-illegal-behavior. Mayor Adams acknowledged the errors at an April 2, 2024 press conference and left the chatbot online.

[13] Colin Lecher and Katie Honan, “Mamdani to Kill the NYC AI Chatbot We Caught Telling Businesses to Break the Law,” The Markup and THE CITY, January 30, 2026: https://themarkup.org/artificial-intelligence/2026/01/30/mamdani-to-kill-the-nyc-ai-chatbot-we-caught-telling-businesses-to-break-the-law

[14] Moffatt v. Air Canada, 2024 BCCRT 149 (British Columbia Civil Resolution Tribunal). The tribunal held Air Canada liable for inaccurate fare information provided by its website chatbot and rejected the airline’s argument that the chatbot was a separate legal entity responsible for its own actions: https://canlii.ca/t/k2spq

[15] Anthropic, “Anthropic confidentially submits draft S-1 to the SEC,” June 1, 2026: https://www.anthropic.com/news/confidential-draft-s1-sec. OpenAI announced its own confidential S-1 submission in early June 2026, noting that timing remained undecided.

[16] OpenAI’s annualized revenue crossed 20 billion dollars in 2025. Reporting has placed OpenAI’s total infrastructure commitments as high as 1.4 trillion dollars.

The state outsourced judgment but retained liability.

The modern state justified itself as the institution of final decision.

It would decide who received a pension and who did not. It would decide which businesses could operate, who crossed the border, who received the subsidy, who qualified for the treatment, who was admitted to the public housing waitlist, who got the license, who faced the sanction. Decision authority was the substance of sovereignty. To govern was to allocate: to determine, through visible law and accountable human procedure, who received what from the collective systems that citizens paid for and depended on.

Something fundamental has shifted in that arrangement. Not because states are retreating. States are, by many measures, larger and more present in daily life than at any prior point in modern history. They spend more, regulate more, intervene more, and monitor more than their predecessors did. The apparatus has not shrunk.

What has changed is where the decisions actually happen.

More and more, the acts that determine whether a person receives credit, whether an application is approved, whether a benefit is granted, whether a vendor is eligible, whether a student qualifies, whether a worker is hired, whether a defendant is detained, whether a patient is prioritized are not made by a civil servant sitting across a desk. They are generated by systems: scoring models, eligibility algorithms, risk engines, procurement filters, hiring platforms, diagnostic tools, classification architectures, and compliance layers that process inputs and produce outputs before any human official acts.

The state no longer always decides.

It certifies, audits, supervises, and absorbs the consequences.

That is the allocation state: the political form that emerges when governance functions have migrated into technical systems, but the liability and legitimacy of those functions remain publicly assigned to the state. The allocation state is not the state that decides everything. It is the state that must justify decisions made elsewhere. The threshold is not the mere use of software. It is reached when the state can no longer independently reconstruct or readily override the systems that produce its allocations, yet remains the institution the affected person must ask for an explanation.

This is not merely algorithmic government. Algorithmic government describes a tool. The allocation state describes a position. The tool changes what the state does. The position changes what the state is.

From Fiscal Crisis to Institutional Crisis

The previous essay in this series described the state’s fiscal predicament after labor thins: the payroll system that once made citizens legible to tax authorities is losing coverage as production reorganizes around AI, and states must search for new revenue bases in an economy whose surplus increasingly accumulates in structures they cannot easily reach.

The fiscal crisis is only one layer of what is happening to the state.

The deeper institutional crisis runs underneath it. The state is not only struggling to fund itself from an economy that has moved. It is also losing direct command over its own allocation functions. Fiscal capacity and allocative capacity are related but not identical. A state can in principle solve the revenue problem while the allocation problem deepens, finding new ways to extract from AI-generated surplus while the systems through which it governs become progressively less its own.

The allocation state names that second, quieter transformation.

The State Does Not Disappear. It Changes Position.

The story of AI and government is usually told in one of two registers.

The first is fear: surveillance states, automated repression, algorithmic control, the dystopian merger of state power and machine speed. The second is aspiration: digital government, AI-assisted services, efficient public administration, faster and more accurate decision-making at scale. Both registers share a common premise: the state remains the primary actor in its own governance functions. One version is empowered by AI; the other is corrupted by it. In both, the state is still the entity doing the governing.

Working through what is actually happening requires setting that premise aside.

What is underway is neither the empowering nor the enslaving of the state. It is the repositioning of the state from the seat of allocation to the frame around it. The state provides the mandate, absorbs the legitimacy demands, certifies the process, manages the appeals, backstops the failures, and defends the outcome in court. But the outcome was produced elsewhere.

This is the key structural distinction: the visible sovereign and the hidden execution layer have separated. Citizens see the state and hold it accountable. The judgment they are holding it accountable for was generated somewhere the state does not fully control.

This is a change in constitutional position, not only in administrative technique. Constitutions describe the state as the repository of public authority: the entity that taxes, spends, licenses, restricts, and distributes in the name of the public. When the systems making those decisions are private, technically complex, unelected, and difficult to audit, the constitutional description becomes progressively less accurate as a guide to where power actually sits.

The state retains the legal form of sovereignty while exercising less of its substance.

States have always relied on private actors for certain functions: road construction, military logistics, outsourced IT. The novelty lies in the depth and character of the current migration. What is moving upstream is not infrastructure, which the state can rebuild, or logistics, which the state can redirect. What is moving upstream is judgment, the cognitive core of the allocation function that was supposed to be the distinctive contribution of accountable human governance.

When the judgment migrates, what remains?

The Judgment-Liability Split

Every allocation decision has two parts: the judgment that produces the outcome, and the liability that attaches to the outcome. Liability here is not only the narrow legal kind. It is the broader answerability that attaches to the state regardless of where the judgment was made: legal exposure, political accountability, and the public duty to explain.

For most of the modern state’s history, both parts lived in the same institution. A human official exercised judgment and was answerable for it. The civil servant who denied a claim, approved a license, or sanctioned a business could be questioned, audited, and overruled. The decision had an author. That authorship was the mechanism through which democratic governance claimed to be distinguishable from arbitrary power.

Algorithmic systems separate these two parts.

The judgment moves into the system. The welfare algorithm determines the benefit. The predictive tool informs the detention decision. The hiring filter removes the résumé before any human sees it. The procurement system excludes the vendor before any official reviews the bid. Each system encodes a set of choices, priorities, and weightings that determine the output. Those choices were made by someone: a private vendor, a development team, a training dataset, a procurement contract. But the choices are not visible to the applicant. They are often not fully visible to the agency deploying the system. They were made before the specific case being decided was ever presented.

The liability, however, remains with the state.

When the algorithm produces a wrongful denial, citizens do not sue the vendor’s training team. They sue the agency. When the risk model generates a discriminatory pattern, the political accountability runs to the minister, not the model developer. When the automated welfare system removes a benefit erroneously, the grievance is directed at the state, because the state is the institution that promised to administer the system fairly and in accordance with law.

This is the core dysfunction of the allocation state: legal sovereignty and political accountability for decisions the state neither fully made nor can fully explain. The gap between accountability and authorship grows as technical systems become more complex, more entangled with private vendors, and more difficult to audit from the outside.

A human official who makes a bad decision can be questioned, and can explain her reasoning. An algorithm that produces a bad output may have no single reasoning that can be surfaced. It was produced by a model trained on historical data, with weights adjusted through optimization processes, combined with real-time inputs, and processed through a pipeline that no one in the agency fully designed or currently understands.

The decision has an output. It does not always have an accessible author.

The Legitimacy Wrapper

Faced with this gap, the state adapts through legitimacy production.

It generates process: the appearance of accountable governance around decisions that were actually made by systems operating faster and more opaquely than the governance overlay can follow. The state builds audit requirements, algorithmic impact assessments, explainability mandates, human-in-the-loop rules, appeal procedures, vendor certifications, procurement standards, and regulatory frameworks.

None of these instruments is fake. Audit requirements produce at least partial visibility. Appeal procedures allow some corrections. Human review catches some errors. Certification processes screen some bad systems.

But they do not restore what was lost.

The original condition was that a human exercised the judgment, could explain it, and could be held responsible for it. The post-algorithmic condition is that a human reviews or approves a judgment produced elsewhere, under time and volume pressures that often make review nominal rather than substantive, and then holds nominal accountability for a decision the human did not author. The form of accountability remains. The substance has migrated.

What the state is constructing, of necessity, is a legitimacy wrapper: a set of processes and representations that allow the system to maintain the appearance of accountable public allocation even after the effective authority over many allocation decisions has moved into private technical infrastructure.

This is adaptation under constraint, not cynicism. No modern state can refuse algorithmic systems and return to purely manual administration. The volume, complexity, and speed of modern governance make that reversion impossible. The real choice is between algorithmic governance with more or less honest legitimacy architecture around it.

The risk is a subtler one. A state that becomes skilled at wrapping systems in the language of accountability, without restoring its substance, may come to mistake the wrapper for the thing it is wrapping. It may mistake compliance documentation for actual oversight. It may mistake algorithmic audits for democratic accountability. It may mistake the presence of an appeal mechanism for the genuine capacity of citizens to contest decisions made by systems they cannot see or understand.

The wrapper is not the house. The state that settles for the performance of accountability, rather than working to reconstruct its substance, stops building.

Judgment Migrates. Liability Stays.

Three sectors already demonstrate the pattern at scale, each revealing a different dimension of the same structural condition.

Welfare and benefit administration. The Netherlands built a System Risk Indication system called SyRI, designed to identify welfare fraud risk by combining personal data drawn from multiple government sources. In 2020 the District Court of The Hague ruled it unlawful, finding it violated the right to private life under Article 8 of the European Convention on Human Rights: the legal framework was insufficiently transparent and verifiable, and the data processing was disproportionate to the aims pursued. [1] SyRI shows that the split can open even without a private vendor. The state built the system, authorized it, and remained answerable for it. Yet the person it flagged still faced a classification she could not see or meaningfully test, produced by a process no official reconstructed for her. When the system failed the legal test, the state dismantled it. The people it had wrongly flagged had no recourse against the process that produced them.

Criminal justice. Predictive risk assessment tools are used across multiple U.S. states to inform bail, sentencing, and parole decisions. COMPAS, a proprietary tool produced by a private vendor, is the most documented example. A 2016 ProPublica analysis found that COMPAS scores were twice as likely to incorrectly flag Black defendants as high risk compared to white defendants who did not go on to reoffend. The vendor disputed the methodology. Courts disagreed on whether defendants had a right to inspect the underlying algorithm. [2] Throughout the legal argument, the accountability for the sentence ran through the judge. The state signed the sentence. The system determined the input. Whether the defendant could inspect the model at all was itself contested in court.

Healthcare prioritization. In 2019, researchers found that a widely used commercial algorithm allocating healthcare resources showed significant racial bias, providing lower risk scores to Black patients who were comparably ill to white patients, resulting in fewer Black patients being referred for additional care. The source of the bias was structural: the algorithm used healthcare spending as a proxy for healthcare need, and because Black patients had historically received less care due to reduced access, the proxy systematically underestimated the severity of their conditions. [3] The vendor adjusted the algorithm after the study was published. The hospitals that had deployed it remained accountable for the outcomes they had produced in the interim.

Taken individually, each case might appear as a correctable implementation failure. Taken together, they reveal the structural condition the split predicts: in each case the judgment was generated in a system its subjects could not inspect, whether the vendor was private or the state itself, while the mandate and the fallout stayed public.

The judgment-liability split is not a future concern. It is already the operating condition of government across every advanced economy. [4]

The Efficiency Is Private. The Failure Management Is Public.

The allocation state does not only wrap private systems in legitimacy. It also backstops them when they fail.

When the state deploys algorithmic systems that later produce systematic errors, it cannot disown the outcomes.

The efficiency gains from automation accrue to whoever deployed the system: hiring costs fall for the firm, underwriting margins improve for the lender, fraud-detection payroll falls for the agency in the short run. The costs of failure flow disproportionately through the state: remediation programs, litigation exposure, enforcement actions, and political damage arrive at the public institution even when the system that generated the problem was built and sold by a private vendor.

The efficiency is private. The failure management is public.

There is a deeper version of this asymmetry. Where the systems performing the judgment are themselves rented, and increasingly they are, they are intelligence environments the state leases rather than owns, running on compute infrastructure it does not control and cannot easily replicate. In those cases the state is not only outsourcing a task. It has become a tenant of its own allocation functions, paying for positional access to capacity held by someone else. When the lease is where the judgment lives, the efficiency accrues to whoever owns the estate, and the legitimacy burden stays with whoever signed the mandate.

That asymmetry is the structural consequence of deploying privately owned allocation systems in public functions without retaining sufficient operational control to prevent failures rather than only remediate them afterward.

The Allocation State Is Not Omnipotent

A possible misreading of this argument would treat the allocation state as an updated version of the total administrative state: more pervasive than any Weberian bureaucracy, delegating superficially while retaining ultimate control. That reading inverts the actual condition.

The allocation state is frequently dependent on the systems it nominally governs.

This is what distinguishes it from both the classical welfare state, which administered its own allocation functions directly, and the authoritarian surveillance state, which builds or captures technical systems under state authority. The allocation state depends on private vendors, foreign-owned infrastructure, proprietary models, and technical systems it lacks the internal capacity to fully evaluate, replace, or override.

A government agency that has integrated an AI hiring tool cannot easily remove it when its staffing capacity has been reduced on the assumption of automation assistance and its institutional knowledge of alternative approaches has atrophied.

Each adoption of an external system appears rational at the moment: faster processing, lower costs, better accuracy than available alternatives. Each adoption also increases the agency’s dependence on systems whose parameters it cannot set, whose training data it does not own, and whose replacement it cannot easily afford. The dependency compounds across time. Each new integration makes the next exit more expensive.

The allocation state in the Global South illustrates the extended form of this condition. Across large parts of sub-Saharan Africa, South Asia, and Latin America, states outsource allocation to multilateral donor systems, NGO compliance architectures, IMF conditionality frameworks, and World Bank project management requirements. Decisions about what healthcare infrastructure gets funded, which populations qualify for emergency food distribution, which governance reforms unlock budget support are made by institutions not accountable to the citizens they affect. The state signs the agreements, implements the programs, and absorbs the political consequences. The allocation logic was written elsewhere.

Advanced economies are importing a variant of this condition through vendor dependency rather than donor dependency. The structure is the same. The judgment originates outside the state. The liability stays inside it.

Access Is Allocated Before the State Arrives

There is a second dimension of the allocation state less visible in the domain of government services but no less consequential.

The same logic that repositioned the state relative to its own public systems has also produced a private allocation regime that shapes life chances before the state’s systems come into play at all. An earlier essay in this series argued that ranking is itself allocation, and that the decisive act moves upstream of the market: hiring platforms filter résumés, scoring systems decide which applications advance, and a person is sorted long before any public institution is involved. That argument stands. What matters here is what it does to the state.

At no point in that chain is there a publicly accountable allocation system. Each step is private, automated, and governed by the internal policies of whoever built it.

By the time a person reaches the domain where public allocation would apply, welfare benefits, public housing, healthcare coverage, educational access, the private allocation system has already substantially determined the material context in which the public system will find her.

The state allocates at the residual layer. The upstream allocation has already happened.

This is why the allocation state cannot be understood only as a story about government deploying AI. It is about the relationship between private algorithmic allocation systems and public ones, and about which is doing the more consequential work. Public remediation inherits the shape of private allocation. The allocation state manages the residual. It does not design the field.

Sovereignty Becomes Override Capacity

Sovereignty, in the classical understanding, is the capacity for final decision. The sovereign is whoever can say “this is what happens” and make it so.

In the allocation state, that capacity is fragmenting across a chain of actors that no single institution controls. The private vendor built the model. The state deployed it. The state does not fully understand it. The vendor cannot be compelled to explain it completely. No court can inspect it comprehensively. No legislator designed the parameters that produce its outputs. The final decision is distributed across technical choices, contractual relationships, training procedures, and deployment configurations assembled over years by different organizations with different interests.

This is not the end of sovereignty. It is its disaggregation into a form that existing constitutional vocabulary does not yet adequately describe.

Recovering effective sovereignty over the allocation state requires not merely the legal authority to regulate, but four operational capacities that most states currently lack.

Interpretive capacity: the state must know, in substantive rather than formal terms, what the system is doing and why, which requires genuine technical expertise inside government rather than the ability to commission audits no official can evaluate.

Override capacity: the state must be able to stop or modify the system without paralyzing the functions that depend on it, which requires maintaining institutional alternatives even while delegating execution.

Replacement capacity: the state must preserve the ability to switch vendors or rebuild systems before crisis arrives, not after, which means treating that optionality as a governance asset rather than an unnecessary redundancy.

Design capacity: the state must shape the architecture before procurement hardens dependency, which means participating in defining the systems it deploys rather than accepting products built for commercial markets and adapting them to public purposes.

These are the operational definition of what governing means when judgment has migrated into technical infrastructure. Most states, currently, do not have all four. That gap is not a failure of political will. It is the consequence of a transition that moved faster than institutional adaptation could follow, deploying systems whose accumulating dependencies were not visible until they had already formed.

Rights and Scores

Beyond shifting where decisions happen, the allocation state changes the language in which allocation is expressed.

For most of the modern state’s history, allocation was described in legal and moral vocabulary. A person was entitled to a benefit, eligible under a statute, subject to a duty, protected by a right. That language was normative. It organized the relationship between the individual and the state in terms that could be contested, interpreted by courts, and revised through democratic processes.

Algorithmic systems operate in a different vocabulary. Risk scores. Eligibility indices. Fraud probability thresholds. Hiring fit ratings. Benefit tier assignments. Compliance pass/fail determinations. This vocabulary is quantitative, probabilistic, and optimized for operational efficiency. It does not describe the individual. It classifies the individual against a population model. The person is not a rights-holder with specific entitlements. She is a data point whose classification positions her in a distribution.

The legal vocabulary asks: does this person have a right? The algorithmic vocabulary asks: what is this person’s score?

These are not different ways of saying the same thing. Algorithmic allocation does not merely automate legal judgment. It changes the grammar of public obligation from rights to scores. A person can have a legal right and still receive a low score. A person can fail an eligibility index and still be legally entitled. The two systems can produce contradictory outcomes, and when they do, the question of which governs is increasingly the central question of administrative law, civil rights enforcement, and democratic legitimacy.

The allocation state is not merely a state that uses machines to administer law. It is a state in which the tension between normative and operational vocabularies has become a structural feature of governance, not a temporary technical problem that better systems will eventually resolve.

The systems will improve. The tension will remain.

The Allocation State and What Came Before

The preceding essays traced how the AI economy reorganizes the relationship between production and distribution, between labor and claim, between market access and eligibility, between capital and the justification for its returns, between surplus populations and the institutions managing them, and between fiscal states and the assets they can reach.

The allocation state is where those transformations converge at the institutional surface of governance.

Each transformation leaves a remainder no private actor will hold: a benefit to administer, a system to certify, a dependency to define, a political consequence to absorb, a population to stabilize, a revenue base to rebuild.

All of these functions converge on the state, not because the state is the most powerful actor in the AI economy (it often is not), but because the state is the institution society has assigned accountability for outcomes that no other institution has been assigned accountability for.

The allocation state is the form this assignment takes when the state lacks the operational capacity to match its assigned accountability. It is the state held responsible for outcomes it cannot fully govern, legitimizing systems it does not fully understand, managing populations whose claims it cannot fully honor, taxing a surplus it cannot fully reach, and providing a legitimacy wrapper around allocation architectures designed by actors answering to different principals than the citizens living inside the outcomes they produce.

The state has not disappeared. It has been repositioned.

The state does not control the system.

It is what the system calls when it needs to be legitimate.

Notes

[1] Rechtbank Den Haag [District Court of The Hague], NJCM c.s. v. De Staat der Nederlanden, ECLI:NL:RBDHA:2020:865, February 5, 2020. The court ruled that SyRI violated Article 8 of the European Convention on Human Rights, finding the legal basis insufficiently clear and the data processing disproportionate to the aims pursued. The Dutch government subsequently dismantled the system. https://uitspraken.rechtspraak.nl/details?id=ECLI:NL:RBDHA:2020:865

[2] Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner, “Machine Bias,” ProPublica, May 23, 2016. The analysis found that COMPAS recidivism risk scores produced an asymmetric error structure: Black defendants who did not reoffend were significantly more likely to be classified as high risk than white defendants who did not reoffend. The vendor disputed the statistical methodology. Courts disagreed on whether defendants had a right to inspect the underlying algorithm. https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing

[3] Ziad Obermeyer, Brian Powers, Christine Vogeli, and Sendhil Mullainathan, “Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations,” Science 366, no. 6464 (2019): 447-453. The algorithm used healthcare spending as a proxy for healthcare need. Because Black patients historically received less care due to reduced access rather than lower need, the proxy systematically underestimated the severity of their conditions. https://www.science.org/doi/10.1126/science.aax2342

[4] Virginia Eubanks, Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor (St. Martin’s Press, 2018). Provides the most sustained empirical account of the judgment-liability split operating across welfare, child protective services, and public housing: private or semi-private algorithmic systems producing outcomes for which public institutions bear accountability.

The modern state was built on a paradox it never had to explain.

It governed millions of people it could not directly observe, their incomes, transactions, assets, and behavior mostly hidden inside households, firms, and private relationships. It taxed a vast and opaque economy. And yet it funded itself, for most of a century, at rates that built welfare states, militaries, infrastructure systems, and public research capabilities that no previous political form had achieved.

It could do this because of a single administrative invention that most people experience as tedium: the paycheck.

The wage was not only the mechanism through which workers received income, as this series has argued. It was the mechanism through which the state could find workers at all.

The payroll system was one of the most effective information machines ever built into a capitalist economy.

Employers reported wages before workers filed returns. Taxes were deducted at source. Social insurance contributions moved automatically. The state did not need to chase income because the employment relationship itself was a standing disclosure requirement.

What is coming under stress in the AI economy is not only a tax rate problem, or even a conventional tax base problem. It is an information problem. The wage worked as a fiscal instrument because production was legible. It moved through employers, who were registered, accountable, and already processing payroll. Output had addresses. The state knew where to look.

AI reorganizes production in ways that make it progressively less legible to that architecture, not by fraud, not by evasion in the ordinary sense, but by structural drift. The economy is becoming more legible to machines and less legible to states. Call it the legibility inversion. That is the problem at the center of the fiscal question.

And it arrives at the worst possible moment: when states must fund the management of surplus populations, the stabilization of legitimacy, and the transition costs of an economy remaking itself, all while the payroll-based revenue system that funded everything is quietly losing its grip.

The Payroll System as Information Architecture

To understand what is being lost, it helps to understand what the payroll system actually was.

Modern fiscal states were built on a series of compounding information advantages. Wages were visible by design: they moved through registered entities with legal obligations to report. The employer became an unpaid arm of the revenue authority. The state’s information asymmetry, it must know what you earned before it can tax it, was solved, for wage earners, not by auditing but by delegation. The counterparty to the wage transaction already had independent legal duties to report it.

This made collection extraordinarily efficient. Evasion required collusion between employer and employee, which the threat of audit on both sides largely suppressed. The information was timely, deducted before the worker ever handled the full sum. And it was comprehensive: payroll covered not only income tax but social insurance, health contributions, unemployment premiums, and pension levies. The fiscal state built its entire welfare architecture on top of this infrastructure. The dependence persists: personal income taxes and social security contributions still supply roughly half of all tax revenue across the OECD, and in the United States, individual income and payroll taxes together account for about 85 percent of federal receipts. [1]

E. P. Thompson’s account of industrial time-discipline, the way the clock-time of wage labor reorganized the moral experience of daily life, captures one dimension of the payroll system’s power. But it also reorganized the state’s administrative capacity. The wage did not merely train workers to inhabit time differently. It trained the state to find its citizens through the employer, which meant the state could fund public goods without building a surveillance apparatus capable of reaching every private transaction directly. [2]

That is the arrangement now under stress. Not because employers are hiding wages more aggressively. But because the relationship between output and employment is loosening, and with it the coverage of the fiscal information architecture that depended on it.

The Inversion

The AI economy is producing an asymmetry the tax state is not built to handle: the economy is becoming more legible to algorithmic systems and less legible to fiscal states.

Consider what the large platform and AI companies know about economic activity. They see transactions before they become taxable events. They process payments before they become reportable income. They observe labor before it becomes employment: gig work, creator revenue, freelance contracts, platform selling. They hold the data on which the productive system increasingly depends. Their systems can classify, route, price, and rank economic actors in real time at enormous scale.

The state, by contrast, must wait. It waits for annual returns. It waits for audit cycles. It waits for treaty negotiations with other jurisdictions. It builds enforcement capacity measured in years while the transactional infrastructure it is trying to reach moves in milliseconds.

This is not only a matter of speed. It is a matter of architectural position. The fiscal problem is not that the state has lost money. It is that the state has lost position. The payroll state sat inside the main transaction of industrial capitalism: employer, worker, wage, withholding, contribution, benefit. The fiscal system was embedded directly in the relationship that organized production. The AI economy moves the decisive transaction elsewhere. The platforms sit between economic actors and their markets. They own the intermediary layer where value is generated, recorded, and routed before it ever surfaces in a form the fiscal state can reach. A payment processed through a platform’s rails, attributed to IP held in one jurisdiction, routed through a cloud layer in another, and sold to users in a third has already passed through more private information architecture than most tax authorities will ever inspect.

The problem is not opacity in general. Someone can see the economy very clearly. It is just no longer the public authority.

AI intensifies this because it concentrates value further upstream. The surplus of synthetic production accumulates in compute estates, model IP, cloud infrastructure, and the interface layer through which users and markets interact. These are not factories with a fixed address. They are positional assets whose value is real but whose location is a question of legal architecture, not physical geography.

The OECD’s Pillar Two global minimum corporate tax, a floor of fifteen percent, was the most serious international attempt to slow base erosion. It was designed for the platform era. The AI era may make it look modest. The gap between where value accumulates and where states can extract it is not a consequence of inadequate rate-setting. It is a consequence of mismatched architectures. [3]

The Paycheck Thins

While the surplus migrates upward into structures the state cannot easily reach, the base it can reach is contracting.

Labor’s contribution to production does not need to collapse for the payroll-based fiscal system to weaken. It only needs to drift. Firms producing the same output with fewer junior employees are not dramatic events. They are allocation decisions spread across thousands of hiring cycles, quietly not replacing people who leave, compressing workflows that previously required several people into one, using AI to eliminate the first-draft function that once justified the entry-level hire.

Each individual decision is invisible at the macro level. The cumulative effect is an erosion of the coverage of the payroll system, not a cliff, but a slope. The early-career employment decline in AI-exposed occupations that opened this series, around sixteen percent relative to less-exposed peers, reads differently from a fiscal seat. If that decline holds across professions and cohorts, it lands where payroll was historically most complete: formal, reported, mid-tenure employment in professional and analytical roles.

The gig economy was the preview. Platform work replaced employment contracts with contractor arrangements, which are structurally less covered by payroll systems. Self-employment, project-based work, creator monetization, and informal service provision all represent forms of labor that the fiscal architecture treats less cleanly: more evasion, more underreporting, more reliance on the individual’s own disclosure rather than the counterparty’s.

AI does not only add to this by eliminating some jobs. It adds to it by making the surviving work less structurally legible to payroll. The worker who supervises AI outputs rather than producing first drafts, the professional whose billing is now half what it was because clients do their own AI-assisted analysis first, the contractor whose project work has fragmented across platforms: none of these appear as unemployment. They appear as reduced coverage.

The state built its obligations around an economy where stable employment was the normal condition of working-age adults. That assumption is becoming less true every year, not through rupture but through structural drift.

The State Inherits What Firms No Longer Need

The fiscal problem after labor is not only that revenue weakens. It is that expenditure pressure rises as revenue weakens.

A private firm can reduce headcount and improve its margin. It can automate junior workflows, compress the professional layer, optimize for output per employee. From the firm’s perspective, these are efficiency gains. From the state’s perspective, they are cost transfers.

The displaced junior analyst, the professional whose billing rate has compressed, the credential-holder who cannot find the entry-level role that was supposed to follow from the education the state subsidized: they do not disappear from the political system. They remain citizens. They still require healthcare, housing, income support, and public order. Their claims on the state do not diminish because the economy has decided it needs less of them.

The economy may need fewer workers. The state does not need fewer citizens.

This is the fiscal expression of the surplus-human problem the previous essay described. Containment, transfers, credential extension, therapeutic systems, platform participation, is not free. It has to be funded. And it has to be funded from a system whose principal revenue base is the employment relationship, which is simultaneously the relationship being compressed.

The state that manages an AI economy is being asked to spend more on its surplus population while extracting less from the system producing the surplus. That is not merely a fiscal tension. It is a structural contradiction that no rate adjustment can close.

The Ground That Cannot Be Offshored

States will adapt. They always do, imperfectly and belatedly. The fiscal history of capitalism is partly a history of states chasing value wherever it migrates: from agricultural rents to industrial capital to platform revenues to whatever comes next.

The emerging logic of post-labor taxation follows a single thread: find what cannot be moved, and tax it.

Compute is the most promising candidate, for the same reason land was the original tax base. The AI compute estate, chips, data centers, energy systems, cooling infrastructure, fiber corridors, is physically anchored in ways that the mobile IP and transfer-pricing structures of the platform era were not. A data center cannot be relocated to Ireland through a licensing agreement. It requires land, grid access, water, permits, and a long-term capital commitment to a specific geography. It is visible, assessable, and immovable.

And it matters fiscally not only because it is hard to move. It matters because it can be measured. Compute has meters: energy draw, chip deployment, data-center capacity, inference volume, cloud contracts. The payroll system succeeded because wages were countable at the point of payment. A post-labor fiscal system will need an equivalent point of measurement, some machine-readable proxy for synthetic production. The question is not only where value is created. It is where value becomes visible enough to claim.

Measuring where computation happens is not the same as deciding computation should be taxed. The meter locates synthetic production. It does not settle what the state ultimately reaches, whether rents, profits, revenues, or transactions. Reading the two as one invites the crude version of the idea, a flat charge on raw compute, when the harder move is locating synthetic production at all.

Several European countries have introduced or are considering data-center energy levies, framed sometimes as carbon pricing, sometimes as infrastructure contributions from systems that impose significant costs on local grids and water supplies. The fiscal logic beneath both frames is identical: extract from the physical substrate of synthetic production because that substrate has an address.

Digital services taxes, levies on revenues generated from users in a jurisdiction regardless of where the firm books profit, represent a cruder instrument, one that the United States has consistently opposed when applied to American firms. But they contain the same insight: if profit can be located anywhere, tax the consumption it depends on, which cannot be moved.

The most ambitious proposals involve taxing compute capacity directly, a severance levy on AI capability operated within a jurisdiction, analogous to the resource extraction taxes applied to oil or minerals. The precedent for this framing is Norway’s sovereign wealth fund: the premise that petroleum rents belong partly to the public because petroleum is a common inheritance, and that private extraction from common inheritance carries a public obligation. The inheritance grammar proposed earlier in this series for post-wage income applies here to the fiscal question: AI surplus is built on public science, public law, public infrastructure, and collective data. Extracting it without a public return is not merely inefficient. It is private appropriation from a civilization-built asset base. [4]

Sovereign AI fund proposals, in various forms across Europe, the Gulf, and parts of Asia, represent the ownership version of this logic: rather than taxing AI surplus after private accumulation, the state acquires equity stakes in AI infrastructure as a condition of market access, subsidization, or regulatory clearance. The state becomes a co-owner rather than a creditor.

None of these instruments is fully formed. Each generates its own political economy of resistance. Tax too aggressively and the infrastructure migrates. States compete for data centers with subsidies and streamlined permitting, which creates a race to the bottom that can leave the public net-negative on extraction. Tax the wrong surface, revenues rather than surplus, transactions rather than value, and dominant firms absorb the cost while smaller actors cannot.

The danger of bad extraction is not only fiscal inefficiency. It is that the tax reproduces the concentration it claims to correct. Complex, compliance-heavy levies favor the largest firms, which can absorb administrative costs and structure around whatever instrument the state deploys. A poorly designed AI tax can become another form of the barriers to entry that already make the compute estate inaccessible to most actors.

The Objection From History

The obvious objection is that none of this is new. States have taxed novel economic forms before. They taxed railroads, factories, oil, corporations, financial income, and eventually platforms. Fiscal systems lag transformations and then catch up. Why should AI be different?

The objection deserves a real answer, because it is half right. States will adapt here too; the previous section is a catalogue of the early attempts. What the objection misses is that AI weakens two fiscal architectures at once. It weakens payroll visibility from below, because less production has to pass through stable employment. And it weakens corporate territoriality from above, because more surplus accumulates in mobile IP, cloud contracts, model access, and interface control. The state is squeezed from both directions: the worker becomes less legible as a worker, while the surplus becomes less locatable as profit.

Every previous fiscal adaptation chased value after it moved. This one has to rebuild the instrument by which value becomes visible at all. Railroads had track. Factories had chimneys. Oil had wells. The payroll system had the employment contract. The AI economy’s equivalents exist, but they report to private systems first.

The sharper form of the objection presses harder: employers were themselves conscripted once, turned into unpaid arms of the revenue authority, and platforms, payment processors, marketplaces, and cloud providers can be conscripted the same way. That is correct, and it is the likely path. But it concedes the mechanism rather than defeating it. The state has not lost the information in principle. It has lost the position from which the economy once reported itself without being asked, and rebuilding that position means turning a set of transnational, privately governed intermediaries into instruments of collection, as payroll once turned the employer.

That is a different kind of problem, and rate-setting does not solve it.

The Fiscal Constitution of AI

Underneath the technical questions, compute levies, energy surcharges, digital services taxes, transfer pricing enforcement, sovereign wealth funds, is a political question the fiscal crisis will eventually force into the open.

What does the public get from a system that has learned to produce without needing as many of the public as it once did?

This is not, at base, a question about tax rates. It is a question about the relationship between production and membership. The progressive income tax was not only a revenue instrument. It was a political statement about the relationship between individual success and collective conditions, an acknowledgment that no wealth was purely private in origin. Payroll taxes funding social insurance were not only premiums. They were a statement about shared risk across the working population, a translation of the wage relationship into a social contract with temporal structure: you contribute while you can, you claim when you cannot.

The fiscal state after labor will need to make comparable political statements about AI surplus. Not because the old statements were always correct, or because the institutions they built were always efficient, but because a productive system that extracts large surplus without paying for the collective conditions it depends on, public science, public infrastructure, educated workforces, stable political environments, the accumulated data of millions of private lives, is extracting from civilization without returning to it.

The fiscal constitution of the AI era is already being written, badly and in fragments. Energy levies, minimum taxes, transfer pricing rules, digital services taxes, sovereign fund proposals: these are early drafts. They will require many revisions. The largest firms will extract carve-outs. The political process will lag the economic transformation. The instruments will be crude relative to the complexity of what they are trying to reach.

That is how fiscal constitutions have always been written. The income tax was crude in its first versions. Payroll taxes took decades to reach their mature form. Corporate taxation still has not resolved the question of where profit is located in a multinational firm, a question the platform economy made worse and the AI economy is making harder still.

But the question the fiscal state is now asking is sharper than it has been in decades. Not: how do we tax income? That question assumed a world where income was attached to labor, and labor was attached to employment, and employment was attached to a jurisdiction. All three of those attachments are loosening simultaneously.

The sharper question is: who owns the ground?

This series has already answered it. In agrarian economies the ground was land: visible, immovable, taxable. In industrial economies it was capital. In the AI economy it is compute, expensive to build, slow to replicate, physically anchored, and the precondition for synthetic production. The state that can establish a public claim on what synthetic production generates from a publicly built inheritance has a revenue base. The state that cannot will be left funding containment with instruments designed for an economy that no longer exists.

But locating the ground is the easier half of the problem.

The harder half is the one this essay began with.

The wage was the state’s information system as much as its tax base. It made citizens legible to fiscal authority, social insurance systems, and democratic institutions. The AI economy makes citizens more legible to platforms and less legible to states.

The fiscal reconstitution of the AI era is ultimately about reversing that inversion, not to restore the old information architecture, which will not come back, but to establish a new one capable of translating synthetic production into public revenue.

Whatever instruments emerge, levies, funds, equity stakes, minimum taxes, they will only work if the state can see what it is taxing. The payroll system did not succeed because its rates were clever. It succeeded because it made the economy report itself.

That is the rebuild the fiscal state actually faces. Not a new rate, and not even, in the end, a new base. A new sensor.

The state that learns to read synthetic production will fund itself inside the AI economy.

The state that does not will govern an economy it can no longer see.

Notes

[1] OECD, Revenue Statistics 2025, OECD Publishing, 2025: in 2023, social security contributions accounted for 25.5 percent and personal income taxes for 23.7 percent of total tax revenues on average across OECD countries. https://www.oecd.org/en/publications/revenue-statistics-2025_3a264267-en.html. For the United States: Congressional Research Service, “Overview of the Federal Tax System in 2024,” R48313: in fiscal year 2023, the individual income tax generated 49 percent and payroll taxes 36 percent of federal revenue. https://www.congress.gov/crs-product/R48313

[2] E. P. Thompson, “Time, Work-Discipline, and Industrial Capitalism,” Past & Present 38 (1967): 56-97. Thompson’s account of the moral transformation of time under industrial capitalism captures how the wage reorganized not only compensation but the entire temporal architecture of social life, including, as a secondary consequence, the administrative legibility of workers to fiscal institutions.

[3] OECD, Tax Challenges Arising from the Digitalisation of the Economy, Global Anti-Base Erosion Model Rules, Pillar Two, 2021. https://www.oecd.org/tax/beps/tax-challenges-arising-from-the-digitalisation-of-the-economy-global-anti-base-erosion-model-rules-pillar-two.htm

[4] On the Norwegian precedent: Norway’s Government Pension Fund Global, established 1990, now manages approximately $1.7 trillion in assets accumulated from petroleum revenue. The inheritance framing, that common assets generate a public claim, is the conceptual precedent for applying similar logic to AI-generated surplus. On the emerging framework for AI-era public finance: Anton Korinek and Lee Lockwood, “The Future of Tax Policy: A Public Finance Framework for the Age of AI,” Brookings Institution, February 2026, drawing on their working paper of January 8, 2026. https://www.brookings.edu/articles/future-tax-policy-a-public-finance-framework-for-the-age-of-ai/. On the state-as-co-owner model: the European Commission’s InvestAI initiative, announced February 2025, targets 200 billion euros in mobilized public-private AI investment, including a 20 billion euro facility for AI gigafactories structured as a layered fund with a public first-loss tranche; the Commission’s technological sovereignty package of June 2026 proposes a European equity capacity to take stakes in strategic technology firms.

He had not been arrested by Belarus. He had not been convicted of a crime in Serbia. He had been flagged by an Interpol Red Notice, a police cooperation request issued through an international system that treats criminal accusations as standardized inputs. Belarus had labeled him a fugitive on tax charges he denies. Serbia processed the alert. [1]

For a year his life was reorganized around a form. Seven months in Belgrade Central Prison, then house arrest with a tracking bracelet and one hour outside a day. The punishment was not only the possibility of extradition. It was the process itself, running through institutions that did not need to share Minsk’s politics to act on its signal.

In July 2024, Interpol deleted the notice, finding it incompatible with the provisions of its own constitution that bar the organization from political cases. [2]

The process did not stop.

Serbian proceedings ran three more months. An appeals court overturned the extradition order in September but sent the case back for retrial, and Hniot stayed under house arrest. He walked free on 31 October 2024, a year to the day after his arrest, because Serbian law caps pre-extradition detention at a year. The case was still open when he left the country. [1]

The accusation had been withdrawn at the source. The machinery kept running on momentum.

Authoritarian regimes have discovered that the most effective way to reach a dissident abroad is not always to send an agent. It is to submit a form. File the right paperwork, in the right format, through the right international channel, and the institutions of liberal democracy may do the rest. Border systems flag passports. Banks freeze accounts. Visa offices open reviews. Universities, employers and nonprofits hesitate.

The regime does not need to win. It only needs the process to run.

In the age of AI, that process is becoming faster, cheaper and harder to see. What authoritarian states will produce more of is not propaganda. It is machine-readable suspicion.

A blacklist, a terrorism label, a criminal charge, a state-media smear, a coordinated reporting campaign: in the pre-AI world these were political attacks, damaging reputations and creating bureaucratic friction. In the AI-mediated world they become data, absorbed into banking systems, visa screening, compliance software, hiring platforms, grant reviews and automated risk scores.

The regime no longer needs Western institutions to believe the accusation.

It only needs their systems to ingest it.

Procedural Trust

To understand the exploit, one must understand the system it exploits.

After World War II, liberal democracies built their international order around a sound moral intuition: power should be forced through procedure. Courts, treaties, banking rules, migration systems and police cooperation were designed to slow arbitrary force into channels that could be reviewed and contested. Procedure created a hidden vulnerability of its own.

Liberal institutions often assume that official-looking claims, even from adversaries, maintain some relationship to evidence and legal reasoning. A criminal charge is a criminal charge. A Red Notice is a Red Notice. The system is built to process such inputs and produce outputs: approvals, denials, delays, escalations.

Authoritarian regimes learned to exploit that assumption.

Freedom House has documented 1,375 direct physical incidents of governments reaching across borders to silence dissidents between 2014 and 2025: assassinations, abductions, illegal deportations, family intimidation. In at least eleven of the 2025 cases the perpetrator government used an Interpol notice to do it. [3]

But the most consequential shift may be less visible. It is administrative.

Authoritarian states learned to translate political hostility into bureaucratic form. A dissident does not need to be captured if he can be made institutionally toxic. A journalist does not need to be censored if she can be turned into a compliance problem.

The accusation becomes more powerful when it stops looking like propaganda and starts looking like a file.

That is the exploit: repression formatted for liberal processing.

Exported Blacklists

Mikhail Khodorkovsky, once Russia’s richest man and imprisoned by Putin for a decade before going into exile, described the mechanism with unusual precision in May 2026, testifying before the European Parliament’s special committee on the European Democracy Shield. [4]

He and his colleagues in the Russian Anti-War Committee have been designated terrorists by the Kremlin. Most European states, he told the committee, understand this as a political decision with nothing to do with security. In practice it means that on every trip, and in every interaction with a bank or migration authority, there is a terrorism flag in his file. For thousands of less prominent exiles the consequences are duller and heavier: banks close accounts because compliance departments fear any link to a name marked suspicious, migration authorities delay status extensions because they cannot interpret a Russian accusation, passports go unrenewed until people are stateless in practice. Europe’s financial system, he argued, has become an unwitting instrument of the repression it opposes.

A Russian blacklist becomes a European banking problem. A Kremlin terrorism label becomes a migration delay. A politically motivated designation becomes a compliance risk, not because the institution endorses the claim, but because the claim has entered its data environment and triggered its risk logic.

This is not traditional censorship. It is repression through interoperability.

Igor Pestrikov’s case shows the shape it takes. A Russian businessman and shareholder in a magnesium and rare-earths producer, he left in 2022 after refusing to supply metal to buyers the state had designated, and settled in southern France. Russia did not publish a Red Notice against him. It used a diffusion, a request circulated directly between national police bureaus, which moves faster and draws less scrutiny than a notice the target can contest. For the two years it was live his bank accounts were frozen and he could not rent an apartment. Interpol’s review commission eventually deleted it as politically motivated. [5]

No bank officer accused him of anything. A screening system returned adverse results and institutions declined the exposure.

The pattern repeats wherever a risk system sits between a person and a service. No one has to say: we believe the dictatorship. Each says only: this is complicated.

In a risk-managed institution, complicated often means not worth it.

The authoritarian state creates the label. The liberal system supplies the enforcement surface. The punishment arrives through the administrative caution of institutions that would never endorse the politics behind it.

The API

The API metaphor is not decorative. An application programming interface is a standard way for one system to send instructions to another: an input format, a receiving system, an execution layer. Send the right request, in the right format, and the receiving system acts.

Authoritarian states have learned to exploit all three. The input format is the official-looking signal: a criminal charge, a terrorism designation, an Interpol request, a state-media article. The receiving system is liberal institutional infrastructure: banks, visa offices, compliance vendors, platforms, universities, grant makers. The execution layer is not a police order. It is a risk decision: delay the visa, close the account, freeze the grant, restrict the profile.

The volume is rising. Interpol issued 15,548 Red Notices in 2024, a 27 percent increase and the largest annual total on record, while its own compliance task force refused or cancelled 2,462 notices and diffusions that year, up 54 percent, also a record. [6] More inputs arrive and more bad ones are caught, but the catching happens after the processing has begun.

Files given to BBC World Service and Disclose by a whistleblower, published in January 2026, sharpened the picture. Interpol’s review commission removed at least 322 notices in 2024 as unjustified, and Russia drew more complaints to it than any other member state. Fewer than one in ten of roughly 86,000 active Red Notices are published at all, and measures constraining Russian abuse were quietly relaxed during 2025. [7]

Hniot’s notice was corrected. The harder question is how much punishment lands before a correction arrives, and how little it undoes.

A secure API assumes some inputs are hostile and validates accordingly. Liberal democracy still runs too many systems as if official-looking inputs are valid by default. That assumption once made procedure possible. Now it makes procedure exploitable.

Risk Laundering

Banks, payment processors and compliance vendors are not villains in this story. They are operating rationally inside a system never designed to distinguish political persecution from genuine risk at scale. Risk systems are not the problem. They are being asked to process signals from regimes that manufacture risk as a political weapon.

Authoritarian regimes do not need to invent the risk system. They only need to contaminate it. The Financial Action Task Force, which sets global standards for fighting financial crime, opened a review of its own unintended consequences in February 2021. The stocktake it published that October sorted the damage under four headings: de-risking, financial exclusion, undue targeting of nonprofits, and curtailment of human rights, the last focused on due process and procedural rights. [8]

A standards body examined its own rules and found they were eroding procedure. The diagnosis came from the institution that wrote the standard.

De-risking is the named mechanism: banks withdrawing services from whole categories of customer rather than pricing the risk of any individual one. The categories that absorbed the damage were nonprofits, diaspora populations and people in conflict-affected regions, which is to say the populations transnational repression selects for. [8]

A bank does not have to conclude that an activist is guilty. It only has to conclude that she is operationally complicated.

This is what risk management can launder. A Kremlin designation need not be cited directly if it has already shaped the data environment around the person. A Chinese state-media smear need not be believed if it reappears downstream as adverse media. The machine does not repeat the lie. It operates the uncertainty the lie creates.

Machine Suspicion

AI does not create this exploit. It changes its physics.

The pre-AI version depended on bureaucratic friction. A blacklist had to be noticed by a compliance officer, translated, escalated and interpreted by someone with enough context to evaluate it. Each step was a chance for a person to pause or override. AI removes many of those pauses.

The signals Khodorkovsky described also stayed fragmented: a criminal file in one system, state-media coverage in another, an unrenewed passport in a third. In the AI-mediated world the fragments become legible to one another. Name-matching connects aliases across languages. Compliance software flags unresolved allegations.

Adverse-media screening is where this concentrates. LSEG’s World-Check media tool applies machine learning and automated tagging across more than thirteen thousand vetted sources in twenty-four languages, clusters the results into discrete events, and delivers them to compliance platforms through an API. [9] At that point the metaphor stops being a metaphor.

Nor is the automation merely a commercial preference. Successive European anti-money-laundering directives widened the range of firms required to screen open-source media and pushed those checks toward automation. The pipeline is built, mandated and running, and whatever an authoritarian jurisdiction produces in text enters it as text. [10]

No institution has to believe the regime. Each has only to downgrade the person by one notch: enhanced review, delayed approval, account closure, grant hesitation, reputational caution.

A model does not say: this dissident is a terrorist. It says: elevated risk, unresolved allegations, adverse media present, enhanced review recommended. Not belief. Probabilistic suspicion. And probabilistic suspicion, distributed across enough institutions, produces political outcomes without any single institution making a political decision.

An authoritarian lie no longer has to survive as a lie. It can survive as a probability.

Hostile Inputs

There is an obvious objection: not every criminal case from an authoritarian state is fabricated. Not every adverse-media flag is propaganda. Democracies cannot ignore all foreign-origin risk signals.

The answer is not blanket dismissal. It is adversarial weighting.

A terrorism designation from an independent judiciary and a court filing from a captured legal system should not enter downstream risk systems at the same default weight. Cybersecurity already runs on this principle: a login attempt from a known malicious network is handled differently from one on a trusted device. A secure system does not abolish communication. It validates the source, checks the payload and limits the damage any single input can cause.

The line is not always clean. In May 2025, El Salvador obtained Red Notices against two Salvadoran lawyers living in Spain, Ivania Cruz and Rudy Joya. UN experts identified them as retaliation tied directly to the lawyers’ human rights work, and Interpol revoked them as politically motivated. The two still faced extradition proceedings in Spain and did not secure asylum until 2026. [11]

A scheme keyed to a short list of usual suspects would have missed that case, and revocation, once again, did not stop the process the notice had started. The design problem is harder than sorting regimes into columns. What has to be weighted is the quality of the producing system: whether its judiciary can be overruled by its executive, whether a designation carries reasons, whether the target can see the accusation and answer it.

The goal is not to make dissidents untouchable, but to stop authoritarian regimes turning liberal procedure into an enforcement arm.

Patching Procedure

Fixing this does not require dismantling procedure. It requires making procedure adversary-aware. Five reforms would help.

Provenance labeling. Risk signals from regimes with documented records of political persecution should carry origin flags before entering compliance, migration, employment and grant systems. The source of an allegation should travel with the allegation.

Corroboration requirements. High-consequence decisions such as account closures, visa denials, grant cancellations and platform removals should not execute on a single authoritarian-origin signal without independent corroboration.

Source auditability. Compliance vendors should disclose whether their adverse-media summaries draw on state media, regime-linked outlets or opaque source chains in illiberal jurisdictions. Provenance is the foundation of judgment, not a technical footnote beneath it.

Fast-track review. Banks, platforms, payment systems and immigration offices need escalation paths for people plausibly targeted by transnational repression, separate from the standard commercial dispute queue.

Meaningful appeal rights. Anyone denied banking, visas, grants or essential services by an automated risk system should receive human review and a substantive explanation.

That last reform is the one most likely to fail, and the failure mode deserves naming. Human review placed above an automated judgment frequently works as a liability buffer rather than a check, supplying a face for the decision without supplying judgment about it. A reviewer who cannot see the source of the flag, cannot overrule the model and has no time to read the file is not a safeguard. He is a signature. Appeal rights are worth something only where the reviewer can see the provenance the first reform requires, can rule against the system, and leaves a record for doing so. Otherwise the appeal is one more piece of procedure the exploit can run through.

These reforms do not ask institutions to abandon caution. They ask them to become cautious about the right thing. The risk is not only that a dangerous person slips through. It is that a dictatorship learns how to make innocent people administratively radioactive.

The Process Runs

The postwar liberal order was built around trust in process. Authoritarian regimes learned to weaponize that trust, and AI is turning the exploit into infrastructure.

The exiled journalist does not disappear because one institution makes a tyrannical decision. She disappears administratively because many institutions make cautious ones. A dissident does not need to be convicted if he can be risk-scored. A journalist does not need to be censored if she can be deplatformed by procedure. An exile does not need to be imprisoned if his identity, banking, travel, work and reputation can be quietly degraded across systems that all claim neutrality.

The old danger was propaganda: a lie people might believe.

The new danger is administrative: a lie systems might process.

The future of authoritarian power may not always look like a knock on the door. It may look like an automated decision no human owns, based on a label no one verifies, produced by a regime no one admits they obey.

Democratic institutions were built to fight force. They are less prepared to fight formatted suspicion. The authoritarian API works not because the lie is believed, but because the lie gets processed. Procedure, without adversarial awareness, will keep running it.

Notes

[1] PEN International, case file “Andrej Hniot,” and “Serbia: Belarusian filmmaker and journalist Andrej Hniot to be retried,” December 2024: the arrest on 30 October 2023, seven months in Belgrade Central Prison, transfer to house arrest on 5 June 2024, and the Court of Appeal’s order of 11 September 2024 sending the case back for retrial. On the release and the statutory one-year ceiling on pre-extradition detention, see Associated Press reporting of 1 November 2024 and Balkan Insight, “Belarus Activist Freed From Detention in Serbia Leaves for EU,” 1 November 2024, quoting defense counsel Filip Sofijanic that proceedings remained pending and that Belgrade’s High Court was still awaiting documentation from Belarus when Hniot departed.

[2] Pozirk, “Interpol deletes red notice for filmmaker Hniot,” 19 July 2024, reporting the Belarusian Association of Journalists’ account of the Interpol General Secretariat’s letter to the EU delegation in Serbia, which found the notice inconsistent with Articles 2 and 3 of the Interpol Constitution.

[3] Freedom House, Collaboration and Resistance: Tracking Transnational Repression in 2025, April 2026. The database records 1,375 direct, physical incidents by 54 governments across 107 host countries between 2014 and 2025, including 126 new incidents in 2025. Detention (49) and unlawful deportation (48) were the most common tactics, and in at least 11 such cases perpetrator governments used Interpol notices, which the report reads as evidence that Interpol’s reforms have not yet closed the avenues available to member governments.

[4] Mikhail Khodorkovsky, testimony to the European Parliament’s Special Committee on the European Democracy Shield, 5 May 2026, published as “Khodorkovsky in the European Parliament: How the Kremlin Exports Repression to Europe.” Khodorkovsky also asked the committee to recommend guidance on banking de-risking, developed with DG FISMA and national regulators, obliging banks to distinguish genuine security threats from politically motivated accusations by authoritarian states.

[5] BBC World Service and Disclose, joint investigation published 26 January 2026. Pestrikov’s account, given publicly for the first time, describes frozen accounts and refused tenancy across the two years the Russian request was active, and its eventual deletion by the Commission for the Control of INTERPOL’s Files. On the distinction between notices and diffusions, the same reporting describes member states using Interpol’s messaging channels to trace people abroad in place of a notice the subject can challenge.

[6] INTERPOL, Annual Report 2024; see also Red Notice Monitor, “Increasing number of Red Notices issued,” 2025. 15,548 Red Notices published in 2024, a 27 percent rise and the highest annual total on record; 2,462 notices and diffusions refused or cancelled by the Notices and Diffusions Task Force, against 1,598 in 2023.

[7] BBC World Service and Disclose, 26 January 2026; and Amnesty International, “Global: Misuse of Interpol red notices to target dissidents a grave institutional failure,” 26 January 2026. The figure of at least 322 removals by the Commission for the Control of INTERPOL’s Files in 2024, and Russia’s position at the head of complaints to that commission, derive from the leaked files. Disclose reports that fewer than 10 percent of approximately 86,000 active Red Notices have been disclosed publicly, and that as of September 2024 the largest holders were Russia (4,817), Peru (4,457) and Tajikistan (3,493).

[8] Financial Action Task Force, High-Level Synopsis of the Stocktake of the Unintended Consequences of the FATF Standards, 27 October 2021. The project was launched in February 2021 and examines four themes: de-risking, financial exclusion, undue targeting of NPOs, and curtailment of human rights with a focus on due process and procedural rights. Recommendation 8, governing the treatment of non-profit organizations, was revised in 2023.

[9] LSEG, product documentation for World-Check One Media Check, describing machine learning and intelligent tagging applied to more than 13,000 vetted sources across 24 languages, clustering of content into discrete events, and delivery by platform or API.

[10] The Fifth Anti-Money Laundering Directive widened the range of obliged entities required to conduct open-source media checks and introduced automation requirements for adverse-media screening; the Sixth added predicate offences that extended the screening scope further. See also European Center for Not-for-Profit Law, How AI Is Powering Transnational Repression, 2026, a scoping study covering algorithmic risk assessment, automated blacklisting and AI-enabled surveillance in the transnational-repression context.

[11] Freedom House, Collaboration and Resistance: Tracking Transnational Repression in 2025, April 2026, recording the May 2025 Red Notices against Ivania Cruz and Rudy Joya, the UN experts’ finding of a direct connection to their human rights work, Interpol’s revocation on political-motivation grounds, and the continuation of Spanish extradition proceedings until asylum was granted in 2026.

The birthrate never recovered. Italy is older than its politicians expected. Germany has more care facilities than classrooms in some regions. Spain has become a country of empty interiors and crowded coastal cities. France still argues about identity, housing, policing, and Islam, but now does so inside a society where the median voter is older, the welfare state is more expensive, and the old promise of integration through work is less convincing than it used to be.

For most of the early twenty-first century, this was supposed to settle the migration debate. Europe was aging. Africa was young. The Mediterranean was not only a border. It was a demographic equation. One side had old people and labor shortages. The other had young people and need. However bitter the politics became, the economic argument seemed durable: Europe would need workers, and migration would supply them.

But by the 2040s, the equation looks less certain.

Europe still needs people. It needs nurses, builders, engineers, technicians, caregivers, soldiers, founders, children, and citizens who believe the future belongs to them. Automation has not abolished human need, and no serious society can be run as a warehouse with a flag.

But the old panic has softened. Ports move with fewer hands. Warehouses run with fewer pickers. Farms use more machines. Translation is cheap. Clerical systems are thinner. Public agencies process more through software. Hospitals still need human judgment, but scheduling, monitoring, triage, documentation, and routine support have been absorbed into systems. Care remains morally human, but parts of care have become infrastructural.

The result is not abundance, and not the end of work. It is something stranger: a society that has shed its most routine labor and remains aging, strained, and functional.

That is the future that changes migration.

The Bargain Was Labor

The migration argument of the labor civilization depended on a sentence that sounded practical enough to survive moral disagreement: they will work. That sentence turned admission into exchange. The migrant was not only asking to receive protection, schooling, housing access, welfare eligibility, legal standing, and eventually political membership. He was also bringing labor into a society that needed it.

Labor was never the only door. Families reunified, refugees were sheltered, empires left obligations. But labor was the justification both sides of the argument could read.

This is why labor shortages mattered so much to the pro-migration case. A farm without pickers, a hospital without nurses, a care home without staff, a restaurant without kitchen labor, a construction site without crews: each turned migration from a moral preference into an economic necessity. The host society could dislike the cultural consequences and still accept the bargain because the bargain was legible.

The anti-migration case attacked the other side of the same settlement. It argued that labor was not the only variable. Housing, crime, schools, welfare, identity, trust, and public order also mattered. Work did not automatically compensate for social strain. A worker could fill a job while also changing the neighborhood, the classroom, the welfare system, the policing environment, and the meaning of national belonging.

But even this objection remained inside the labor world. It did not deny that migrants could work. It argued that work was not enough.

Both sides therefore shared a hidden assumption: the migrant mattered because the worker mattered.

That shared assumption is the one now coming apart, before the politics is anywhere close to resolved.

The Worker Was the Passport

The modern border was never only a line around territory. It was a filter around membership. To cross into a rich country was to enter a thick inheritance of wages, rights, courts, schools, hospitals, public order, social insurance, political claims, and intergenerational possibility.

Labor made that entry explainable. Call it the labor passport: the worker’s usefulness was the document that turned need into belonging.

The migrant could say: I am not only asking to be carried by this system. I am entering it by helping it function. My work is my justification. My children will inherit the national story because I first entered through the economic one.

This was not merely economic. It was moral grammar. Work translated need into legitimacy. It made the poor more than recipients, the foreigner more than a stranger, and the future citizen more than an act of generosity. The host country could imagine incorporation because the migrant had a role before he had full belonging.

That grammar is now under pressure. If AI and robotics reduce the need for human labor across enough sectors, the migrant does not merely lose bargaining power. The labor passport begins to lose its authority. The host society may still feel compassion. It may still need some workers. It may still admit refugees, spouses, students, founders, nurses, and skilled professionals. Sectoral migration can survive, even grow. But the broad argument that migration solves the demographic future becomes less convincing.

This is not because migrants are uniquely unnecessary. It is because necessity itself is being redistributed.

Native citizens are not protected from this. They are merely already inside the line. A warehouse worker replaced by automation, a graduate filtered by AI hiring systems, a call-center employee displaced by agents, a junior analyst compressed by software, and a migrant whose labor is no longer needed are not the same political figure. But they are symptoms of the same deeper shift: labor is losing its power to explain why a person belongs.

The border can still decide who enters. It cannot restore the old reason entry mattered.

The Demographic Panic Fades

Population decline used to be frightening because the industrial state was built on bodies. Bodies staffed factories, paid payroll taxes, filled armies, cared for the old, consumed houses and cars, and gave the nation a sense of forward motion. A shrinking population meant fewer workers, fewer taxpayers, fewer caregivers, weaker consumption, and a thinner military pool.

That logic has not vanished. It is still true in many sectors and many places. But it is no longer the whole truth.

A society with advanced automation does not experience demographic decline in the same way as a society without it. A country with AI administration, robotics, reliable energy, compute infrastructure, automated logistics, and deep capital can absorb population decline differently from a country whose only surplus is young labor. The scarce input changes. Sometimes the bottleneck is no longer the worker. It is electricity, chips, land near power, water, cooling, grid permission, institutional trust, or the right to build the systems that make fewer workers necessary.

This changes the migration question. The claim that Europe must import labor to survive assumes that labor remains the central constraint. But if the productive system moves its binding constraint from bodies to substrates, demographic decline no longer carries the same meaning. It remains a social problem, a fiscal problem, a care problem, and a legitimacy problem. It is no longer automatically a production problem. If anything, automation sharpens the fiscal side of the question: machines produce output but pay no payroll taxes, and a taxed wage is still easier for a state to see than machine income. The migrant who works may become more valuable to the treasury, not less.

That distinction matters. A country may still need children without needing population growth in the old economic way. It may still need citizens without needing continuous labor expansion. It may still need human beings for meaning, loyalty, defense, families, and care, while needing fewer human beings for routine production.

The old fear was that aging societies would run out of workers. The new fear is that they will run without needing enough of them.

The Young World Loses Its Leverage

Now imagine the other side of the Mediterranean.

A young African state enters the same future with a growing population, weak industrial capacity, fragile public institutions, unreliable electricity, limited control over compute, and schools that credential more people than the economy can absorb. Its young people want what young people have always wanted: work, status, marriage, housing, dignity, movement, and a believable future.

The scale is not rhetorical. By 2050 the World Bank expects roughly 740 million more working-age people in Sub-Saharan Africa, the fastest increase anywhere on earth. Up to 12 million young people reach the labor market each year; about 3 million formal wage jobs are created to meet them. The pressure is already there, before automation touches it.

For decades, the development story told them that demography could become leverage. A young population could be a demographic dividend. If institutions improved, infrastructure developed, and global capital arrived, workers could move from farms to factories, from factories to services, from services to higher-value sectors. Migration could relieve pressure through remittances. Manufacturing could absorb labor. Education could turn population into capacity.

That sequence is exactly what the coming automation threatens.

If rich countries automate faster than poor countries industrialize, the development ladder does not merely become harder to climb. It shortens while people are still waiting to step onto it. Manufacturing absorbs fewer workers. Services are mediated by platforms. Translation becomes cheap. Call centers shrink. Administrative outsourcing compresses. Low-end coding becomes more automated. Credential work becomes less scarce. Even migration routes narrow because destination countries have found substitutes for some of the labor they once imported. The political coalition narrows with them: the employers who once lobbied to open the border stop showing up to argue for it.

The poor country still has resources. It may have minerals, land, ports, solar potential, military geography, carbon sinks, migration pressure, and instability that richer states want contained. But it may have less leverage through labor. Nor will the line run neatly around the continent. A few states will turn minerals, energy, and position into real leverage; their neighbors will be managed.

This is the cruel break. The old development promise was not only that poor countries would receive aid. It was that their people would eventually be needed by the world economy. First as workers, then as consumers, then as citizens of modern states.

The new promise may be colder: not incorporation, but maintenance.

A population can be fed, vaccinated, connected, counted, biometrically identified, educated through screens, stabilized by aid, governed by imported software, and prevented from moving. It can be kept alive without being given a path to power.

That is not old colonialism. It does not require foreign governors or annexation. It does not even require hatred. It can be done through development banks, resource contracts, cloud providers, humanitarian platforms, security partnerships, debt programs, border deals, climate finance, digital identity systems, and migration compacts.

The country remains sovereign in form. Its future is administered through systems it does not control.

The Flag Stays. The Operating System Moves.

The nation-state may survive longer than sovereignty.

A weak state can keep its anthem, parliament, ministries, elections, courts, and seat at international organizations while losing control over the systems that determine its future. Food systems, payment rails, cloud infrastructure, identity layers, energy finance, logistics corridors, insurance, mining rights, education platforms, migration quotas, security training, and AI governance can all be shaped elsewhere.

The country remains visible as a state. Operationally, it becomes a surface. What it keeps is formal sovereignty; what it loses is operational sovereignty, continuity under conditions it no longer sets.

This is why the future does not need to call itself empire. Empire is too visible. It creates responsibility. It invites resistance. It requires administration. Managed dependence is cleaner. The weak state signs the agreement. The lender sets conditions. The platform processes identity. The donor stabilizes the crisis. The security partner trains the force. The resource company extracts. The cloud provider hosts. The border agreement contains movement.

No single actor rules the country. The country is governed through dependence. What coordinates it is not a capital but interoperability: standards, financing conditions, and shared exposure to the same risks.

This is how a world can preserve the form of sovereignty while hollowing out its operation. The line on the map remains. The meaningful boundary moves elsewhere: into compute, energy, capital duration, standards, institutional trust, and access to the substrate.

The future border may not simply divide countries from countries. It may divide populations incorporated into the operating layer from populations administered outside it.

The New Humanitarian Empire

The darkest future is not that the rich world abandons the poor world. Abandonment is visible. Management is harder to name.

A managed population can receive aid, mobile money, food imports, vaccines, digital schooling, biometric identity, security support, climate adaptation funds, and periodic emergency rescue. It can be monitored for famine, violence, disease, migration risk, radicalization, and debt distress, stabilized before collapse and contained before arrival.

The system can truthfully say it is helping. That is what makes the structure powerful. A person can be helped and demoted at the same time. A country can receive support while losing position. A population can be protected from catastrophe while being denied entry into the systems that decide the future.

The old industrial world exploited labor. It needed bodies to work, so it fought over wages, unions, factories, plantations, mines, and class power. What is forming now is a colder relation. Some populations may no longer be valuable primarily as labor. They may be valuable as resource holders, risk pools, migration sources, security problems, climate exposure zones, or humanitarian obligations.

Exploitation required incorporation into production. Management does not. This is the difference between being used and being maintained.

The old poor feared being exploited. The new poor may face something worse: not being needed enough to exploit.

Migration Becomes a Claim on Membership

People will still move.

They will move because local institutions fail, because climate pressure rises, because violence spreads, because screens make rich-world life permanently visible, because development no longer feels believable, and because managed life is still life without a future.

But the destination countries will hear the demand differently.

In the labor civilization, migration could be heard as a request to work. In the world taking shape, it will increasingly be heard as a request for membership. That is much more threatening. A worker can be assigned a place. A member has claims.

This is why migration politics may become harsher even as economies become more automated. The issue will not only be jobs. It will be housing, schools, benefits, policing, religion, public order, status, and the right to make demands on a system whose own citizens feel less secure.

The migrant becomes the visible outsider asking for entry. The native citizen becomes the insider slowly discovering that entry no longer guarantees weight. Both are facing the same hidden question: who counts when contribution no longer explains membership?

To describe this is not to endorse it. No one deserves exclusion for being unneeded; legitimacy runs on what societies believe they require.

That question cannot be solved by saying citizens come first. Citizenship decides priority. It does not explain purpose. A society can defend its border and still fail to form its own people. It can exclude outsiders and still produce insiders who are provisioned, entertained, monitored, and politically flattered without being needed by the systems that govern them.

The migrant is not the exception to the future.

He is its preview.

Citizens Can Be Managed Too

Closing the border cannot solve the deeper problem. A country can stop migration and still produce managed citizens. It can defend national membership and still lose the institutions that made membership meaningful.

The old system made membership through work, family, education, service, locality, and institutional passage. People were not merely counted. They were formed. They were given roles that connected discipline to standing and standing to responsibility. The system often failed, excluded, humiliated, and exploited, but it still offered a grammar of incorporation.

A society can preserve the shell while losing the formation. It can give citizens benefits without roles, entertainment without consequence, connection without authority, education without admission, and political speech without operational power, telling them they are sovereign while routing the real decisions through systems they do not understand and cannot contest.

This is where the migrant and the native begin to converge. The migrant may be managed first because he is easier to classify as outside. The native may be managed later because the system no longer needs enough of him to keep the old bargain intact.

The divide this opens is not rich versus poor in the old sense. It is the divide between the formed and the managed: those trained, trusted, and admitted into systems that still require judgment, and those provisioned by systems they do not operate.

Migration is the first place this becomes politically visible because borders force the question early. Who may enter? Who may claim? Who must be processed? Who gets a path? Who receives aid without membership? Who becomes a permanent object of administration?

But the border is only the surface. The deeper question is whether entry into a rich society still means entry into a meaningful role.

The Real Border Is Not the Map

The old border asked where the state ended. The new border asks where meaningful participation begins.

That border may run through passports, but not only through passports. It may run through compute access, energy security, institutional trust, platform permission, educational formation, legal standing, biometric identity, security clearance, and the right to make claims that must be answered rather than merely processed.

A rich-country citizen with no formation may have formal membership but little operational weight. A foreign engineer inside trusted infrastructure may have more real participation than a native citizen outside the systems that matter. A resource-rich country may possess land but not sovereignty. A small advanced state may possess little land but deep operational sovereignty because it sits inside the substrate.

This is the migration question after labor. Not whether people move. Whether movement still leads to membership.

The humanitarian answer says every human being has dignity. It is right to say so. But dignity alone does not tell a civilization how membership works when labor no longer performs that function.

The nationalist answer says citizens come first. It is politically powerful. But citizenship alone does not explain what citizens are for when the productive system needs fewer of them.

The development answer says poor countries need investment, education, and institutions. It remains true. But it weakens if the world economy no longer offers mass incorporation through labor-intensive ascent.

The technocratic answer says automation will create new jobs. It may create some. It may create important ones. But a society cannot base its legitimacy on the hope that every displaced claim will be reabsorbed somewhere else.

The deeper question is simpler. Can a civilization preserve membership after necessity?

For two centuries, the worker carried the passport. Labor was the document that turned need into belonging, the reason a stranger could become a member and a member could believe he was needed. That document is expiring, and not only for the migrant.

The rich world may discover that it no longer needs more workers. Then it will discover that this was never only a migration problem. It was the first rehearsal for a world where being needed is no longer the reason to be let in.

Migration was an argument about labor. It is becoming an argument about membership, and membership is the one thing a border can police but cannot create.

Notes

[1] The demographic background is drawn from the United Nations World Population Prospects 2024, the twenty-eighth official UN round of population estimates and projections, covering 237 countries and areas.

[2] The migration-policy background is the OECD’s International Migration Outlook 2025, which continues to analyze migration primarily through labor-market inclusion, policy change, and sectoral shortages, including health-professional migration.

[3] The automation premise is not that all human work disappears, but that exposure is broad and that demography itself accelerates adoption: the economies aging fastest have adopted automation fastest. See Daron Acemoglu and Pascual Restrepo, “Demographics and Automation,” Review of Economic Studies 89, no. 1 (2022).

[4] World Bank, Sub-Saharan Africa regional overview: a net increase of about 740 million working-age people by 2050, the fastest of any region, with up to 12 million youth entering the labor market each year against about 3 million new formal wage jobs currently created annually. The World Bank’s October 2025 Africa’s Pulse states the 2025–2050 increase as more than 620 million; the figures differ by baseline year, not by direction.

[5] The management machinery already exists in early form. The EU’s 2023 memorandum of understanding with Tunisia and its 2024 strategic partnership with Egypt tie financing packages to migration containment, while donor-funded biometric identity systems already administer registration for populations no destination state intends to admit.

The deeper shift is simpler and stranger. For a few centuries, every major institution assumed intelligence was scarce, human, expensive, and slow to train. Schools certified it. Firms hired it. Wages priced it. States taxed it. Professions guarded it. That assumption is now failing, and the institutions built on top of it are starting to go blind.

This project has one claim at its center: cognition is becoming infrastructure. Judgment, search, writing, diagnosis, coding, coordination, the things people used to do inside their own heads, are turning into services that can be rented, embedded, ranked, denied, or governed from upstream. The machines don’t have to replace us for that to matter. They only have to change where intelligence lives.

Once intelligence moves, the institutions built to measure it lose their grip. A school that graded effort now faces a machine that produces competent work on demand. A labor market that paid for scarce skill now faces tools that close skill gaps. A tax system that reads wages now faces value draining into compute and capital. None of this requires the machines to be dangerous. Obedient AI is enough.

A note on what you’ve arrived at

This is not a newsletter. Newsletters are written to be current, and their pieces accumulate without composing. This is a corpus, and it is built to be read against itself.

It works at two distances from the same event. There is a near-term layer that diagnoses the transition we are already inside: political economy, state capacity, class, legitimacy, infrastructure. And there is a long-horizon layer that describes the structure this settles into over decades and centuries. Neither sits above the other. They are two lenses on one object, and they are kept deliberately separate, because the argument you can make about 2029 is not the argument you can make about 2129, and blending them produces confident nonsense in both directions.

Underneath that sits a discipline most public writing skips. Every concept has one home essay that defines it, and when it reappears elsewhere it points back rather than being quietly redefined. Every statistic anchors one argument, once. New pieces are checked against the existing ones before they are written, so the corpus builds instead of circling. That is a slow way to work and it costs reach. It is also the only way a hundred essays turn into an argument rather than a pile.

What the usual debate misses

Three stories dominate. Each is partly true and too small.

The productivity story says AI makes us more efficient, output rises, the economy adjusts. Probably right about output. But producing more is not the same as distributing it, and a society can grow richer while more of its people lose bargaining power and any reason to feel needed.

The replacement story says AI takes jobs, so we need retraining or UBI. Also partly right. But the wage was never only a paycheck. It was how modern societies handed out income, status, routine, adulthood, tax revenue, and a sense of belonging. Weaken it and you get a legitimacy problem, not just an unemployment number.

The safety story says models might deceive or escape, so we need alignment and control. That matters. But even a perfectly safe model can concentrate power. The harder question is what obedient AI lets institutions do.

The core idea

A synthetic civilization is not a world where everything is fake. It’s a world where the core operations of society increasingly run through artificial, model-mediated systems. Knowledge, judgment, allocation, administration, markets: more and more of what people used to do directly now passes through machines that decide what can be seen, priced, rewarded, and accessed.

So the useful question stops being “what can AI do?” and becomes “what parts of civilization now depend on AI-mediated cognition?” Ask it that way and the transition stops looking like a tech story. It looks like a story about state capacity, political economy, class, and legitimacy.

Four claims to carry around

The wage was a legitimacy machine. Work, earn, pay, belong: that bargain organized modern life. AI doesn’t need to erase every job to break it. It only needs to make human development less economically necessary. What you’re left with is a population that is still present, still equal, still entitled, and quietly less needed by production. That’s a legitimacy problem wearing a labor-market costume.

The state was built to read wages. Fiscal systems learned to see the economy through payroll, because payroll is visible and regular. When value migrates upstream into compute, chips, model access, and automated surplus, the state keeps reading a surface where less and less of the money actually is. A government that can see wages but not compute is a monarchy still taxing land after the wealth has moved to the factories.

The market becomes an interface. The old picture had a consumer standing outside the market, choosing. Now agents search, rank, filter, and recommend, and platforms decide what appears at all. The market survives. What changes is that it now runs through a layer that shapes what you see before you choose. The live political question shifts from “what did the market decide?” to “who built the menu?”

Power is going physical again. This is the twist people miss. When cognition becomes infrastructure, power doesn’t float off into the cloud. It drops back down to the things the cloud is made of: energy, land, cooling, chips, latency, and the kind of capital that waits a decade to pay off. Data centers start to behave like territory. Whoever controls the substrate controls what runs on top of it, which leads to the last claim.

And behind all four sits the real one:

The deepest power is allocation. Not raw intelligence, allocation. Who gets compute, ranking, credit, visibility, a diagnosis, an opportunity. Who gets filtered out before any human decision-maker even looks. AI systems don’t only answer questions; they organize access. A society can stay formally democratic while its real allocation decisions move into permissions, rankings, and eligibility filters that ordinary citizens cannot see or contest.

Why now, and the thing to remember

The lag is already visible. Students use AI faster than schools can redesign the test. Firms automate faster than governments can redesign the tax code. States adopt these systems faster than the law can say where responsibility lives.

But the old world doesn’t end in a single dramatic scene. People still work, vote, pay, and raise children. The surface holds. Underneath it, the operating system changes: more cognition outsourced, more judgment mediated, more allocation automated, more legitimacy borrowed from systems no one quite understands.

That’s the point worth passing on. Power isn’t collapsing in this transition. It’s relocating. The machine still works, and that is exactly the problem. A broken machine gets fixed. A working one that has quietly moved the important decisions somewhere you can’t see is much harder to argue with. The frontier lab doesn’t need to conquer the state. It only needs to become the environment the state, the market, and the citizen all run on.

Why this project exists

For the past eight months I have spent a great deal of time in the rooms where this is being argued out. Substack, X, Bluesky, the papers, the podcasts, the comment threads. Frontier researchers with the best available view of the models. Silicon Valley, which is building the substrate. Politicians who will have to write the rules. Economists who own the tools for measuring what happens next.

What struck me was not the disagreement. It was the shape of it.

Almost every axis of the public argument is binary, and almost every binary is about the technology rather than about power. Open weights or closed. Accelerate or pause. Safe or dangerous. Jobs destroyed or jobs created. Good for humanity or bad for it. Serious people sit on both sides of each of those, and you can win any one of them outright without having said a single thing about who allocates compute, or what a tax system does when payroll stops reporting the economy, or what happens to a democracy whose real decisions have migrated into rankings and eligibility filters. The debate is being conducted in a vocabulary that cannot see its own object.

Part of that is honest specialization. The people closest to the models can tell you what the systems will do next, and have no professional reason to think about fiscal capacity. Economists can model the labor shock and rarely reach the legitimacy shock underneath it. Policy people draft the rule without naming the layer at which it will later be quietly reset. Each constituency holds a real piece. Very few are holding the map.

I have looked hard for someone else holding it: a body of work that takes the labor question, the fiscal question, the infrastructure question, the allocation question, the legitimacy question and the long-horizon question and keeps them inside one frame, with a consistent vocabulary, in public, at length. I have not found it. If it exists, send it to me. I would genuinely rather read that than keep writing this.

This is an independent project. No institution, no funding, no press, no algorithmic momentum. It is not built for reach, and it grows slowly, which is the correct outcome. The readership is small and unusually good: people who work on these questions professionally, who read closely and say little. That is the audience this was built for, and it is the one that has been steadily arriving.

The project is pseudonymous, which rules out audio and video and makes correspondence slow. Written questions, disagreements, and counterexamples are still the fastest way to reach me. And if one person you know should be reading this and isn't, send it to them.

Where to read next

The near-term layer is the place to start. Seven essays, in order:

1. Output Without Income. The anchor: production keeps working while the machinery that shares out its gains comes apart.

2. The Wage Was a Legitimacy Machine. Why the wage was never just income, and what breaks when it thins.

3. The Market Becomes an Interface. Why markets don’t disappear, they get mediated, and why that moves the real decision upstream.

4. Capital Without Justification. What happens to ownership’s moral story once labor stops being the thing that earns.

5. The Compute Estate. Why compute becomes the scarce ground of the age, owned and rented like land.

6. The Return of Physical Power. Why the age of infrastructure drags power back down to energy, land, cooling, and chips.

7. Who Elected Anthropic? The legitimacy question underneath it all, when a private lab writes public rules.

You don’t have to agree with every essay. The point is to change the level of the question. AI is not only a tool, a sector, a labor shock, or a safety problem. It is becoming the environment the rest of civilization runs inside. That is the transition, and this is the map.

You are eighteen. You are bright, you are not rich, and you did not get into the school whose name would have done half the work for you. You can borrow six figures against a degree in a field that is being automated while you study it, or you can sign a piece of paper.

Sign the paper and the following happens. Someone feeds you, houses you, and pays you from the first week, before you are worth anything to anyone. Someone trains you in a technical specialty, at their expense, on their equipment. Someone puts you through a background investigation and, if you pass, grants you a clearance that no ordinary employer can hand a twenty-two-year-old. Someone gives you responsibility at twenty that a graduate trainee will not touch at twenty-six. You come out at twenty-two with four years of paid operational work, references from people who watched you do the job, and no debt. If you still want the degree, they will buy it for you: four years of active duty clears the threshold for the full Post-9/11 GI Bill, which pays in-state public tuition outright and adds a housing allowance on top. [1]

The catch is real and it is not small. For four years they own your location, your schedule, your appearance, your conduct, and your right to leave. They can also send you somewhere you do not come back from, or do not come back from whole, and no bonus prices that. You cannot resign because your supervisor is a fool. You may not get the specialty you were promised, and if you fail the school they will simply put you somewhere else. It is not a four-year apprenticeship with a refund policy.

Weigh it honestly against the alternative and it holds up. For a large number of capable eighteen-year-olds, this is now the better deal.

And that is the part worth stopping on. Not that the advice is good. That it is true.

Nobody made the military better

The interesting thing about the pitch is that it has not changed. Almost none of it is new. The military was making that exact offer thirty years ago, and thirty years ago it was a fallback, the thing you did when the other doors were shut. The doors were the point. The military was what remained after them.

The offer did not improve. The alternatives collapsed around it.

They collapsed in a specific way, and it is the part that gets missed when this turns into a debate about whether college is worth it. The real question is what a young person is supposed to walk through in order to become someone who can be trusted with things, and who is still building that passage.

What formation actually requires

Think about what it takes to turn an eighteen-year-old into a competent twenty-two-year-old. Not to certify them. To make them.

It takes someone willing to carry a person who is, for a long stretch, a net cost. Someone who will pay them before they produce, correct them when they fail, and stay attached to the outcome long enough that the correction matters. Formation is expensive and it is slow, and no institution does it out of goodwill. Institutions form people when they cannot get out of the consequences of not forming them.

Foreclosure alone will not do it. A prison cannot get rid of you either, and it forms nobody, because it needs nothing from you at the end. The institution has to be stuck with you and it has to need what you would become. Then it has no choice but to build it.

That is the whole mechanism. Formation survives where exit is foreclosed.

Hirschman named the escape route sixty years ago: when leaving is cheap, an institution stops having to listen and stops having to repair. [2] He was writing about the member who walks out. The version that matters now is the one where the institution holds the exit and the person cannot.

Look at what happened to the places that used to form people, and you find the same thing every time. Not a loss of virtue. An acquisition of an exit.

The ordinary university found the simplest one. Its revenue arrives at enrollment, before it has formed anything, and it bears no cost whatsoever if the person it graduates turns out useless. Its reputation is set by who it admits, not by who it produces. It selects, and it vouches, and it has quietly stopped doing the thing in the middle, because nothing in its position requires it to. A university that graduated nobody employable for four straight years would notice a decline in applications. A university that graduated nobody employable for one year would notice nothing at all.

You can watch the consequence in the numbers. In the first quarter of 2026, about forty-one percent of recent American graduates were working in jobs that do not require the degree they bought, and the unemployment gap between young degree-holders and young people the same age without a degree had closed to a point and a half. [3] None of that reaches the university as a signal it has to answer. The money was collected at the start.

The employer found a different exit and found it recently. Formation was never the employer’s goal, only a byproduct of needing the work done. The junior rung existed because the boring, careful, half-skilled tasks had to be performed by somebody, and the somebody got better while performing them. Being a side effect of production is exactly what made it reliable and exactly what made it disposable. The moment the tasks could be done without a person, the rung went and the training went with it. It was scaffolding that happened to hold people up, and it came down the moment it stopped holding anything else. This is the ladder that is gone, described from the demand side. Somebody has to want the junior work done in order for the junior to become senior. Nobody does anymore.

And then there is the elite university, which looks like the exception and is in fact the purest case in the set.

It seems to form people. The seminars are small, the standards are enforced, the graduates are formidable. But run the mechanism and ask where its exit is, and the answer is that its exit is the admissions office. An institution that selects hard enough does not have to form anybody. It takes in the already-formed: the products of decades of enforced friction that somebody else paid for, in households and schools and summers designed to produce exactly the person the admissions committee is looking for. It receives them finished. It confers the name. It sends them onward, and the outcomes are excellent, and almost none of the earnings difference was made on campus.

That last claim is checkable, and it has been checked. Compare students by the schools that accepted them rather than the schools they attended, and the earnings premium for going to the more selective one largely disappears: the ones who turned it down and went elsewhere ended up earning about the same. [4] There is one exception, and it is the finding underneath the finding. Students from poor families do gain. They are the ones who arrive without the decades of friction already behind them. The elite university adds something to the students it admits fewest of, and almost nothing to the ones who fill its class.

The ordinary university escaped formation by not caring about outcomes. The elite university escaped it by controlling the input so tightly that outcomes take care of themselves. It charges the most and it forms the least, and nobody notices, because its graduates really are exceptional and were before they arrived. So the rule does not have an exception. It has a loophole, and the rich have it.

The last complete institution

Which leaves the one that cannot get out.

Yes, the military can discharge you. It can wash you out of a school, reclassify you, deny you advancement, separate you. But look at what discharging you accomplishes, which is nothing. It does not produce the twenty-two-year-old the service needs, and there is no lateral market for a competent young sergeant. Nobody else is making them. The service cannot buy one, cannot poach one, cannot contract one in.

And underneath that, the reason no such market can exist: the person may be ordered to kill, and may be ordered to die. A contractor can be paid to accept the risk. No contractor can be placed under the duty to obey, to remain, and to bear it without limit, and that duty is the thing being bought. It has to be someone the institution made, and made from the start, because what is being built is not a skill set. It is a person who stays in the seat when leaving is the only rational move. Everything decent about the military and everything frightening about it comes out of that one fact.

So it pays you before you are useful, and corrects you when you fail, because it will be living with the result either way.

It is also the only place left that still does all four things under one roof. It selects you, with a test and a physical and a background check. It forms you, over years, at its own cost. It vouches for you, in a way that is legible to employers and to the state, and the clearance is the purest instrument of that vouching. The government’s own adjudication standard calls it the whole-person concept: an examination over a sufficient period, weighing the variables of a person’s life, all available information, favorable and unfavorable, ending in an affirmative determination. [5] Compare that to a transcript. A degree is a standardized record that you completed a curriculum. A clearance is an individual judgment that you can be trusted. And then the service places you, into work that exists, at a rank that means something.

Select, form, vouch, place. That used to be one continuous passage, and it ran through half a dozen institutions. Now the four steps have come apart. The university selects and vouches and skips the middle. The employer places and no longer forms. The credential market vouches without knowing anything.

Other institutions still form people, and the mechanism tells you which. The medical residency forms, and it cannot buy a lateral resident either. The union apprenticeship forms, and it binds both parties by design: the shop is stuck with you and needs the journeyman it is making. The religious order forms, and it takes you in and keeps you. But the certificate program forms nobody, and neither does the bootcamp. They charge tuition, they vouch on completion, they carry no one through the years when a person costs more than they return. They are the ordinary university on a shorter clock. The line does not run between college and the trades. It runs between institutions that are stuck with what you become and institutions that are paid before anyone finds out.

And every survivor on that list is a door you must already be qualified to reach. The residency wants a medical degree. The apprenticeship wants a place in a queue and somebody to vouch for you. The order wants a vocation.

The military screens too, and harder than the pitch admits. On the Pentagon’s own count, only about twenty-three percent of Americans aged seventeen to twenty-four can enlist without a waiver: weight, drugs, a record, a diagnosis, a test score take out the rest. [6] The services move that line when they need bodies, but most of the cohort does not clear it.

Look at what it does not screen for, though. Not what your parents did. Not what school you went to. Not who will vouch for you, and not what you can pay. This is the last complete institution in the country whose entrance exam is not, in some laundered form, the class you were born into. That is a narrower claim than the pitch makes, and it survives contact: not open to everyone, but the only one not closed by inheritance.

What it costs to be kept

The institution that cannot leave you is the institution that owns you: one fact, seen from two sides. It invests in you because it is stuck with you, and it controls you for the same reason. The care and the coercion come out of the same root. That is why every institution that still forms people looks, from the outside, faintly monstrous: the residency that works you past the point of safety, the order that governs your hours and your speech, the apprenticeship with its hazing and its closed shop. The unbearable parts are not a flaw in the arrangement. They are the arrangement.

And it forms you for itself, which is the concession the pitch never makes. The institution builds the person it needs, and you overlap with that person only as far as its needs happen to run. It will teach you to be reliable under pressure and it will not much care whether you learn to refuse. It can produce obedience and call it judgment. Some of what it makes of you does not travel, and some of what it makes of you should not. The formation is real. It was never designed with you in mind.

Which brings up the study that damages this argument.

The Vietnam draft lottery handed economists something rare: service assigned at random, no selection to correct for. A decade after the war, white veterans were earning roughly fifteen percent less than comparable men who never served. [7] Not a wash. A penalty.

Read one line further and the pattern turns familiar. Nonwhite veterans showed no lasting earnings loss at all. Set that beside the college finding and the two studies point the same way. The elite university adds almost nothing to the affluent and something real to the poor. Service cost the men who had civilian ladders waiting and cost nothing to the men who did not. Neither institution is a machine for raising your wage. Both seem to matter most to the people who arrive without an alternative already prepared.

Take the damage seriously, then notice what it damages. Formation is not an earnings premium and never was. The institution that forms you is not optimizing your income; it is producing what it needs. Four years is four years, and the man who spent them elsewhere was compounding something. What you get back is that you were made into somebody, by someone with a reason to bother, in a world that has largely stopped bothering. Whether that trade is worth it is the question, and anyone who tells you the arithmetic is obvious is selling something.

You cannot have the one without the other. Anyone offering you the formation without the possession is offering you a university.

A selection or a term

So there are two doors left into a formed adulthood, and only one of them actually forms you.

The first is bought, and what you are buying is the selection. The name, the sorting, the confirmation that you were already the sort of person who ends up fine. It is narrow by design, because the narrowness is the product, and it is mostly opened before you are old enough to know it exists.

The second is served. You pay in years and in autonomy, the institution takes possession of you for the duration, and at the end you are handed back with the skills and the clearance and the bearing of someone who has been through something.

Buy a selection, or serve a term.

What is gone is the thing that used to sit between them. The ordinary state university, the graduate scheme at the regional firm, the apprenticeship anybody could walk into, the long middle where an unexceptional young person of decent capability went in one end and came out the other as somebody. That road is now a toll booth without a road behind it, charging the price of a passage it no longer performs.

This is why the joke works, which is also why it is not really a joke. Telling a bright, broke eighteen-year-old to enlist sounds like a provocation, or like something a recruiter says. It is a description of what is left.

The part that should worry you

Here is the objection, and it is probably right.

The military’s junior rungs are made of exactly the material that is being eaten. Signals processing. Intelligence support. Network administration. Log review, first-line triage, the patient careful entry-level work that a young analyst has always done badly at first and well by the third year. That is not adjacent to what the models are taking. That is the center of it.

Some of it survives longer than the rest. The clearance is a judgment about a person and not a task, and no model can be adjudicated. The leadership pipeline is a genuine bottleneck, because someone still has to command. But the mass of junior technical work underneath both is being hollowed out on the same schedule as everybody else’s, and the service that finds it can run its intelligence shop with a third of the analysts will run it with a third of the analysts. It will just take longer to notice, because it does not answer to a quarterly number.

And there is a trap inside the part that looks most protective. The state will keep a human in the seat longer than a company would, because a model cannot hold a clearance, cannot be court-martialed, cannot be the name on the finding when something goes wrong. But that requirement preserves the position, not the passage. A person retained so that failure has a face is not a formed adult, and accountability can be satisfied by a signature. The billet survives. The years of doing the work badly and then less badly, which were the only reason the billet ever formed anybody, do not. That is the quiet way this door closes. The posting remains, and the thing that used to happen inside it is gone.

So the door narrows before it closes, and more slowly than the others, because the military’s need for the twenty-two-year-old is structural rather than economic, and structural needs are stubborn. But it is the same force, and the honest version of the advice has a clock on it.

Which is not a reason to tell the eighteen-year-old no.

It is open now, one of two that are, and the other mostly required being born through it. If you are eighteen and capable and unrich and the letter did not come, the thing in front of you is an institution that will take you as you are, pay you while you are worthless, teach you something real, put its name behind you, and hand you back at twenty-two with a start the university would have sold you at a loss.

Take it while it is there. Go in with the specialty in writing and the years counted and no illusions about who owns your calendar for the duration, because they will own it, and that is not the price of the formation. It is the mechanism of it.

The last institution that cannot get rid of you is the last one that will make you into someone.

Notes

[1] U.S. Department of Veterans Affairs, “Post-9/11 GI Bill (Chapter 33) rates.” Eligibility for 100 percent of the full benefit requires at least 1,095 days (36 months) of active duty; a standard four-year enlistment clears it. The full benefit covers net tuition and mandatory fees at a public in-state institution, with a monthly housing allowance and a books-and-supplies stipend on top. https://www.va.gov/education/benefit-rates/post-9-11-gi-bill-rates/

[2] Albert O. Hirschman, Exit, Voice, and Loyalty: Responses to Decline in Firms, Organizations, and States (Harvard University Press, 1970). Hirschman’s subject is the member or customer who can leave, and what their departure does to an organization that would otherwise have had to hear them. The application here inverts the parties: the institution holds the exit, and the person cannot.

[3] Federal Reserve Bank of New York, “The Labor Market for Recent College Graduates,” data through Q1 2026. Underemployment among recent graduates (those working in jobs that do not typically require a bachelor’s degree) stood at 41.5 percent; unemployment for graduates aged 22 to 27 was about 5.7 percent, against 7.2 percent for those in the same age group without a four-year degree, down from 7.7 percent the previous quarter. https://www.newyorkfed.org/research/college-labor-market

[4] Stacy Berg Dale and Alan B. Krueger, “Estimating the Payoff to Attending a More Selective College: An Application of Selection on Observables and Unobservables,” Quarterly Journal of Economics 117, no. 4 (November 2002): 1491–1527. Matching students by the selectivity of the schools that admitted them, rather than the school they attended, removes the effect of the unobserved characteristics on which elite colleges select. The earnings premium for attending the more selective school largely disappears under that correction. The authors find one significant exception: students from low-income families did earn more for having attended the selective school. The follow-up study, using Social Security administrative earnings data for the 1989 entering cohort, reaches the same result (Dale and Krueger, “Estimating the Return to College Selectivity over the Career Using Administrative Earnings Data,” Journal of Human Resources 49, no. 2 (2014): 323–358). https://www.nber.org/papers/w7322

[5] Security Executive Agent Directive 4, “National Security Adjudicative Guidelines,” signed by the Director of National Intelligence on December 10, 2016, effective June 8, 2017. SEAD 4 establishes the single common adjudicative criteria across the executive branch and sets out the whole-person concept, under which an adjudicator weighs all available information about a person, favorable and unfavorable, past and present, to reach an affirmative determination of eligibility. https://www.dni.gov/files/NCSC/documents/Regulations/SEAD-4-Adjudicative-Guidelines-U.pdf

[6] U.S. Department of Defense, 2020 Qualified Military Available Study (Office of the Under Secretary of Defense for Personnel and Readiness). Approximately 77 percent of Americans aged 17 to 24 would not qualify for military service without a waiver, up from 71 percent in the 2017 study. The most common single disqualifiers are being overweight (11 percent), drug and alcohol abuse (8 percent), and medical or physical health (7 percent); 44 percent of disqualified youth are disqualified for more than one reason. Waivers are widely used and the standards move: roughly 17 percent of recruits received a medical waiver in 2022, up from 12 percent in 2013.

[7] Joshua D. Angrist, “Lifetime Earnings and the Vietnam Era Draft Lottery: Evidence from Social Security Administrative Records,” American Economic Review 80, no. 3 (June 1990): 313–336. The randomly assigned induction risk removes the selection bias that contaminates ordinary comparisons of veterans and nonveterans. In the early 1980s, white veterans earned approximately 15 percent less than comparable nonveterans; Angrist finds no evidence of a lasting earnings loss for nonwhite veterans. The estimates cover conscripts in a wartime, draft-era force and do not transfer directly to a modern volunteer enlistment in a technical specialty. They are cited here as the strongest available causal evidence against the argument, not as a description of it. https://economics.mit.edu/sites/default/files/publications/Angrist%201990%20-%20Lifetime%20Earnings%20and%20the%20Vietname%20.pdf

The state does not announce surplus.

It manages it.

There is no policy document that says the economy no longer needs you. There is no official declaration that your claim on the system has become administratively inconvenient. The signal never arrives directly. What arrives instead is a set of arrangements: transfers, platforms, therapeutic frameworks, attention systems, civic rituals, and algorithmic participation that absorb the presence of people the productive order has learned to operate without.

This is not malice.

It is adaptation.

When the wage weakens as the institution that turns contribution into claim, when markets move their decisive acts upstream into interfaces, when capital captures surplus from systems built on collective inheritance, the political system faces a structural problem it cannot name openly: what to do with the people whose claims remain after their necessity declines.

The surplus human is not someone without needs. She is not someone who has stopped wanting recognition, income, standing, or a role. She is someone whose claims remain intact while the system has quietly learned to operate around her.

Strictly speaking, there are no surplus humans. There are surplus claims: to income, to standing, to recognition, to a role, claims the productive system no longer knows how to redeem through necessity. That is why the phrase is so violent. It locates in the person what actually lives in the classification.

And she still votes.

That is what makes the politics of containment different from simple exclusion. The surplus population is not outside the political system. It is inside it, with full formal membership, with grievances, with ballots, and with an accumulating sense that the system has stopped making good on the promises it extracted compliance to honor.

The system still needs these people politically, morally, and symbolically after needing fewer of them productively.

That is the contradiction.

They may be less central to production, but they remain central to legitimacy. They may be less necessary to output, but they remain necessary to social peace. They may lose leverage inside the economy, but they do not lose their formal claim on the political order.

A society can automate tasks faster than it can automate belonging. It can reduce labor dependency faster than it can reduce the moral claims of the people labor once explained. It can thin the wage without yet building another grammar of membership.

Managing that interval, not resolving it, not explaining it, not eliminating it, is the central political task of the AI economy. And the toolkit for managing it is already taking shape.

The Claim Outlives the Necessity

The old political economy had a convenient alignment: people who were economically necessary also had political weight.

Workers could strike because production depended on them. Professionals could demand status because institutions needed their judgment. Young people could expect absorption because organizations required junior labor to reproduce senior capacity. The worker, the professional, and the apprentice all occupied recognizable positions inside the productive order.

That alignment is loosening.

AI does not need to eliminate labor to change this. It only needs to make large categories of human contribution less scarce. A firm can preserve visible output with fewer juniors. A sector can expand productivity while hiring less. An institution can keep delivering services while reducing the cognitive dependency that once gave workers leverage. The exposure estimate that opened this series, roughly 60 percent of jobs in advanced economies affected, some complemented and others facing substitution, is only the surface. The deeper change is that even workers who remain employed may find that the economy’s dependence on their specific contribution has weakened.

They are still present. They are simply less necessary.

The claims, meanwhile, do not disappear. Votes remain. Political voice remains. Consumption power, however thin, remains. The capacity for disruption remains. The expectation of recognition remains. The sense of entitlement to a recognizable role remains, because modern society spent a century teaching people that work was the bridge between existence and belonging.

That bridge does not dissolve the moment the productive system learns to cross it differently.

The political system cannot simply ignore the people left on the wrong side of that crossing. It cannot explain them away. It has to manage them. And because they vote, management has to be sophisticated enough to maintain at least the appearance of recognition while the underlying condition persists.

This is why the post-work problem is not idleness.

It is recognition.

People do not merely need income. They need a socially recognized reason to believe their presence matters. A transfer can prevent destitution. A platform can produce engagement. A credential can extend preparation. A wellness system can help manage distress. But none of these automatically restores the role that wage society once promised: a place where effort, status, income, discipline, and belonging appeared to form one coherent arc.

When that arc breaks, people do not become passive.

They become difficult to explain.

The Containment Toolkit

The welfare state was built around the necessary worker. Its premise was simple: the people who labored near the center of production had to be protected, educated, insured, and represented, because the economy needed them and the polity depended on them.

What is being assembled now starts from a darker premise. It manages people whose claims remain after their necessity declines. It does not integrate them into production. It stabilizes them outside the old grammar of usefulness: transfers without centrality, participation without leverage, therapy without explanation, credentials without absorption.

This successor arrangement has no name, because naming it would require admitting what it is for. So it borrows the welfare state’s vocabulary instead. Containment does not announce itself as containment. It announces itself as support, opportunity, wellness, engagement, empowerment, and care. The language is always that of provision. The structure is that of management.

The toolkit has five components.

The first is income stabilization: transfers, credits, supplements, UBI pilots, and platform-mediated earning that prevent destitution while doing little to restore the leverage or recognition that wages once conferred. These systems are genuinely necessary. They prevent acute suffering. They also decouple income from standing. This is welfare as maintenance: the material problem addressed, the legitimacy problem left intact.

The second is platform participation: the conversion of surplus presence into engagement metrics, content creation, digital identity performance, and algorithmic circulation. Platforms discovered before governments did that people who are no longer economically central still have enormous amounts of time, attention, and need for recognition, and that discovery became a business model. The creator economy, the gig economy, and gamified productivity tools all provide systems that feel like contribution without conferring the standing that wage labor once produced.

The third is therapeutic normalization: the expansion of wellness systems, resilience discourse, and individual self-management practices to absorb what are structurally produced conditions. This does not mean care is false or unnecessary; many people need it, and much of it is real. The political point is different. The vocabulary does the work. Resilience asks the individual to absorb what the institution has offloaded. Burnout describes as a personal energy problem what is often a structural recognition problem. Mindfulness teaches the worker to observe her anxiety without asking what the anxiety is accurately perceiving. Employers expand wellness programs alongside thinning career ladders; universities add counselors as the credential’s redemption value falls. None of this is coordinated. It is the path of least resistance: treating the distress is cheap and scalable. Treating the structure is not for sale.

The fourth is credential extension: the continuous expansion of degree requirements, certification programs, badge systems, and reskilling initiatives that keep people in preparation mode longer. It delays the moment when the gap between preparation and absorption becomes undeniable. It generates revenue for the education and training industries. And it allows institutions to continue demanding performance from people whose prospects they are, structurally, no longer in a position to honor.

The fifth is civic ritual: voting, national service proposals, participatory governance experiments, and digital democracy platforms that maintain the appearance of consequential participation while the decisive acts of allocation move upstream into interfaces, procurement systems, and technical architectures that no democratic process meaningfully governs.

Each component is real, and each provides something. None resolves the underlying condition.

Why Containment Works, For a While

Containment is not stable.

But it is durable.

It persists not because it satisfies people, but because it fragments the conditions that would produce a coherent political response to their situation.

The surplus human’s condition is produced by multiple overlapping systems: compute ownership, market interfaces, AI-mediated hiring, platform distribution, credential inflation, declining wage leverage, and the weakening of the entry-level roles through which institutions once absorbed and socialized the young. No single actor is responsible. No single villain can be identified, and no single reform would address the aggregate.

This fragmentation is politically significant for the reason the tenant condition made visible earlier in this series: the environment has no single landlord. There is no focal antagonist around which a movement can coalesce, only an aggregate condition that no individual actor is responsible for producing.

The surplus human may be angry. She will find it difficult to direct that anger coherently at its structural source. The platform provides content that gives her anger an object. The transfer gives her enough to prevent the desperation that produces organized disruption. The credential system gives her a task that feels productive. The therapeutic framework tells her that her difficulty is partly about mindset. The civic ritual tells her that she participates in the decisions that shape her life.

Together these systems do not eliminate the problem.

They manage it at a level below political crisis, most of the time.

When the System Has No Face, People Attack Its Body

When containment frays, the anger does not remain formless.

It finds targets.

And increasingly, the targets are not abstract systems or dispersed corporate actors. They are the physical infrastructure of the AI economy: the data centers, the energy contracts, the chip foundries, the cloud campuses, and the fiber corridors that make synthetic production possible.

This is already happening. In Virginia, the Prince William Digital Gateway, planned as the largest data center campus in the world, was approved by an outgoing county board in December 2023 after a 27-hour public hearing, voided by a circuit court in August 2025 for defective public notice, and abandoned by the county in April 2026 after years of resident litigation and organizing. By that point, the share of Virginia voters comfortable with a new data center in their community had collapsed from 69 percent to 35 percent in three years, and one industry tracker counted 48 data center projects blocked or delayed across the United States in 2025 alone. In Arizona, the Tucson city council voted unanimously in August 2025 to refuse annexation and water access for Project Blue, a 290-acre data center campus linked to Amazon, after weeks of public mobilization in a drought-stressed desert city. In Congress, members of both parties have moved from rhetoric to bills: the Clean Cloud Act would set emissions standards for data centers, and the bipartisan AI-Related Job Impacts Clarity Act would require companies to report the jobs AI eliminates. [1]

This is not irrational. It is the political economy of visible infrastructure meeting the political economy of invisible displacement. The data center is concrete, local, and photographable. The algorithmic process that filtered a job applicant out of consideration before any human reviewed the resume is not. The cooling tower uses water the community can measure. The model that depreciated a professional skill set operates in a cloud region the affected worker cannot locate on a map.

When economic displacement is structurally diffuse but its infrastructure is physically concentrated, the infrastructure becomes the target.

This dynamic has historical precedent. Luddite machine-breaking was not random vandalism. It was skilled workers attacking the specific capital equipment whose introduction was destroying their trade. The machines were the visible, attackable form of a broader economic transformation whose ultimate authors and beneficiaries were considerably harder to reach. [2]

What makes the advanced-economy version of this dynamic distinctively volatile is something the Global South comparison makes visible.

The condition this series tracks as emerging, a widening gap between the claims people hold and the place the formal economy will grant them, is not new. Billions outside the OECD have lived a version of it for decades, absorbed through informal labor markets, remittance dependence, NGO service delivery, and the particular containment toolkit of societies that never fully industrialized before deindustrialization arrived. These populations were never economically unnecessary. Many were indispensable in practice, through informal markets, household production, and precarious services, while being denied a stable place inside the official grammar of employment, taxation, and protection.

Lagos, Cairo, Dhaka, and Jakarta have long housed populations the formal economy could not absorb at scale and the state could not fully account for. What is different in advanced economies is not the condition but the political infrastructure surrounding it: the ballot, the zoning board, the congressional hearing, the class action, the investigative press.

Populations excluded from formal absorption in the Global South were often managed through scarcity, informality, and weak provision. Their advanced-economy counterparts are managed through abundance and strong institutions. The export is upward, and the destination has more tools for converting grievance into political pressure.

None of this is grounds for optimism. It means the political consequences will be louder, more institutionally disruptive, and harder to contain through the informal mechanisms that absorbed displacement elsewhere.

The contemporary version is not machine-breaking, at least not yet. It is zoning fights, permit battles, legislative restrictions, and the slow accumulation of local political resistance to the physical plant of AI: the same permission politics that now constrains the buildout itself. The political form is democratic and procedural. The underlying logic is the same: attack the thing you can see when the thing that harmed you cannot be named or reached.

The important political implication is that this targeting, however understandable, does not address the underlying legitimacy problem.

Blocking a data center does not restore wage leverage. Restricting AI energy use does not rebuild the junior layer. Passing algorithmic accountability legislation does not answer the question of what grammar should replace the wage as the primary explanation for why people have a recognized place inside the productive order.

Infrastructure resistance is real politics.

It is not a settlement.

The Grammar That Has Not Been Built

The AI economy has produced a bilateral legitimacy crisis at its center.

Labor can no longer say “I made this” with the force it once had. Capital can no longer say “I own this and therefore deserve the full return” without the moral thinness of that claim becoming more apparent as the surplus concentrates in systems built on collective inheritance.

The surplus human lives inside that double failure.

She is the person for whom neither grammar works. The wage no longer explains her income if she has any. The ownership claim does not apply because she does not own the systems she depends on. And no new grammar has been built that can tell her, and tell the political system, why she has a recognized place inside the world that synthetic production is creating.

One direction for that grammar is inheritance: the claim that AI surplus depends on accumulated civilization, on public science, public law, collective data, shared infrastructure, and inherited knowledge, and that membership in that civilization therefore carries a claim on some portion of what it produces.

That grammar does not require treating contribution as meaningless. It requires admitting that contribution has always depended on systems no individual built alone.

This grammar has begun appearing at the edges of mainstream policy. Even some proposals from frontier AI actors now gesture toward public stakes in AI-linked assets, citizen dividends, robot taxes, or social wealth funds. The motive may be strategic positioning as much as principle. But even as positioning, the gesture concedes the structural point: AI surplus cannot be narrated only as private corporate property. The inheritance is too visible. The collective substrate is too large. The concentration is too stark. [3]

But grammars take time to become politically legible. They have to be narrated, contested, institutionalized, and eventually taken for granted before they can organize stable legitimacy. The inheritance grammar is not yet politically usable at scale. It remains too unfamiliar, too weakly institutionalized, and too easily mistaken for either charity or confiscation.

In the meantime, the containment system fills the gap.

Provision Without Membership

The obvious objection to this account is that it is too cynical. Transfers are not containment; they are what a decent society owes people the market has failed. Therapy is not containment; it is care. Education is not containment; it is opportunity. Civic participation is not containment; it is democracy doing what democracy does. On this reading, the toolkit described above is not a management system. It is the achievement of civilized society.

The objection is partly right, and it is worth conceding exactly how. None of these systems is containment by nature. Each becomes containment by substitution. A transfer becomes containment when it replaces a claim on production without creating a new claim on membership. Therapy becomes containment when it treats structurally produced dislocation as an individual adjustment problem. A credential becomes containment when it extends preparation after the destination has narrowed. Civic ritual becomes containment when people are invited into politics after the decisive acts have migrated beyond its reach.

The same institution can be care in one historical setting and containment in another. The difference is whether it restores membership or manages the absence of it.

Provision without membership is management.

Containment Is Not Stability

The error is to mistake containment for equilibrium.

A society can suppress the political consequences of a legitimacy failure for a long time. Transfers keep people fed. Platforms keep people engaged. Credentials keep people occupied. Therapeutic frameworks keep private distress from becoming shared grievance. Civic rituals maintain the grammar of participation. Infrastructure resistance gives anger a local, legal, manageable form.

But suppression is not resolution.

The condition that containment manages accumulates rather than dissolves. As more workers find that AI can do more of what once gave them leverage, as more young people find that the junior layer has thinned before any alternative pathway was built, as more professionals find that their judgment has become supervisory rather than generative, the population whose claims exceed their recognized necessity grows.

The gap between what those people were told to expect, from education, from work, from participation in a productive society, and what the system is now offering becomes harder to paper over with any combination of transfers, platforms, and therapeutic reassurance.

The political science literature on economic anxiety confirms the direction. Autor, Dorn, Hanson, and Majlesi find that trade-induced manufacturing displacement increased political polarization, with stronger effects in more exposed congressional districts. [4] Case and Deaton document rising mortality among working-class Americans without college degrees as manufacturing employment collapsed and the social structure that organized those communities deteriorated. [5]

These are not studies of AI.

They are studies of what happens when the economy stops explaining people.

AI extends that condition more broadly, more unevenly, and faster. Containment does not resolve the surplus human problem. It defers it while the conditions compound.

The political system that manages this condition successfully for a decade may find that it has created larger and more volatile problems in the next one. The anger of people the system has learned to operate around does not disappear because it is absorbed by the attention economy or channeled into zoning disputes. It accumulates grievance, loses specific form, and waits for a political framework capable of giving it direction.

That framework will not necessarily be constructive.

The surplus human is not someone without needs.

She is someone whose claims remain after the system has learned to operate around her. And a civilization that responds to that condition only with management, without explanation, without a grammar that makes her presence intelligible inside the world being built, is not stable.

It is delayed.

More people each year stand at the edge of a productive system that functions without fully needing them, holding claims the system has not yet found a language to honor or a mechanism to redeem.

Containment keeps them standing there.

It does not answer the question of why they belong.

Containment is what a civilization builds when it still needs obedience from people it no longer knows how to need.

Notes

[1] On the Prince William Digital Gateway: the Prince William County Board of Supervisors approved the rezonings in December 2023 after a 27-hour public hearing; Circuit Court Judge Kimberly Irving voided the rezonings in August 2025 for defective public notice; the Virginia Court of Appeals upheld that ruling on March 31, 2026; and the board voted on April 14, 2026 not to appeal further. Prince William Times, “Prince William County drops legal fight for PW Digital Gateway,” April 2026, https://www.princewilliamtimes.com/localnews/breaking-prince-william-county-drops-legal-fight-for-pw-digital-gateway/article_e7152a10-cbf9-4a7a-80ef-55fa9e13770b.html. On collapsing public support: a Washington Post-Schar School poll published in April 2026 found 35 percent of Virginia voters comfortable with a new data center in their community, down from 69 percent in 2023. Data Center Watch, a tracker run by 10a Labs, recorded 48 data center projects blocked or delayed across the United States in 2025, representing roughly $156 billion in planned development. On Tucson: AZ Luminaria, “Tucson City Council rejects Project Blue data center amid intense community pressure,” August 6, 2025, https://azluminaria.org/2025/08/06/tucson-city-council-rejects-project-blue-amid-intense-community-pressure/. On federal legislation: Clean Cloud Act of 2025, S. 1475, 119th Congress, introduced April 10, 2025 by Senators Sheldon Whitehouse and John Fetterman, https://www.congress.gov/bill/119th-congress/senate-bill/1475; AI-Related Job Impacts Clarity Act, S. 3108, 119th Congress, introduced November 2025 by Senators Josh Hawley and Mark Warner, https://www.congress.gov/bill/119th-congress/senate-bill/3108.

[2] The classic account of Luddite machine-breaking as skilled workers’ rational response to capital-sponsored deskilling and displacement is E. P. Thompson, The Making of the English Working Class (1963), Part Two. Thompson establishes that the frame workers were targeting was not machinery in general but the specific equipment being used to undercut their trades. On surplus labor management in the Global South: Mike Davis, Planet of Slums (Verso, 2006) documents how cities across Africa, Asia, and Latin America absorbed populations that formal labor markets could not employ, through informal economies, NGO provision, and the containment logic of urban peripheries. The ILO’s World Employment and Social Outlook reports consistently find that more than 60 percent of employment in low-income countries is informal, reflecting a long-running condition in which large populations remain economically active while lacking stable incorporation into formal employment, taxation, and social protection.

[3] For the frontier-actor gesture: Sam Altman, “Moore’s Law for Everything,” March 2021, proposing an American Equity Fund financed by taxes on large companies and land, https://moores.samaltman.com; Cullen O’Keefe et al., “The Windfall Clause: Distributing the Benefits of AI for the Common Good,” Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (2020). The specific mechanisms in the broader family of proposals vary, from social wealth funds to robot taxes to citizen dividends, but their shared premise is that AI-generated surplus cannot be narrated only as private corporate property when the underlying capability depends on public science, collective data, shared infrastructure, and inherited civilization.

[4] David Autor, David Dorn, Gordon Hanson, and Kaveh Majlesi, “Importing Political Polarization? The Electoral Consequences of Rising Trade Exposure,” American Economic Review 110, no. 10 (2020): 3139-3183.

[5] Anne Case and Angus Deaton, Deaths of Despair and the Future of Capitalism (Princeton University Press, 2020).

Update — August 6, 2026: This essay received an honorarium in the Boyd Institute’s Debt and Deficits Essay Contest. In its announcement, Boyd highlighted the essay’s identification of “a real risk, tax receipts actually falling as output grows” and its “well-fleshed-out solution.” [Read the announcement here.]

The United States has a spending problem and a revenue problem both. It also has a third problem, underneath those two and priced by almost no one: a seeing problem.

Every serious fiscal plan in circulation is an argument about two levers. Spend less: reform entitlements, cap discretionary growth, means-test, raise the retirement age. Take more: lift rates, broaden the base, close loopholes, tax wealth. The debate is old and the positions are worn smooth. What almost no plan examines is the thing both levers depend on, the quiet assumption underneath the whole apparatus: that the state can still see income clearly enough to tax it.

Fiscal capacity begins there, with orientation, not with rates. Every act of taxation is downstream of an act of sight: you cannot claim what you cannot see, and you cannot see more clearly by looking harder at the wrong place. A state that loses track of where value forms does not just collect less. It loses the ability to act on the economy at all.

America escapes the fiscal trap by doing what it did once before: attaching the tax state to the dominant measurement layer of the economy. In the twentieth century, that layer was payroll. In the machine economy, it is compute. The policy is Federal Compute Withholding: a low, provider-side levy on large-scale AI training and inference, collected where machine output is already metered, and used dollar-for-dollar to reduce payroll taxes on human labor.

That is not a side payment to the existing tax system. It is a tax-base transition. If the machine economy is going to replace part of the wage economy, then the fiscal state has to move part of its sensor from wages to compute before the base learns to disappear.

For a century it could see. The reason was the wage.

Modern fiscal capacity did not grow out of a smarter tax code. It grew out of an accident of measurement. When income arrives as a wage, it arrives already visible. The employer computes it, reports it, and withholds against it before the worker ever touches the money. Payroll withholding turned the private economy into a machine that reports itself. The state did not have to chase most of its revenue. The revenue arrived pre-counted, at the source.

This is not a minor feature of the system. It is most of the system. In the United States, individual income and payroll taxes together supply roughly eighty-five percent of federal revenue [1]. Across the OECD, personal income tax and social contributions make up about half of all tax revenue [2]. Payroll was not just income. It was the reporting layer of industrial capitalism. The fiscal state is, structurally, a wage-reading instrument. Its power to fund itself is inseparable from labor income being the most legible thing in the economy.

This is legibility in James Scott’s sense: a state acts only on what it has first made readable. Scott’s modernizers made forests, land, and people legible in order to govern them. The fiscal state made the wage legible in order to tax it. The principle is old. Applying it to the tax base, and watching that base turn unreadable, is the new part.

That legibility is now migrating out from under it.

The base is leaving the sensor

The standard worry about automation is that fewer workers will pay less tax. The real problem is stranger and harder. As intelligent systems compress labor, value does not simply shrink. It moves. It accrues to capital, to equity, to the owners of compute, and increasingly to autonomous systems whose output has no wage line at all.

Capital income was always less legible than wages. It is harder to withhold, easier to shift across borders, and timed to the taxpayer’s convenience through the choice of when to realize a gain. Agentic output is worse still, because there is no employer in the loop to report it. A human employee produces a W-2. A model produces output. An agent completes a task. A firm substitutes software for staff and the wage line disappears, but the value does not show up in an equally taxable form.

So the coming fiscal shock is not only that the tax base gets smaller. It is that value increasingly takes forms the fiscal apparatus was never built to see. You cannot set a rate on a base you cannot observe. The trap tightens from the revenue side in a way no rate change can reach, because the instrument that made rates enforceable was the wage, and the wage is thinning.

This is why the deficit debate keeps feeling both urgent and strangely stale. It is trying to solve a sensor failure with rate changes.

The official numbers largely assume this away. The Congressional Budget Office’s latest ten-year baseline still expects individual income and payroll taxes to hold at between 82 and 84 percent of federal receipts straight through 2036 [6], and under ordinary macroeconomic assumptions that is a defensible call. But it means the scorekeeper’s central case is that the wage sensor keeps working exactly as it has for a century. None of this says the decay is already visible in the receipts. It is not, and the early-2020s labor data is too contaminated to read cleanly either way. The narrower claim is the one that bites: every deficit projection, every debt-to-GDP curve, every fiscal cliff we argue about inherits the assumption that it will keep working. If it is wrong, it is not wrong at the margin. It is wrong about the one input the entire forecast is built on. The baseline cannot price the possibility that the base itself becomes unreadable, because its models have no way to represent a measurement layer that decays. That is not reassurance. That is the risk, sitting unpriced inside the official numbers.

Why the usual menu misses

None of the standard answers is foolish. Each is just aimed one layer too high.

Growth is the most seductive, and it is not wrong. Growing the denominator faster than the debt reprices is the one durable exit, and every serious account of this trap, this sprint’s included, ends there. But growth answers the denominator, and revenue is a numerator problem. A productivity boom does not rescue the sensor, because it accrues where the sensor is weakest: output can rise while the taxable-at-source share of it falls. The official baseline concedes as much from the other side, quietly assuming a productivity lift from AI while still treating the wage sensor as if it will read the machine economy exactly as it read the wage one. You can grow the economy and blind the tax state in the same decade.

This is also why the optimistic AI-growth story does not remove the problem. Suppose artificial intelligence helps push the United States toward a forty-trillion-dollar economy over the next generation, or even higher if the most aggressive productivity forecasts prove right. That would ease the denominator of the debt ratio, but it would not automatically repair the numerator. If the new output arrives as wages, the old fiscal state survives. If it arrives as cloud margin, model rent, equity appreciation, automated enterprise surplus, and offshore-booked software profit, America can become richer while the tax state becomes blinder. Growth solves the denominator. It does not automatically solve the sensor.

Taxing the rich runs into arithmetic and administration. Even aggressive high-income proposals raise limited revenue relative to the long-run fiscal gap [3], and they press on the visible, realized, domestic base, which is exactly the base most able to defer, reclassify, litigate, borrow against, or leave. It is leaning harder on the part of the economy most trained in becoming hard to see.

Entitlement reform is real and, on the spending side, unavoidable. But it does nothing for revenue legibility. It slows the outflow. It does not restore the state’s sight.

Debt brakes and fiscal commissions impose discipline on a base that is going dark. A rule that forces balance is only ever as good as the revenue it can actually see. And interest compounds all of it: net interest on the public debt has now crossed the trillion-dollar line for the first time and overtaken defense as a federal spending category [4]. Rising interest on a shrinking legible base is simply the trap closing on schedule.

Behind all five is the same unexamined premise: a working sensor. Remove that premise and the entire debate is being conducted about the wrong variable.

There is a sharper point hiding in this, and it should unsettle anyone who trusts markets. A tax code that reads wages precisely but cannot read compute is not neutral, and it is not pro-market. It is a standing subsidy for whichever form of value becomes hardest to see. Left alone, the system will quietly privilege machine-mediated output over human-mediated work, not by anyone’s design but by blindness. That is not the market deciding. It is a thumb on the scale in favor of whatever hides best from the tax collector.

The state has not lost money. It has lost position.

Put the diagnosis plainly. The fiscal trap is downstream of a measurement failure.

For a century, the private economy reported itself through payroll. The employer was the state’s unofficial fiscal interface, converting millions of private transactions into one legible public stream before the individual was ever paid.

That arrangement was historically contingent. It depended on mass employment, large firms, standardized wages, and income arriving in forms that could be captured before the individual received it. The modern tax state did not merely tax labor. It was built around labor as the reporting layer of capitalism. If that reporting layer thins, the state does not merely need a new rate. It needs a new interface.

Artificial intelligence threatens that arrangement not by eliminating work overnight but by changing where the marginal dollar appears. It may not arrive as a salary at all, but as cloud margin, software rent, equity, or a task an agent completes without ever entering payroll.

The deficit is the symptom the models can see. The blindness is the disease they cannot.

Which reframes the escape. The task is not to press harder on what remains visible. It is to rebuild the instrument that decides what is visible at all.

Where value becomes visible again

There is one honest answer to where post-labor value can still be seen, and it is physical.

Compute.

Compute is the rarest thing in the new economy: value that is already metered. Every training run and every inference call is counted, because counting is how the service is billed. The meter is not a policy anyone has to invent. It is already running, humming in the data center, because the product cannot be sold without it. Compute is not just infrastructure. It is the reporting layer of machine production.

And unlike capital income, compute is anchored. A data center is bound to a grid interconnection, a water supply, a parcel of land, and a latency envelope. It behaves less like a bank account and more like a mine, a railroad, or a port: expensive to build, slow to move, locally permitted, publicly visible, and dependent on physical infrastructure. This is the same lesson the monetary essays in this sprint reach from the other side. Money is a claim that ultimately rests on a physical constraint stack. A base can be made to feel weightless for a while, but the constraint underneath does not repeal, it only waits. Compute is where that stack surfaces in the AI economy and becomes countable. The physical anchoring is not a detail. It is what makes the value both legible and harder to route away.

US data centers already consume a material and rising share of national electricity. The Department of Energy’s 2024 data center report, produced by Lawrence Berkeley National Laboratory, estimated that data centers consumed about 4.4 percent of US electricity in 2023, with demand projected to rise sharply by 2028 [5]. That is not yet a full tax base. But it is a map of where the new taxable surface is forming.

The successor to payroll withholding is therefore not exotic. It is the same architecture applied one layer higher.

The employer withheld against the wage. The compute provider should withhold against machine output.

Federal Compute Withholding

The policy should be called what it is: Federal Compute Withholding, the successor to payroll withholding for the machine economy.

Not a robot tax. Not a wealth tax. Not a punitive AI tax. A withholding regime.

And before it is a revenue measure, it is an institutional move. Payroll withholding made wages visible decades before anyone fought over the rate. Federal Compute Withholding does the same for machine output. It builds the eye first, while the base is still concentrated, physical, and counted, so the country is not forced into a panic tax after the base has already fragmented.

A robot tax fails because nobody can define the robot. A wealth tax chases a base engineered to be mobile, contested, and litigated. A general corporate tax still waits for profit to be declared after the value has passed through accounting, geography, debt, transfer pricing, and timing.

Federal Compute Withholding would collect at the layer where the value is already counted.

The basic design is simple.

First, the collection point should be the provider, not the user. Hyperscale cloud firms, AI infrastructure providers, and large model-serving platforms already meter training and inference for billing. They know how many accelerator-hours are sold, how much inference is served, which customers are using capacity, and where the service is being delivered. They are the closest functional equivalent to the employer in the payroll system.

Second, the base should begin with high-scale AI compute services above a threshold. The regime should not touch a graduate student fine-tuning a model, a small startup buying modest cloud credits, or a university lab. It should apply to providers selling or internally deploying large-scale training and inference capacity above a defined annual threshold. Start where the market is concentrated and legible.

Third, the measurement should use a hybrid safe-harbor system: AI compute service revenue where available, accelerator-hours where revenue is bundled, and energy-adjusted capacity where internal deployment makes revenue hard to observe. The perfect unit does not exist, but the payroll tax did not require perfect measurement of human value either. It required a durable reporting point. Compute has one.

A few concrete cases fix the boundary. A frontier model sold through a hyperscaler’s cloud, an API call billed through Azure or Bedrock, is covered as AI compute service revenue, metered the way the provider already bills it. A hyperscaler serving its own model internally, where no arm’s-length price exists, is covered through the accelerator-hour safe harbor. A startup below the threshold, a university lab, an ordinary SaaS product with an incidental AI feature, and consumer or gaming hardware fall outside the regime entirely. National-security and sovereign deployments sit under an explicit carve-out, so the exemption is a decision the state makes rather than a hole the base falls through.

Fourth, the rate should begin low and adjust only when the wage sensor weakens. The purpose is not to punish AI adoption. A starting levy in the low single digits on covered AI compute services would be enough to build the reporting infrastructure without strangling the sector. The rate would rise only on a defined trigger, a rolling multi-year index of labor income as a share of national output, so the levy tracks the measured erosion of the wage base rather than anyone’s forecast of it. If the wage sensor holds, the levy stays near zero. The tax activates as replacement, not as panic.

Fifth, sourcing should follow the market served, not merely the legal address of the invoice. If a model serves US users, substitutes for US labor, or is deployed by a US business into the US market, routing the billing entity through Ireland or Singapore should not erase the meter. This is the lesson the tax state learned too late from digital advertising, intellectual property, and platform profits. Build source rules before the base learns to disappear.

Sixth, the revenue should be used to lighten the tax on work, not to fund a new program. The cleanest version routes compute receipts directly against the payroll tax: for every dollar collected from covered AI compute, one dollar reduces payroll taxes on human labor, as close to dollar for dollar as the numbers allow. This is not a new spending stream dressed up as a trust fund. It is a transfer of the tax base itself, off the decaying wage sensor and onto the compute sensor, at something near revenue neutrality.

It also inverts the politics. A compute levy framed as punishment invites a fight. A compute levy that visibly cuts the payroll tax on every working American is a different proposition, and a far more durable one. The first rule of fiscal adaptation should be simple: do not tax the worker harder because the machine became harder to see. The point is not to grow the state. It is to stop taxing the thing that is disappearing and start reading the thing that is not.

This is the missing institutional move. Do not wait for AI value to become profit, compensation, or capital gains. Collect a small share when it passes through the only layer where it is still counted out loud.

How Congress could build it

Phase one is reporting first, with the tax near zero. Large AI compute providers would report covered training, inference, accelerator-hours, energy-adjusted capacity, internal deployment, and the market served. This is the payroll-reporting moment: before the country argues over how much to collect, it establishes where the new base is.

Phase two is low withholding above a high threshold. A one to three percent levy begins only for providers selling or internally deploying high-scale AI compute, with clear exemptions for research, universities, small startups, ordinary SaaS features, consumer hardware, and national-security deployments.

Phase three is the payroll offset. Receipts automatically reduce employer-side payroll taxes, worker payroll taxes, or both. The political promise should be visible on the paystub: machine-output receipts buy down the tax on human labor. The rule is simple to say and hard to argue with. Do not tax the worker harder because the machine became harder to see.

Phase four is triggered scaling. If labor income falls as a share of national output over a rolling multi-year period, the compute rate rises within a statutory band. If the wage sensor holds, the rate stays low. The mechanism activates as replacement, not as panic.

How much could it matter?

No single reform closes the fiscal gap. But every durable fiscal regime begins by finding the base it can actually see. Payroll withholding did not matter because its first year was large. It mattered because the state attached itself to the measurement layer that would dominate the next century. Compute withholding is the equivalent institutional move for the machine economy.

At the early sensor stage, suppose covered AI compute services reach one trillion dollars in annual value in the 2030s. A three percent withholding rate raises about thirty billion dollars a year, enough to fund a visible payroll-tax cut without raising a single rate on a worker. That is the floor, not the point.

At mature machine-economy scale, the base changes. If covered AI compute and machine-output services reach three trillion dollars in the 2030s, a five percent rate raises on the order of one hundred and fifty billion dollars a year, and closer to two hundred billion at the aggressive end of the statutory band. That is not a rounding error dressed up as a reform. One hundred and fifty billion is roughly a full percentage point off the combined payroll tax, a cut every worker would see on a paystub. It is on the order of a twelfth of the federal deficit, and a real bite out of net interest, the fastest-growing line in the budget, drawn from a base that did not exist twenty years ago. And if AI value scales the way its own boosters project, the base and the levy scale with it inside the statutory band, without anyone having to invent a new tax.

And if AI really is one of the engines of a forty-trillion-dollar American economy, the stakes rise again. Such growth does not rescue the tax state if the marginal trillion is formed outside payroll and booked through compute-mediated rents. In that world, failing to meter the compute layer would be one of the great fiscal mistakes in American history: a larger economy passing through a smaller fiscal aperture.

The important number is not the first-year levy. It is the claim on the next dominant tax base before that base becomes as mobile, litigated, and offshore as digital profits became. The fiscal mistake America should not repeat is waiting until a new economy has matured, fragmented, and lawyered itself away from the tax system before trying to see it.

The income tax could rely on payroll because payroll was already institutionalized when the modern state scaled. The digital economy was allowed to mature first, and only afterward did the state grasp how completely profit could be shifted and booked away from where the value was made. Ireland is the monument to that delay: a handful of American firms now route global profits through Irish subsidiaries, and a large share of the state’s corporate-tax revenue rests on the booking decisions of just a few of them. Compute is still early enough to be instrumented before it learns the same trick.

The point is not that compute is already as large as payroll. It is that payroll was once just a mechanism too. The fiscal state became powerful when it attached itself to the dominant measurement layer of the economy. If the dominant measurement layer moves, fiscal capacity has to move with it.

The objections that actually bite

Compute is mobile too.

Less than capital. An income stream can be booked in Dublin from a laptop. A training cluster cannot serve a latency-sensitive American market from a jurisdiction with no power, no grid connection, no land, and no proximity. Inference can move across borders, but not frictionlessly and not without performance, regulatory, security, and energy constraints. The anchor is the whole point. It is a stronger anchor than any the income tax has left. And where inference does move, the meter follows the market rather than the invoice. A firm can train in Virginia and route its API calls through a cheaper cluster abroad, but if the model serves US users the receipt is sourced to the US regardless of which data center answers the request, fixed by provider attestation of the served market and backstopped by the physical footprint of training. The routing is visible to whichever provider bills for it. The state’s task is only to require that the provider report it, the way an employer reports a wage.

This taxes the frontier and slows the thing driving growth.

A genuine tension. The design has to respect it: low initial rates, high thresholds, exemptions for research and small operators, and automatic adjustment tied to actual payroll erosion rather than ideological hostility to AI. But the counterfactual is not a world with no tax. It is a revenue collapse that arrives regardless and forces cruder instruments under worse conditions. Better a small meter built early than a panic tax built late.

The incidence falls on users, not owners.

Partly, yes. So did payroll. The first question a fiscal instrument has to answer is not who ultimately bears the cost in a textbook model. It is whether the state can see the value at all. Visibility comes before incidence, because without visibility there is nothing to allocate. Distribution can be corrected downstream. Blindness cannot.

This is a gross-receipts tax, and economists are right to dislike those.

Fair, and worth meeting head on. A levy on compute turnover can cascade the way gross-receipts taxes do, and charging revenue rather than margin sits heavier on low-margin, compute-heavy uses than a profits tax would. The design can blunt this by crediting the compute a provider buys in against the compute it sells, taxing the value added at each layer instead of stacking the charge layer on layer, which is how every value-added tax in the OECD already avoids this exact problem. But the deeper answer is that the honest alternative is not a clean profits tax that reaches this value. It is a profits tax that keeps missing it, because the value is shifted, deferred, and booked elsewhere before any profit is declared. A blunt instrument aimed at a base you can see beats an elegant one aimed at a base you cannot. Payroll was blunt too, and it funded the modern state anyway.

Measuring value at the compute layer is hard.

It is the one place it is least hard. Raw compute is an imperfect proxy for value, but it is measured more reliably than most of the capital income the tax state currently tries to chase. Tokens, accelerator-hours, energy draw, utilization, and model-serving revenue are not metaphors. They are operational quantities. The state does not need metaphysical precision. It needs a durable withholding point.

If the Wage Sensor Holds

The bet is straightforward: wage income is losing, or may soon lose, its place as the base the state can most easily see. If labor income stabilizes, payroll receipts hold through the 2030s, AI value flows back into wages faster than it leaves them, and inference decentralizes so completely that no metering chokepoint remains, then Federal Compute Withholding is not the next fiscal layer. The old one survived.

But a temporary burst of payroll revenue during a data-center construction boom would not prove that. It would be the old sensor at its brightest moment, not evidence that it can read the machine economy after construction gives way to operation. Nor would it refute the proposal to say compute withholding cannot replace the whole tax system. Payroll withholding did not abolish every other tax. It made the modern revenue state administratively possible.

That is the standard here. Not omnipotence. Institutional succession. If payroll was the fiscal interface of industrial capitalism, compute is the most plausible fiscal interface of machine production.

The escape

The escape from the fiscal trap is usually pictured as a hard choice finally made: the grand bargain, the ceiling held, the entitlement cut nobody wanted, the tax increase nobody wanted to vote for. Some version of those choices may still be necessary. But a decision about the rate is worthless if the state can no longer see the base. This is why Federal Compute Withholding belongs at the center of the fiscal conversation, not at its edge: it is the instrument that makes every later bargain possible in a machine economy.

America’s fiscal problem is therefore not only budgetary. It is institutional. The tax state was built for an economy where value passed through wages before it passed into consumption, savings, or capital. Artificial intelligence is building an economy where more value passes through compute before it ever becomes payroll at all.

The old fiscal state read wages. The new one has to read compute.

That does not mean strangling AI. It means recognizing the obvious fiscal fact hiding inside the technical one: the machine economy already meters itself. It counts its training runs, its inference calls, its accelerator-hours, its energy draw, its utilization, and its capacity. The state does not need to invent a new eye. It needs to stand where the economy is already looking.

This is the choice underneath the arithmetic. A small meter built early, while the base is still legible, is renewal. A crude tax improvised late, under fiscal duress, is managed decline. The base is moving either way. The only open question is whether the state reorients in time to see where it went, and whether it uses that new sight to reduce the tax burden on human labor before the wage base thins.

The United States has a spending problem and a revenue problem both. Underneath them is the one almost no plan prices: a seeing problem.

The escape begins by building the meter before the base disappears.

Notes

[1] Congressional Budget Office, Revenues in Fiscal Year 2024: An Infographic. CBO reports $2.426 trillion in individual income taxes and $1.709 trillion in payroll taxes out of roughly $4.9 trillion in total federal revenue for FY2024 (about 84 percent combined). FY2025 figures confirm the pattern: about $5.2 trillion in total receipts, a $1.8 trillion deficit (5.9 percent of GDP), with individual income and payroll taxes still supplying the large majority of revenue. The combined share dips slightly in FY2025 only because tariff receipts surged. CBO’s forward baseline holds it at 82 to 84 percent through 2036 (see note 6).

[2] OECD, Revenue Statistics 2025. Personal income tax and social security contributions together make up roughly half of total OECD tax revenue.

[3] Brookings Institution, “Can Taxes Alone Fix Long-Term Deficits?” (2026), finds that even an aggressive combination of high-income and corporate measures (taxing capital gains as ordinary income, a 77 percent estate-tax rate, a 35 percent corporate rate, and applying Social Security tax to all earnings) falls well short of the roughly 4 to 5 percent of GDP adjustment needed to stabilize the debt, and that closing the gap on the revenue side requires broad-based taxes reaching well beyond high earners. Manhattan Institute (Jessica Riedl), “The Limits of Taxing the Rich,” reaches the same conclusion from the revenue-maximizing side: taxing the wealthy at revenue-maximizing rates yields at most about 2 percent of GDP, and roughly 1 to 2 percent after accounting for macroeconomic responses.

[4] Congressional Budget Office, Monthly Budget Review: Summary for Fiscal Year 2025. CBO reports that net interest on the public debt surpassed $1 trillion for the first time in FY2025; independent analyses note it now exceeds national defense spending by roughly $150 billion.

[5] Department of Energy / Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report. The report estimates data centers consumed about 4.4 percent of US electricity in 2023 and projects data center demand could roughly double or triple by 2028.

[6] Congressional Budget Office, The Budget and Economic Outlook: 2026 to 2036 (February 2026). CBO projects that individual income and payroll taxes remain between roughly 82 and 84 percent of federal receipts through FY2036. As summarized by the Bipartisan Policy Center, “The Fiscal Outlook in CBO’s Latest 10-Year Baseline” (February 11, 2026).

The Democratic establishment keeps treating democratic socialism as an ideology problem.

That is probably too comforting.

Ideologies do not become politically useful because people suddenly read the right books. They become useful when they explain a pain that existing institutions can no longer translate. The socialist language now returning to American politics is not only a theory of ownership. It is a theory of betrayal.

The wage was supposed to make people legible to adulthood.

Work hard. Get credentialed. Enter the labor market. Rent first, own later. Form a household. Become reliable. Become respectable. Become someone whose future could be narrated without irony.

That sequence no longer feels stable to a large part of the Democratic coalition.

Housing costs absorb the wage. Healthcare disciplines the wage. Student debt pre-spends the wage. Childcare breaks the wage. Entry-level work demands experience that entry-level workers cannot yet have. The credential still matters, but it no longer reliably opens the gate. The result is not only economic pressure. It is a collapse of explanation.

People can endure hardship when the system can tell them what the hardship is for.

That is where democratic socialism enters.

It does not arrive as a finished governing program, a national majority, or a return to the twentieth century. It arrives as the political form of people who no longer believe that market participation, credential compliance, and institutional patience will deliver adult life.

The rent is too high.

But that is only the surface form.

The deeper problem is that the old distributive promise has failed before a new system of allocation has been named.

Not a Socialist Moment, a Wage-Legitimacy Moment

The recent socialist surge is easy to misread because American politics still processes socialism as an ideological label.

That is the wrong first unit of analysis.

The visible events matter. Zohran Mamdani became mayor of New York City on January 1, 2026, after running an affordability-centered campaign in the most symbolically important city in American capitalism. His administration then secured a rent freeze on one- and two-year leases for rent-stabilized apartments, covering roughly one million units. [1]

A few days later, Melat Kiros, a 29-year-old democratic socialist, defeated Diana DeGette, a 15-term Democratic incumbent, in the Democratic primary for Colorado’s Denver-based 1st Congressional District. She is favored to win the seat outright this fall. The incumbent was not a conservative Democrat. She was a long-serving progressive. That made the result more important, not less. The upset suggested that the cleavage is no longer simply left versus center. It is old institutional progressivism versus a younger politics of rent, debt, Gaza, labor, and anti-corporate legitimacy. [2]

These are not isolated campus signals. They are institutional signals.

Democratic socialism has learned where the Democratic Party is structurally vulnerable: safe seats, low-turnout primaries, urban rents, generational frustration, tenant politics, union energy, and a party establishment whose language of competence no longer explains why life feels harder after doing what one was told to do.

The first mistake is to ask whether America is becoming socialist.

It is not, in any simple national sense.

The better question is why socialist language has become useful inside the Democratic coalition precisely where the old liberal bargain should have been strongest: educated cities, professional labor markets, dense institutions, public-sector unions, universities, nonprofits, media ecosystems, and service economies.

The answer is not that these places are poor in the old sense but that they are expensive in a new sense: places where the wage still exists, but adulthood feels rented from someone else.

The Party of Renters

The new democratic socialist subject is not the factory worker in the old industrial imagination.

It is the renter with a degree.

The graduate student with no path into secure formation, the nurse or the adjunct close enough to elite institutions to see the gates, but not secure enough to inherit the life those gates once promised.

This is why the movement can look culturally strange to older observers: educated but angry. Institutional but anti-establishment. Morally intense but materially focused. Online, but increasingly capable of door-knocking, canvassing, primary challenges, and local governance.

The demand is not merely for more money but for a different explanation of social membership.

The older Democratic Party told people that institutions were slow but legitimate, markets were unequal but productive, credentials were costly but worthwhile, and moderation was the price of governability. Democratic socialism grows where that explanation loses force.

The housing market is the most obvious reason.

The Harvard Joint Center for Housing Studies reports that nearly half of renter households were cost-burdened in 2024, with more than a quarter severely burdened. Lower-income renters had only about $210 left each month after paying rent. The Federal Reserve’s 2025 household well-being report found overall financial well-being relatively stable, but specifically noted deterioration among young adults, low-income families, and Black adults. [3]

The politics follow from how rent is experienced: not like a normal price.

Rent is experienced as permission to remain.

A worker can believe in markets while buying coffee, clothes, or electronics. Believing in market legitimacy is harder when the market prices the conditions of adult life as a recurring tribute. Housing turns capitalism from an opportunity system into a residency system.

The socialist revival begins there.

Not at the factory gate, but at the lease renewal.

The Old Rent Was Housing

The socialist left has a strong theory of the old rent.

It knows how to talk about landlords. It knows how to talk about insurers. It knows how to talk about student debt, medical billing, corporate PAC money, public ownership, unions, and monopoly power. Its internal language is not shy about this. The Democratic Socialists of America describes itself as a member-driven mass organization and says working people should run the economy and civil society. Its platform calls for building an independent working-class party, taking on the capitalist class through labor and tenant organizing, and guaranteeing basic needs through democratic control. [4]

That theory is politically powerful because it gives structure to diffuse injury.

Rent is no longer a private inconvenience. It has become a governing relation, the terms on which someone is allowed to keep living where they live. Debt, healthcare, and the employer relation are following the same path. Each was once a transaction. Each is becoming a gate, a route, a thing that decides who reaches the next thing they need before politics ever sees the decision being made.

This is why the movement’s politics can travel across issues that appear separate. Rent freeze, Medicare for All, universal childcare, unionization, public transit, student debt cancellation, Gaza, and corporate money are not random left positions. They are all framed as cases where life is being governed by institutions that demand obedience while refusing responsibility.

The left knows how to call that extraction. It knows how to call it rent.

That is its strength, and its limit. Because the next rent does not always look like rent.

The New Rent Is Access

The old landlord owns the building.

The new landlord owns the route.

That is where the socialist vocabulary begins to lag behind the world it is trying to describe. The left knows how to fight the landlord, the insurer, the boss, the bank, the university, and the monopolist. It is less fluent when the decisive power is not a person charging too much, but a system deciding who appears, who is ranked, who is eligible, who is routed, who is filtered, who is priced, who is offered, and who is quietly excluded before politics can see the exclusion.

AI does not simply automate labor.

It changes where the decisions get made.

Before a person receives an opportunity, a price, a loan, a job interview, a search result, a welfare decision, a risk score, an insurance quote, a platform audience, or institutional attention, they may already have passed through systems that structure the possible outcomes. Some of those systems are public. Many are private. Some are formal. Many are embedded into interfaces so ordinary that they do not look political. In practice, this does not always begin with frontier AI. It begins with scoring, screening, ranking, and routing systems that AI is making cheaper, broader, and harder to avoid.

The socialist question has traditionally been: who gets the surplus?

The prior question is earlier: who gets seen by the system through which the surplus becomes reachable?

This is not only an AI problem. It is an institutional relocation problem. The economy continues to function, but the place where effective decisions are made moves upstream. The visible institution remains. The operational decision moves into ranking, scoring, routing, procurement, compute access, model access, standards, licenses, and platform permission.

A city may still debate housing policy.

But landlords increasingly outsource the decision itself to tenant-screening algorithms. In 2024, a federal court approved a settlement after renters in Massachusetts showed that one such algorithm, SafeRent’s scoring tool, had been giving lower scores to Black and Hispanic applicants and to anyone using a housing voucher, regardless of whether they could actually pay the rent. [5] The debate about housing policy happens after the score, not before it.

A university may still admit students. A company may still hire them. But the screening increasingly happens upstream of any human decision. Workday’s own court filings put the number of job applications rejected by its hiring software at 1.1 billion during the period covered by an ongoing federal lawsuit alleging its tools disproportionately screened out older applicants, a claim serious enough that a judge let it proceed as a nationwide class action. [5]

A welfare state may still promise benefits.

But eligibility systems, fraud models, identity verification, platform interfaces, and administrative scoring can decide who is delayed, denied, investigated, routed, or abandoned.

The old distributive state asks what people receive.

The state now forming around AI and platform infrastructure inherits responsibility for decisions it did not fully make.

That is the real crisis for democratic socialism. It wants to use the state to distribute. But AI and platform infrastructure are making the state answer for decisions it does not govern.

Distribution Theory Meets a New Kind of Power

Democratic socialism has a distribution theory.

It believes wealth is produced collectively and captured privately. It believes the state should guarantee basic needs. It believes labor and tenant power should discipline capital. It believes public ownership and democratic control can bring major economic decisions back under social authority.

Whatever one thinks of that program, it is at least aimed at a real failure. The wage no longer distributes dignity reliably. The market no longer distributes formation reliably. The credential no longer distributes admission reliably.

The problem is that visibility now precedes distribution.

If distribution is the question of who receives what, this earlier question is who is positioned to receive anything at all. It is the architecture before the transfer: the ranking before the offer, the score before the loan, the interface before the market, the standard before the law, the permission before the price.

The socialist toolkit was built for visible antagonists, which is why the difference is not academic.

A landlord can be named. A boss can be organized against. A corporation can be taxed. A utility can be regulated. A university can be pressured. A hospital system can be investigated. These remain real targets.

But this newer form of power is often stranger than ownership power.

It can be distributed across vendors, software systems, data brokers, procurement rules, cloud providers, model providers, compliance standards, insurers, payment systems, identity systems, and public agencies that do not experience themselves as sovereign actors. No single institution may decide in the old political sense. Yet the result still governs.

This is how politics becomes harder to locate.

The socialist left says the economy should be democratized.

The AI economy replies by moving the economy into systems whose political nature is not yet publicly legible.

That does not make democratic socialism irrelevant. It makes it incomplete.

The next left will need more than a theory of redistribution. It will need a theory of who gets routed where, and by what.

Who controls the systems that determine whether a person, firm, city, worker, applicant, patient, tenant, or institution becomes visible to opportunity? Underneath that one question sit several more: who audits the categories, who owns the models, who writes the standards, who controls compute access, and who decides which systems become mandatory by convenience before they become mandatory by law.

Who is liable when a public institution adopts a private system and then calls the outcome administrative?

These are not secondary technical questions. They are the new political economy.

The factory was visible. The interface is not.

The Chokepoint Strategy

This is why democratic socialism can matter before it becomes nationally popular.

It does not need to win America to change the Democratic Party. It needs to win chokepoints.

Safe-seat primaries, rent boards, union endorsements, procurement fights, and candidate pipelines are all chokepoints.

This is the movement’s institutional intelligence.

It understands that the Democratic Party is not one electorate. It is a series of gates. Some gates are expensive. Some are low-turnout. Some are culturally dense. Some are controlled by unions, activists, donor networks, local press, neighborhood organizations, and online narrative flows.

A faction does not need broad national approval to change what can be said inside those gates.

That is why the Mamdani and Kiros signals matter. Mamdani’s victory turned democratic socialism from a protest register into a governing experiment in America’s largest city. Kiros’s primary win showed that even a veteran progressive incumbent can be vulnerable if a challenger captures generational impatience, anti-corporate legitimacy, and activist organization inside a safe Democratic district. [6]

This is not yet a national majority.

It is a pressure system inside the party that governs many of the places where the next political economy will be built.

This is where AI enters: blue cities, universities, state agencies, public hospitals, procurement offices, labor unions, and local governments will be among the first institutional surfaces where algorithmic allocation becomes contested. The socialist left may arrive at that fight with a strong instinct against corporate power, but without a precise vocabulary for the systems now translating corporate power into public administration.

The likely first move will be moral.

Ban it. Tax it. Publicly own it. Unionize against it. Prohibit replacement. Demand transparency. Require human review. Create a public option.

Some of those moves will matter.

But the deeper issue is not whether machines replace workers in a clean one-for-one sense. The deeper issue is whether institutions become dependent on systems they cannot interpret, cannot reproduce, cannot refuse, and cannot govern without losing operational capacity.

That is where socialist politics and AI governance will collide.

Democratic socialism wants public power to discipline private capital.

Synthetic civilization relocates effective power into infrastructure that public institutions increasingly need in order to function.

The Wrong Layer

There is a risk that the socialist left fights the wrong layer.

The most immediate danger is not unemployment. It is the seniorization of entry-level work, the fact that the first rungs of a career now demand experience only a career could have given you. The monopoly fights that come naturally to the left may miss the more precise issue underneath them: dependency on infrastructure that cannot be easily exited, ownership questions that turn out to be routing questions, and price fights that turn out to be about permission. Even replacement, the thing everyone worries about first, may not be the durable problem. Formation might be.

A society can survive losing some jobs to automation. It has a much harder time when the systems that used to turn people into adults, that produced judgment, confidence, and standing, stop being built for that purpose at all.

The entry-level labor market already shows this ambiguity. Employer surveys and reporting point in different directions. Some firms say AI adoption supports hiring and raises demand for AI-fluent junior workers. Other data and reporting show a more difficult market for recent graduates, more automated screening, flatter hiring, higher underemployment, and entry-level jobs demanding skills once associated with senior workers. [7]

The dispute itself is informative.

The question is not simply whether AI destroys all junior work, but whether the old initiation ladder can survive when the first tasks of formation are automated, compressed, or reserved for people who already know how to perform without being formed.

That is not only a labor-market issue.

It is a political issue.

A society can redistribute income to people who have been excluded from work. Formation is much harder to redistribute. So are judgment, confidence, institutional trust, apprenticeship, adult status, and the feeling that one has been admitted into the world rather than merely maintained by it.

Democratic socialism is powerful where it names distribution failure.

It will become more powerful if it names formation failure.

But it will become much more dangerous, and much more serious, if it learns to name the power that decides who gets seen at all.

The Rent Is Still Real

There is a strong objection to this argument.

Maybe democratic socialism does not need a refined theory of algorithmic allocation yet. Maybe the old problems are still the main problems. Rent is real. Healthcare costs are real. Childcare costs are real. Corporate concentration is real. Union decline is real. Public infrastructure is real. The housing shortage is real. If people cannot afford rent, telling them that the deeper issue is upstream allocation may sound like evasion.

That objection should be taken seriously.

A politics that skips the landlord in order to theorize the interface will fail. People do not experience abstraction first. They experience the bill, the denial, the rejection, the rent increase, the commute, the job application that disappears into software, the hospital invoice, the grocery receipt, the lease renewal.

The old rent is not fake because a new rent is emerging.

Housing still disciplines life.

But the two are beginning to connect. The landlord increasingly depends on software, financing conditions, insurance pricing, tenant screening, platform visibility, zoning analytics, payment systems, and regulatory interfaces. The employer increasingly depends on automated screening, productivity systems, vendor platforms, model access, cloud infrastructure, and compliance tools. The state increasingly depends on private systems to see, sort, verify, detect, prioritize, and administer.

So the issue is not that the old socialist categories are wrong. The issue is that they are becoming downstream.

The landlord remains. But behind him, another system is beginning to decide which tenants become visible at all.

If the Pattern Breaks

None of this is worth much if it cannot fail. So here is what failure would look like.

The socialist surge would look like a local anomaly, not a shift, if candidates identified with democratic socialism keep losing outside a narrow set of symbolic cities and deep-blue districts, even after getting the national attention and organizing muscle Mamdani and Kiros just proved they can attract.

Rent-centered governance would undercut its own case if it produces visible deterioration in housing supply, maintenance, fiscal capacity, or public services, and the voters who demanded it turn against it. That would mean the movement can win office but not hold legitimacy once it has to govern.

The formation argument weakens if the AI labor transition ends up strengthening early-career hiring rather than compressing it. Firms expanding apprenticeship and junior training broadly, not in isolated cases, would say the crisis is less structural than the argument claims.

The claim about a new locus of power weakens if public institutions keep real control over the systems they adopt: auditable models, exit options, procurement discipline, technical capacity kept in house, accountability that is actually visible. The problem was never private software by itself. It was dependence without competence.

The whole argument would weaken fastest of all if democratic socialist organizations build a working vocabulary of compute, standards, procurement, and algorithmic eligibility before this argument is even out of date, and move past rent and redistribution on their own.

Nothing on that list has happened yet. The pattern still points toward rent as the language, and toward what sits underneath it.

Democratic socialism is rising because the old bargain has lost narrative authority. The wage still exists, but it no longer reliably delivers the life it was supposed to organize. Rent has become the everyday language of that failure. The socialist left knows how to name it.

But the next conflict will not stop at rent.

The coming political economy is not only about who owns the surplus. It is about who owns the systems that decide who reaches the surplus at all.

The left has a theory of who gets paid.

AI is rewriting who gets seen.

Democratic socialism is the political form taken by a generation that no longer believes the wage can deliver adulthood.

Notes

[1] New York City’s mayoral site states that Zohran Mamdani was sworn in as mayor on January 1, 2026. The mayor’s June 25, 2026 statement describes the Rent Guidelines Board vote as a freeze on one- and two-year rent-stabilized leases, and multiple reports place the affected universe at roughly one million rent-stabilized apartments.

[2] On Melat Kiros’s defeat of Diana DeGette in the Democratic primary for Colorado’s 1st Congressional District, see Axios, Colorado Newsline, The Guardian, NBC News, and the Associated Press. Coverage emphasized the shock to House Democrats, the fact that DeGette was herself a progressive incumbent, and Kiros’s backing from DSA, Justice Democrats, and Senator Bernie Sanders. The seat itself does not change hands until the winner of the November general election is sworn in; CD1 is a safely Democratic district.

[3] Harvard Joint Center for Housing Studies, America’s Rental Housing 2026, reports that the renter cost-burden rate was 49 percent in 2024, including 26 percent of renter households with severe burdens, and that lower-income renters had about $210 left after paying rent. The Federal Reserve’s report on household well-being in 2025, published in 2026, says overall financial conditions were generally stable, but financial well-being declined for young adults, low-income families, and Black adults.

[4] The DSA platform describes the organization’s goals in terms of working-class power, labor and tenant organizing, democratic control, an independent working-class party, and a future free from capitalist exploitation. Its homepage describes DSA as a member-driven mass organization committed to working people running the economy and civil society.

[5] On the SafeRent Solutions litigation, see Louis v. SafeRent Solutions, LLC (D. Mass., 1:22-cv-10800), including the July 2023 order denying SafeRent’s motion to dismiss and the November 2024 final approval of a roughly $2.28 million settlement covering Massachusetts housing voucher recipients. On the Workday litigation, see Mobley v. Workday, Inc. (N.D. Cal., 3:23-cv-00770), including the July 2024 order allowing an “agent” theory of vendor liability to proceed and the 2025 order granting preliminary collective certification of the age-discrimination claims. Workday’s own court filings put the number of applications rejected by its software at 1.1 billion during the period covered by the suit, which runs from September 2020 onward.

[6] For Mamdani’s rent-freeze victory, see the NYC Mayor’s Office statement and coverage of the Rent Guidelines Board vote. For Kiros, see Axios, Colorado Newsline, and NBC News. The point here is not that the two cases are identical, but that both show democratic socialist politics operating through institutional chokepoints rather than mass national majorities.

[7] The entry-level labor evidence is mixed. NACE’s 2026 Spring Update projected a 5.6 percent increase in hiring for new college graduates and said AI is becoming an expectation for early-career talent. Business Insider reported on Ramp and Revelio Labs research finding that heavy AI adopters increased headcount and entry-level hiring after adoption. Other reporting describes a difficult recent-graduate labor market, automated hiring friction, high underemployment, and employers demanding more advanced skills in junior roles. The author’s reading is that the uncertainty itself matters because the entry-level job is no longer only a job. It is a formation mechanism.

The letter told Anthropic that a license would no longer be required to release its Claude Mythos 5 model to the entities named in an attached annex, to those entities’ foreign-national employees, and to Anthropic’s own foreign nationals. [1] The annex held a little over a hundred names. Everyone not on it stayed where they had been since June 12, when the same department had ordered the model withdrawn from every foreign person on the planet, including the non-citizens sitting inside Anthropic’s own offices. [2]

The sentence that mattered was not the permission. It was the reservation that followed it. The Secretary kept the right to amend the approved list “at any time.”

That is not a market clearing.

It is a guest list.

The first durable moat of frontier AI may turn out not to be model quality, capital, talent, or chips, but administrative permission.

From price to license

For two years, the question of who could use a frontier model had a settled set of answers, and all of them were private.

The price page tiered you. The rate limit metered you. The partner agreement let some firms in early and kept others in the queue. The product was described as open access, but access was never flat. It was sorted beneath the interface.

The allocator was a company. The lab set the price. The incumbent set the terms. The platform decided who was ordinary, who was strategic, who was early, who was delayed, and who would never see the real frontier at all.

The June letters move the lever into a different hand.

For this class of frontier access, the state has become the allocator, and the instrument is no longer a price but a license.

Export controls have shaped the inputs to this industry for years: chips, fabrication equipment, design software, tooling, and the industrial base beneath compute. It is defensible to call those controls the first real regulation the AI industry faced. They did not govern speech. They governed the machinery that made speech at scale possible.

What changed in June was the object.

The control no longer stops at the hardware. It reaches the model itself, delivered as a hosted service, and through the model it reaches the user.

The June 12 directive was structured as an “is informed” license requirement under the Export Administration Regulations, the rules administered by the Bureau of Industry and Security, drawing on authorities written into the Export Control Reform Act of 2018. [3] Bloomberg published the text four days later. It was the first public use of those authorities against access to an AI model rather than the export of a physical thing.

The mechanism doing the work is older than the technology it now governs.

Under the regulations, releasing controlled technology to a foreign national can count as an export to that person’s home country, even when the technology and the person both sit inside the United States and neither moves an inch. The doctrine has a name from an earlier industrial era: a deemed export. [4]

It was built for a world of schematics, laboratory benches, machine tools, and controlled technical data. A foreign researcher reading a restricted blueprint could be treated, in law, as the blueprint crossing a border.

Applied to a chat interface, the doctrine becomes stranger and more powerful. A prompt typed in California by an engineer on a work visa can be treated as a shipment abroad.

That is the hinge.

Once access is an export, the government that controls exports controls access. And a government that controls access can keep a list.

Access by nationality

When a license turns on the nationality of a firm’s employees, the allocation does not stop at the border. It reaches inside the workforce.

The June 12 directive did not distinguish between a hostile state’s intelligence service and a German postdoc on Anthropic’s payroll. Both were foreign persons. Both were cut off. The company concluded it had no way to comply except to switch the models off for everyone.

The June 26 letter restored access along the same axis by which it had been denied: not to the world, but to a set of named institutions, their foreign-national employees, and Anthropic’s own foreign nationals.

Read structurally, access now sorts by citizenship, by institution, and by the bloc a citizen’s country belongs to.

The long-range version of this is cognitive blocs: cognition hardening into rival spheres the way territory once hardened into empires. That is the fifty-year picture.

The present-transition artifact is smaller, drier, and more important because it is real. It is a letter on department letterhead with an annex stapled to it.

The blocs do not announce themselves first as treaties. They appear as the difference between the names on the list and the names off it.

The alliance cost is the part stated most quietly, because it runs through the firm rather than the border.

An allied government may see its national champions named in the annex while citizens of that same allied state, employed by those same institutions or by American frontier labs, remain foreign nationals under the deemed-export rule. The ally is trusted as an institution while its citizens remain suspect as persons.

That is the new ambiguity of AI alliance politics. The same instrument affirms alliance and strains it. It says: your firm may enter, your passport may not. It makes partnership conditional on a licensing apparatus controlled elsewhere.

Allies do not experience this as openness. They experience it as dependence on decisions taken in Washington. That is not resentment. It is an accurate description of being downstream of someone else’s list. [5]

This is what it means for the routing of intelligence to relocate into a single state’s licensing apparatus. The relationship is no longer mediated only by capability, price, contract, or mutual access. It is mediated by membership.

The list is the moat

The hundred-odd names in the annex are not merely permitted. They are advantaged, and the advantage is durable in a way ordinary commercial leads are not.

Consider the unlisted firm. It cannot buy its way onto the list because the list is not for sale. It cannot out-compete its way onto the list because the gate is not a market. It can only petition, wait, comply, lobby, and hope.

During that wait, its listed rivals build products. They train staff. They redesign workflows. They test failure modes. They integrate the model into internal systems. They learn what the unlisted cannot learn because the unlisted have never had the thing in front of them.

A frontier model is not only a tool. It is an environment of discovery. Access teaches the user what access is worth.

This is why the list becomes a moat.

The listed institution does not merely get to use the model. It gets to accumulate the operational familiarity that later appears as competence, maturity, trustworthiness, and strategic importance. The unlisted institution does not merely lack a product. It lacks the chance to become the kind of institution that regulators can confidently list.

The gap widens by use. The listed pull ahead because they are listed. Then their lead becomes evidence that they belonged on the list. The unlisted fall behind because they were excluded. Then their lag becomes evidence that exclusion was prudent.

That is the allocation flywheel: each turn of access produces advantage, and the advantage justifies the next turn of access.

In the commercial version, the incumbent turned the wheel. The price page, enterprise plan, capacity limit, and preferred partnership did the sorting. The metering was private.

The license layer changes the owner of the wheel. Entry now passes through a public chokepoint, and the firm administering the annex does so not merely as a market actor, but as the government’s instrument.

This is the part competition law was not built to reach.

Antitrust was built to police private exclusion. It looks for an agreement, a conspiracy, a dominant firm abusing its position, a merger that lessens competition, a platform discriminating against rivals. It finds none of those cleanly here.

There is no cartel, because the allocation is a public act. There is no conventional abuse, because the advantage was conferred, not seized. There is no single landlord to sue, because the gate is legal, procedural, and justified by national security. The disadvantage has no villain, only an architecture.

National-security permission is becoming a competitive asset.

That is difficult for a competition authority to name, because the harm is not merely that one firm has more market power. The harm is that the market is no longer the place where the most important allocation happens. A price can be undercut; a list cannot.

The administrative frontier

On the same day Anthropic’s annex was published, its closest competitor released its newest model only to a short list of government-approved partners. [6]

That coincidence matters less as a conspiracy than as a pattern. The market did not produce one frontier with many doors. It is producing several frontiers, each with a list.

Some lists are commercial. Some are governmental. Some are partnership lists. Some are security lists. Some will be formal, published, and appealable. Others will be quiet, contractual, discretionary, or hidden inside procurement channels.

But the shape is the same. The frontier is not simply where capability advances. It is where eligibility is decided.

That distinction matters because eligibility is a different kind of power from price.

A price says: you may enter if you can pay. A license says: you may enter if you are allowed. The first discriminates by wealth. The second discriminates by membership.

That is why the licensed frontier is more politically explosive than the expensive frontier. Expensive access can still pretend to be universal, because everyone can imagine themselves as a future buyer. Licensed access makes the exclusion explicit. It says that some institutions and some persons are on the inside of the future, while others remain outside by administrative fact.

This is not a return to the old digital divide. That divide concerned access to networks, devices, broadband, software, and skills. It asked whether you could connect. The new divide concerns access to machine cognition at the frontier of state-recognized capability. It asks whether you are eligible to work with the most advanced systems. That is a different regime.

When a list is only a list

Two weeks of letters is thin evidence for a permanent structure. Here is what would show it was never one.

If this stays one company, one model, one patched vulnerability met with a single overcorrection, there is no architecture here, only an incident. If the annex grows until being listed confers no real advantage, the moat was imaginary. If access settles onto use case, safety procedure, or technical compliance instead of nationality, institutional membership, and strategic alignment, the sorting principle is not the one named here. If a court, a competition authority, or an allied government finds a way to treat national-security permission as a reachable competitive harm, the gate was not as sealed as it looks. And if unrestricted models reach parity quickly, the list governs something no longer scarce enough to ration.

Those are real possibilities. The June 26 letter described itself as a loosening, and a temporary clamp can look, in its first days, like a new architecture. The companion model, Claude Fable 5, remained restricted, though officials had begun signaling the clamp was easing: its return was described as days away, and the administration had already softened its public posture toward the company. [7] A bipartisan group of House members has already asked whether Anthropic was singled out. [8] Anthropic has disputed the basis for the action, describing the cited jailbreak as narrow and reproducible on other publicly available frontier models. [9]

So the correct claim is not that one letter has permanently reorganized the AI economy. The claim is that an institutional form has appeared. And forms, once available, get reused. A list that grows and shrinks while remaining a list is not necessarily the architecture failing. It may be the architecture learning how to operate.

The future closes by eligibility

Extending export control from the chip to the model is a coherent next step in an old logic, not an aberration. The trigger may have been a security finding rather than a pretext, and the officials involved may understand themselves not as building a new political economy but as responding to a narrow risk with an available legal instrument. That is usually how new regimes begin.

They do not arrive as theories.

They arrive as procedures.

A permission letter becomes a precedent. A temporary annex becomes a template. A compliance workaround becomes a standard. A national-security exception becomes the normal path through which the frontier is accessed.

The future will not be closed by censorship or monopoly in the old sense. It will be closed by eligibility.

That is the deeper significance of the June letters. They reveal a world in which artificial intelligence is not merely bought, subscribed to, or consumed but conferred, and what is conferred can be withheld, conditioned, expanded, revoked, and sorted by the state.

The frontier was supposed to be open country, the place past the edge of the map where the rule had not yet arrived.

The rule arrived first. It came with a door, a guard, and an annex.

Access stopped being something you buy.

It became something you are granted.

Notes

[1] Howard W. Lutnick, Secretary of Commerce, to Tom Brown, Anthropic Chief Compute Officer, June 26, 2026: the letter stated that a license would no longer be required to export, reexport, or transfer in-country the Claude Mythos 5 model to the entities in Annex A and their foreign-national employees, or to Anthropic’s own foreign nationals. The Secretary reserved the right to reevaluate and adjust the scope of the license requirement should circumstances change, and noted that the approved list could be amended “at any time.” Claude Fable 5 was not included. Semafor, June 27, 2026; The Hill, June 27, 2026.

[2] Anthropic public statement on the Commerce export-controls directive, June 12, 2026: the company said it received the directive at 5:21 p.m. Eastern and disabled access to Claude Mythos 5 and Claude Fable 5 for all customers to comply, while its other models were unaffected. Fortune, June 13, 2026.

[3] The directive was structured as an “is informed” license requirement under Part 744 of the Export Administration Regulations, administered by the Bureau of Industry and Security, invoking authorities under the Export Control Reform Act of 2018. Bloomberg published the text of the letter on June 16, 2026. “Is Access to Fable an Export?”, Harvard Law Review Blog, June 2026.

[4] On the deemed-export doctrine and its application to model access: a foreign national’s access to controlled technology can count as an export to that person’s home country even when both remain inside the United States. Andrew W. Reddie, Tech Policy Press, June 2026; “Is Access to Fable an Export?”, Harvard Law Review Blog, June 2026.

[5] Allied officials described their new dependence on access decisions taken in Washington. Semafor, June 27, 2026.

[6] On the same day, a competing lab released its latest model to a limited set of government-approved partners. Semafor, June 27, 2026; The Hill, June 27, 2026.

[7] In the days after the June 26 letter, Anthropic and administration officials signaled a near-term restoration of Fable 5, with its return described as days away. The administration also softened its public posture, the President stating he no longer regarded the company as a national security threat. Axios, June 27, 2026; CNN, June 21, 2026.

[8] A bipartisan group of House members asked the administration whether Anthropic had been singled out and raised broader questions about the policy implications. Just Security, June 2026, with contemporaneous coverage in The Washington Post.

[9] Anthropic disputed the basis for the action, describing the cited jailbreak as narrow and reproducible on other publicly available frontier models. Fortune, June 13, 2026, drawing on reporting by The Washington Post and Bloomberg.

The legitimacy crisis of AI is usually told from one side.

Labor loses necessity. Work loses its moral claim. The wage weakens as an explanation for why people deserve a share of what they help produce. That story is real. But it is only half the problem.

The other half is capital.

If labor can no longer say “I made this,” capital cannot simply answer, “I own this.”

Both claims are weakening. But capital is capturing the surplus anyway. That asymmetry is the bilateral legitimacy crisis at the center of the AI economy.

This essay is not about taxation. It is about the moral architecture of ownership itself.

The AI economy is not merely redistributing income. It is reorganizing the question of what entitles an actor to the surplus that synthetic production generates. The old answer was simple enough: capital deserves its return because it took the risk, built the system, funded the innovation, organized the enterprise, and waited for the reward.

That answer is not false.

It is becoming incomplete.

Capital once had a story. It gathered resources before demand existed. It absorbed failure before profit appeared. It built factories, railroads, firms, laboratories, supply chains, platforms, and markets that no isolated worker could assemble alone. It turned scattered possibility into organized production.

That was the moral force of capital’s claim.

But AI changes the terrain. The surplus now being captured by capital rests on foundations that capital did not create alone. It rests on public science, public law, public infrastructure, collective data, state-backed markets, open protocols, accumulated language, social trust, and the inherited knowledge of civilization itself.

Capital may own the final system.

It did not build the whole world that made the system valuable.

A building can be privately owned without meaning that the owner produced the city around it. A landlord can hold title to a tower without having built the roads, laws, utilities, public safety, labor markets, institutions, and civic order that make the tower rentable.

AI capital increasingly stands in that position.

It holds title to systems whose value depends on a civilization it did not privately produce.

The Old Justification

Capital’s moral claim in a market economy rests on several pillars.

The first is risk. The investor bears the possibility of loss. The entrepreneur bets on an uncertain future. The firm deploys resources before it knows whether returns will follow. This gamble is supposed to justify the reward. If the bet fails, capital suffers. If it succeeds, capital earns the premium.

The second is coordination. Capital does not only provide money. It organizes people, assembles inputs, structures incentives, directs effort, builds institutions, and converts resources into functioning enterprises. The entrepreneur who builds a company is not merely a passive owner. She is an organizer of productive activity that would not otherwise exist in that form.

The third is patience. Capital accepts deferred returns. It tolerates periods when investment is underwater. It waits through uncertainty, delay, failure, and reinvestment. This willingness to wait is supposed to justify the claim on long-run surplus.

The fourth is innovation. Capital funds experiments before their value is proven. It finances new tools, new firms, new processes, new products, and new markets. It creates the space in which invention can become infrastructure.

Together, these justifications form the moral core of the capitalist claim: capital deserves its return because it carried burdens that others would not or could not carry.

This argument is not absurd. It has real content.

The history of industrial development is partly a history of capital absorbing risk that no individual worker, household, guild, town, or public agency was positioned to take. The venture investor who finances a company that might fail is doing something genuinely different from a passive creditor collecting interest on a safe bond. The founder who spends years building an enterprise is doing something genuinely different from someone who merely owns an inherited asset.

Capital is not only extraction.

At its strongest, capital is organized risk.

That is why the AI case matters. The argument for capital does not fail because risk, coordination, patience, and innovation are fake. It fails because the surplus now being captured increasingly exceeds what those justifications can explain.

The old story justified a return to productive risk.

The AI economy is producing returns to positional control.

Those are not the same thing.

The Steelman and Why It Fails

The strongest defense of concentrated AI capital is this: public research created inputs, but private capital created the output.

The transformer paper alone did not build a frontier language model. Public science alone did not produce a consumer interface used by hundreds of millions of people. Open research alone did not assemble data centers, secure chips, hire engineers, build inference systems, negotiate cloud deals, manage latency, create developer platforms, run safety evaluations, integrate payments, satisfy enterprise buyers, and turn abstract capability into something people and firms could actually use.

Integration is not a trivial step.

The gap between a publicly available research idea and a deployed system is real. It takes capital, risk, discipline, organizational will, technical execution, and commercial pressure. It is not obvious that governments or universities would have moved at the same speed or scale. Market incentives, for better or worse, produced a level of concentration, urgency, and operational intensity that public institutions did not demonstrate.

This argument deserves to be taken seriously.

It explains why private firms, not universities or ministries, built the most visible frontier systems. It explains why capital has a claim. It explains why private ownership cannot simply be dismissed as parasitic.

But the argument proves too much.

If the justification for concentrated capital returns is the difficulty of integration, then the return should scale with the value of integration. It should not automatically extend to the full surplus generated by the underlying capability, including the value created by public science, collective data, public infrastructure, institutional trust, and inherited civilization.

A contractor who assembles a building on land cleared by others does not thereby acquire a moral claim on the full value of the location.

A company that integrates AI capability has a claim on the integration.

It does not automatically have a complete moral claim on the civilization-built substrate from which the capability draws value.

That is where the old justification thins.

The deeper problem is that the AI surplus is not primarily behaving like a return to entrepreneurial risk in the traditional sense. The largest returns are flowing to actors that accumulated advantaged positions in compute, data, distribution, cloud infrastructure, model access, chips, talent, and customer interfaces. These positions increasingly function less like ordinary productive risk and more like estate.

A company that owns the dominant chip architecture does not win only because it took a brave bet and others did not. It wins because it controls a bottleneck through which the rest of the economy must pass.

That is not the risk argument.

That is the rent argument.

The Surplus Is Not Purely Private

The foundation of frontier AI is not private.

It is cumulative.

Modern AI rests on three collective substrates.

The first is the knowledge substrate. The mathematical and statistical methods underlying modern machine learning emerged from decades of publicly funded research in universities, government laboratories, and international scientific communities. Backpropagation, transformer architectures, reinforcement learning, deep learning, optimization methods, and the broader toolkit of machine intelligence were built through long chains of research that no single firm privately created. Many foundational papers were published openly. Much of the work was funded by governments, universities, foundations, and academic institutions whose purpose was knowledge production, not capital accumulation. [1]

Mariana Mazzucato’s account of the entrepreneurial state documents this pattern across multiple technology generations: public investment absorbed the early-stage risk of foundational research while private firms later captured the commercial returns. The pattern she identifies in semiconductors, the internet, and pharmaceutical research recurs in AI. The public funded the substrate; private capital assembled the application. [2]

Stanford’s 2025 AI Index confirms the current landscape: the majority of frontier model development now occurs inside private laboratories, but the research lineages on which leading systems are built extend deep into publicly funded academic and governmental science. [3]

The second is the social and data substrate. AI systems are trained on the accumulated output of human communication: language, images, code, documents, websites, books, forums, transactions, searches, conversations, preferences, behavior, and cultural memory. This material was produced by billions of people living inside social systems, not by the firms that later transformed it into model capability. The internet did not become valuable because one company spoke into it. It became valuable because civilization externalized itself into digital form.

The third is the institutional substrate. AI depends on public law, contract enforcement, stable markets, schools, power grids, semiconductor supply chains, telecommunications networks, state procurement, export-control regimes, intellectual-property systems, financial markets, and public trust. The model may run inside a private data center, but the conditions that make the data center usable are legal, infrastructural, political, and civilizational.

None of this means private firms contributed nothing.

They did.

The investment required to assemble compute, talent, capital, data pipelines, safety teams, product layers, distribution, enterprise sales, cloud capacity, and operational infrastructure is genuinely enormous. Private firms did not merely discover AI sitting on the ground. They integrated it, scaled it, packaged it, deployed it, and made it economically real.

But investment in the final assembly layer is not the same as having created the full substrate.

A company can build the tower.

That does not mean it built the city.

A firm can own the model.

That does not mean it produced the civilization that made the model trainable.

The Most Valuable Idea Was Given Away

The pattern becomes concrete when one idea is traced from origin to capture.

In June 2017, eight researchers at Google published a paper called “Attention Is All You Need.” It described the transformer, the architecture that now sits beneath nearly every frontier AI system: the GPT series, Claude, Gemini, Llama, and the models behind most of the AI products the market currently prices. The paper was released openly. Anyone could read it. Anyone could build on it. Many did. [1]

The transformer did not appear from nowhere. It compressed decades of accumulated research, most of it publicly carried. Backpropagation, the training method beneath modern deep learning, was developed in academic settings in the 1980s. The attention mechanism the paper generalized came out of university research on machine translation in Montreal. The neural network tradition itself survived two funding winters because public agencies and nonprofit institutes kept paying for unfashionable work: DARPA and the National Science Foundation in the United States, and Canada’s CIFAR program, which funded Geoffrey Hinton, Yoshua Bengio, and Yann LeCun through the years when neural networks were widely considered a dead end. The dataset that proved deep learning could work at scale, ImageNet, was built by academics on public grants. [1]

So the sequence runs in three steps. Public money carried the research through the decades when no market wanted it. A corporate lab assembled the architecture and published it freely, inside the open scientific culture those decades had built. Then the capture began.

Not through patent. Google asserted no exclusive right over the transformer. The capture ran through what the architecture required in order to become valuable: compute at frontier scale, proprietary data pipelines, engineering talent concentrated by capital, and distribution through interfaces a handful of firms already owned. OpenAI built GPT on an open architecture. Anthropic and Google built on the same foundation. The idea was free.

The estate required to operationalize the idea was not.

This is the moral problem in miniature. The most valuable single input to the AI economy was produced by the open, publicly carried research system and given away. The surplus it generates is collected by the actors positioned to enclose what the idea needs in order to run. The justification for that collection cannot be that capital created the idea. Capital did not create the idea.

Capital owns the conditions under which the idea became operational.

Ownership Captures What Society Made Possible

The AI surplus is now accumulating around bottlenecks.

Nvidia’s fiscal 2026 data-center revenue reached a record $193.7 billion, up 68 percent year over year, with gross margins holding in the 71 to 75 percent range. In the first quarter of fiscal 2027, data-center revenue rose a further 92 percent. Margins of that magnitude, at that scale, in the enabling hardware of a new production system, are not merely evidence of a strong product cycle. They reveal the pricing power that appears when one layer becomes the passage point through which the rest of the economy must move. The previous essay traced that pricing power to its source: the estate structure of compute itself. [4]

That matters morally.

Those margins are not only compensation for invention. They are the price society pays when a collectively enabled intelligence layer passes through a privately owned bottleneck.

Corporate AI investment reached $252.3 billion in 2024, with private investment rising 44.5 percent. U.S. private AI investment reached $109.1 billion, nearly twelve times China’s $9.3 billion and twenty-four times the U.K.’s $4.5 billion. [5]

These numbers are not neutral indicators of technological progress. They describe the capitalization of a new command layer.

The ownership structures of the major AI firms are narrow. The compute estate is held by a small number of hyperscalers. The chip supply chain flows through an even smaller number of critical firms. The venture capital that funded frontier AI rounds concentrated among already-advantaged investors, founders, executives, and institutions. The interfaces through which users encounter AI are controlled by firms that already own distribution, operating systems, cloud infrastructure, enterprise software, search, advertising, payments, and productivity suites.

The surplus does not distribute itself across the civilization that made it possible.

It concentrates where ownership is already positioned.

That is the moral problem.

The AI surplus is built on public science, but captured by private ownership. It is built on collective data, but monetized by private platforms. It is built on public infrastructure, but operated by private cloud providers. It is built on open protocols, but distributed through proprietary interfaces. It is built on state-backed semiconductor research, public procurement, legal protection, and geopolitical strategy, but priced by private firms whose access decisions are not democratically accountable.

The surplus is private.

The conditions were collective.

That gap is the moral problem of AI capital.

When Both Claims Weaken at Once

The problem is not only that capital’s claim weakens in isolation.

It is that labor’s claim and capital’s claim weaken simultaneously while the surplus continues to accumulate.

In the old industrial order, labor and capital disputed a surplus whose moral ownership was contested but whose production was jointly necessary. Capital could not produce without labor. Labor could not organize at industrial scale without capital. The conflict between them was genuine and often brutal, but it had a certain structural honesty. Both sides had a real claim because both sides were visibly necessary.

AI changes that structure.

If AI systems can increasingly perform cognitive tasks that human workers previously supplied, then the traditional labor claim becomes harder to assert in its old form: I contributed my judgment, my effort, my time, and therefore I deserve a share.

That is the labor-side legitimacy problem.

But if the productive system is increasingly built from public science, collective data, inherited language, social trust, state-backed infrastructure, and civilizational accumulation, then the traditional capital claim also becomes harder to assert without qualification: I built this system, and therefore I deserve the return.

That is the capital-side legitimacy problem.

This is the unusual feature of the AI transition.

The two great modern claim structures weaken together.

Labor can no longer rely as confidently on necessity.

Capital can no longer rely as confidently on risk.

Yet the surplus does not pause while legitimacy catches up. It keeps flowing toward ownership. It keeps accumulating around bottlenecks. It keeps rewarding control of compute, models, chips, cloud capacity, data, distribution, payment rails, and interfaces.

As labor’s claim weakens, society asks workers more insistently: what did you do to deserve income?

But when the same question is applied to capital capturing AI surplus, the answer becomes thinner than before.

Not because capital did nothing.

Because what capital did increasingly resembles ownership of accumulated position rather than ongoing productive contribution.

Capital is not contributing more as labor contributes less.

It is capturing more as the moral architecture that justified both claims dissolves together.

The Political Danger of Unjustified Surplus

When surplus concentrates without a publicly legible justification, societies generate resentment that cannot be solved by redistribution alone.

People do not only object to having less.

They object to the story that says others deserve more.

That distinction matters. Poverty is material. Inequality is comparative. But unjustified surplus is moral. It tells people not only that they have less, but that the system no longer has a credible explanation for why someone else has more.

When that explanation fails, the surplus becomes politically volatile regardless of its absolute size.

The political history of capitalism is full of moments when the problem was not deprivation alone. It was the inability of the prevailing distribution to explain itself to the people living inside it.

The political responses to unjustified surplus follow a recognizable pattern. Populist movements direct anger at the institutions and actors visibly capturing returns that society does not recognize as earned. Antitrust pressure intensifies as the connection between market power and positional control becomes harder to ignore. Nationalization demands emerge, not always from coherent ideology, but from the intuition that infrastructure everyone depends on should not be governed purely as private property. Windfall tax proposals proliferate because they acknowledge, implicitly, that some portion of the surplus was not earned in the ordinary market sense.

None of these responses is automatically correct.

Nationalization can destroy value. Windfall taxes can be badly designed. Antitrust can misidentify the source of concentration. Populism can direct resentment toward symbolic targets while leaving the deeper structure intact.

But these responses are not random.

They are symptoms of the same underlying failure: a society that cannot explain why the surplus is distributed the way it is will generate pressure to redistribute it by other means.

The pressure does not wait for a coherent theory.

It arrives as rage.

And rage is a less disciplined instrument than legitimacy.

The danger is not that people will demand too much. The danger is that they will demand in incoherent ways because the system has not given them a coherent account of why the current distribution deserves obedience.

The Claim Structure After Labor

The question is not whether capital should exist.

The question is whether capital can still justify its claim to AI surplus under conditions where that surplus depends on infrastructure that is not purely private in origin, where private ownership increasingly resembles positional control, and where the wage-based moral framework that once balanced capital is weakening at the same time.

This is not anti-capitalism.

It is a crisis of explanation.

Private investment still matters. Risk still matters. Entrepreneurship still matters. Markets still matter. Coordination still matters. The ability to build, integrate, deploy, and scale new systems remains real. A society that destroys those capacities will not liberate the future. It will merely make itself poorer, slower, and more dependent on other people’s infrastructure.

But a society that treats ownership as the final answer will face a different failure.

It will preserve the legal form of property while losing the moral explanation for why property should command the surplus of synthetic production.

When intelligence was primarily a human property, the economy could distribute surplus through wages, careers, firms, and markets because the primary productive input, human cognitive labor, was widely distributed across the population. People could claim income because their work remained close enough to production. Capital could claim return because it remained close enough to risk.

When intelligence becomes infrastructure owned by a small number of actors, both explanations weaken.

The wage no longer explains enough.

Ownership no longer explains enough.

A new claim structure becomes necessary.

The moral vocabulary for that structure is still underdeveloped. Some propose data dividends, payments to individuals whose behavior, language, creativity, and digital traces helped train the systems now generating surplus. In April 2026, OpenAI’s policy framework proposed a public wealth fund that would give citizens a stake in AI-linked assets, with returns distributed to the public. Whatever its motive, the proposal concedes the structural point: AI surplus cannot be narrated only as private corporate property. Others propose compute taxation, data-center levies, windfall-profit mechanisms, sovereign wealth funds, public equity stakes, or broader public ownership of AI infrastructure. [6]

None of these mechanisms is a complete settlement.

They are early attempts to find language for a problem the existing framework cannot name cleanly.

What they share is an implicit recognition: if the surplus of synthetic production depends on collectively created conditions, then the claim on that surplus cannot be purely private, regardless of who owns the final system.

This recognition does not require denying private investment.

It requires denying that private investment is the whole story.

The inheritance grammar matters here. AI is not built only from capital expenditure. It is built from accumulated civilization: science, law, language, infrastructure, institutions, culture, public order, education, data, and trust. If synthetic productivity draws from that inheritance, then membership in the civilization that produced it becomes part of the claim.

That does not abolish capital.

It places capital inside a larger moral architecture.

Capital can still earn returns. Builders can still be rewarded. Risk can still be compensated. Innovation can still be prized. But ownership cannot remain the only surviving grammar of claim after labor, contribution, public inheritance, and social dependence have all been stripped of explanatory force.

The danger is not that capital earns.

The danger is that capital becomes legally absolute precisely as it becomes morally thinner.

Then the system does not say: this surplus is deserved.

It says only: this surplus is owned.

That may be enough for contract law.

It will not be enough for legitimacy.

A civilization can tolerate inequality when it believes the distribution still has a story. It can tolerate wealth when wealth appears connected to creation, risk, sacrifice, and contribution. It can tolerate private ownership when private ownership remains visibly bound to public benefit.

But if AI concentrates surplus in systems built on collective inheritance, while labor loses necessity and capital retreats into title, the old story fails.

The economy will keep producing.

The interfaces will keep allocating.

The compute estate will keep collecting rent.

The owners will keep owning.

But the question underneath will not disappear.

What gives anyone a rightful claim on the world synthetic production creates?

If labor can no longer answer alone, capital cannot answer alone either.

That is the crisis.

Not that ownership exists.

That ownership may become the last answer standing after every better explanation has failed.

Notes

[1] On the research lineage: David Rumelhart, Geoffrey Hinton, and Ronald Williams, “Learning Representations by Back-Propagating Errors,” Nature 323 (1986); Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio, “Neural Machine Translation by Jointly Learning to Align and Translate” (2014), https://arxiv.org/abs/1409.0473; Ashish Vaswani et al., “Attention Is All You Need,” Advances in Neural Information Processing Systems 30 (2017), https://arxiv.org/abs/1706.03762. On public and nonprofit funding carrying neural network research through periods of market neglect, including CIFAR’s Neural Computation and Adaptive Perception program, launched in 2004, which supported Geoffrey Hinton, Yoshua Bengio, and Yann LeCun: https://cifar.ca/research-programs/learning-in-machines-brains/. ImageNet was built at Princeton and Stanford with U.S. federal research support: Jia Deng et al., “ImageNet: A Large-Scale Hierarchical Image Database,” CVPR 2009.

[2] Mariana Mazzucato, The Entrepreneurial State: Debunking Public vs. Private Sector Myths, Penguin, 2013, revised edition 2023. Mazzucato documents the pattern of public investment absorbing early-stage research risk while private firms capture the later commercial returns, across semiconductors, the internet, and pharmaceutical research.

[3] Stanford HAI, 2025 AI Index Report. Documents the institutional landscape of frontier AI development and the research lineages underlying leading systems. https://hai.stanford.edu/ai-index/2025-ai-index-report

[4] Nvidia fiscal 2026 annual results. Full-year revenue reached $215.9 billion, up 65 percent from the prior year; data-center revenue rose 68 percent to a record $193.7 billion; full-year gross margin was 71.1 to 71.3 percent. In Q1 fiscal 2027 (ended April 26, 2026), data-center revenue reached a record $75.2 billion, up 92 percent year over year, with gross margin of 75 percent. https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Fourth-Quarter-and-Fiscal-2026/default.aspx

[5] Stanford HAI, 2025 AI Index Report: Economy. Corporate AI investment reached $252.3 billion in 2024, with U.S. private investment at $109.1 billion. https://hai.stanford.edu/ai-index/2025-ai-index-report/economy

[6] OpenAI policy framework, April 2026, reported by TechCrunch, April 6, 2026. https://techcrunch.com/2026/04/06/openais-vision-for-the-ai-economy-public-wealth-funds-robot-taxes-and-a-four-day-work-week/. On compute taxation, sovereign wealth funds, windfall clauses, and broader ownership mechanisms, see Brookings Institution, “The Future of Tax Policy: A Public Finance Framework for the Age of AI,” February 2026. https://www.brookings.edu/articles/future-tax-policy-a-public-finance-framework-for-the-age-of-ai/

Notice what is missing from this account. Not force. Not the willingness to declare an ending. What is missing is the ending itself.

Iran is only the doorway. The pattern is larger than any war, president, or region. It is the shape of almost everything now. We have built systems powerful enough to prevent collapse and too distributed to produce conclusions. The crisis does not resolve. It does not even fail. It is converted into a condition, and the condition persists.

We Got Good at Preventing Endings

We tend to read this as failure, as evidence that we have grown worse at solving our problems. The opposite is closer to the truth. We have grown extraordinarily good at one specific thing, which is preventing endings. Stalling is not incompetence. It is a competence, and we have become its masters.

Look at how much of contemporary life now refuses to conclude. Wars become frozen conflicts, neither won nor lost, settling into the landscape like weather. Pandemics do not end; at some point they simply stop being mentioned. Emergency powers, declared for a season, remain law for a decade. Budgets are not passed but extended, then extended again. The debt ceiling is raised, the deadline reset, the showdown rescheduled. Inflation is transitory until it is structural until it is simply the price of things. And beneath the scale of nations the same pattern runs through ordinary life: the reorganization that is always underway and never complete, the relationship that neither commits nor breaks, the forty open tabs that will never be read and never be closed. Somewhere we stopped expecting things to conclude, and we did not notice the moment it happened.

The reason is not mysterious. An ending is expensive. To conclude anything, someone has to absorb the cost of finality: to declare the matter over, accept the version of reality that the ending makes permanent, and forfeit every option the open question was still holding open. That is what a verdict is. That is what a surrender is. That is what any real decision is. It closes doors that cannot be reopened, and it assigns the closing to a name.

Modern systems are built so that no one has to do this. Responsibility is distributed so widely that no single part holds enough of it to end anything; authority is real but partial, so that everyone can stall and almost no one can finish. The highest value such a system knows how to protect is optionality: the keeping-open of every door, which is the exact opposite of a conclusion. So crises drift from events into conditions. The objective shifts quietly from resolving the problem to preventing it from getting worse, and management, which was supposed to be the holding pattern, becomes the permanent destination. The limiting factor is no longer power. It is resolution capacity, the ability to align enough of the system behind a single irreversible outcome, and that capacity is what we have lost.

Modernity learned to prevent catastrophe. It did so by abolishing the verdict.

Stalling Is Often Mercy

It is worth saying plainly that this is not all bad, and pretending otherwise would be dishonest. A war that stalls kills fewer people than a war fought to the end. A crisis converted into a managed condition is a crisis that has not been allowed to become a catastrophe. The vast apparatus of indefinite postponement has prevented an enormous amount of suffering, and the people who keep dangerous situations from resolving violently are doing real and difficult work. In the moment, stalling is often mercy.

But a civilization that cannot end things also cannot learn. Endings are how a system finds out it was wrong. A war that concludes delivers a verdict, and the verdict can be studied; a war that stalls delivers nothing, and the questions it raised stay open, unanswered, available to be reopened by anyone at any time. The same holds for every unresolved condition. The correction never comes, because a correction requires a conclusion to correct against. A system that avoids collapse by avoiding resolution slowly loses its resolution capacity, and a system that cannot conclude cannot correct.

This is why the stall, in the end, feels worse than failure. Failure is clarifying. It tells you something is over and forces you to begin again somewhere new. The stall tells you nothing. It only extends, indefinitely, everything that was meant to be temporary, until the temporary becomes the texture of the age. We are not living through a run of separate crises. We are living inside the permanent ambience of crises that were never permitted to end, each one humming in the background, none of them resolved, all of them maintained.

The Hunger Goes Looking for a Person

And the hunger for an ending does not disappear when the system stops supplying one. It accumulates. People can endure hardship that is going somewhere far longer than they can endure a hardship that merely continues, and a population held for years inside an unresolved condition does not grow calm. It grows desperate for closure of any kind. This is the most dangerous appetite a society can carry, because eventually someone arrives who promises to deliver the ending the system cannot: cleanly, finally, by force if that is what finality requires. It is an old pattern. Exhausted societies have always turned, in the end, to the figure who promises to cut the knot, and the promise lands hardest precisely where ordinary procedure has proven it cannot. The demand for conclusion, denied by the institutions, goes looking for a person. That is what the inability to end eventually produces. Not peace, but a market for whoever will promise the ending no one else can.

The newest and most powerful engines of this condition are the systems we have built to run by themselves. A process that depends on human attention eventually exhausts it, and exhaustion forces a decision. A process that runs on its own never tires. Fatigue was one of the oldest forcing mechanisms we had. Wars ended when armies could no longer be fed, sieges when the besiegers ran out of patience, ordeals when the people sustaining them simply gave out. The limit was rarely the will to continue. It was the body and the budget and the attention reaching the end of what they could hold, and an ending arrived because someone, somewhere, could no longer carry the open question another day. Automation does not merely make decisions faster. It removes that floor. It abolishes the exhaustion that used to make an ending necessary, and the suspension now holds for as long as the power stays on, routing around each breakdown, absorbing each shock, keeping the surface functioning while the underlying question goes permanently unanswered. This is the quiet danger named elsewhere in this body of work, that the machine still works, and precisely because it still works, the reckoning never arrives. The more capable our systems become at sustaining a condition, the less able we are to end it. Competence at continuation is the new face of paralysis.

So the endings recede, one at a time. The war always sixty days from over. The emergency always about to be lifted. The decision always nearly made. We call this stability.

It is not. A system that can no longer end anything has not become safe. It has become permanent.

Notes

[1] The June 2025 Twelve-Day War ended in a ceasefire the U.S. president announced in all-capital letters, declaring the conflict over; it held until U.S. and Israeli strikes on Iran resumed on 28 February 2026, beginning the 2026 war. Britannica, “2026 Iran war,” https://www.britannica.com/event/2026-Iran-war.

[2] The ceasefire announced on 7 April 2026 was originally intended to last two weeks while the two sides finalized an agreement. CNN, “How many times has Trump claimed an Iran deal is around the corner?” (9 June 2026), https://www.cnn.com/2026/06/09/politics/times-trump-iran-deal-close; Britannica, “2026 Iran war,” https://www.britannica.com/event/2026-Iran-war.

[3] On 21 April 2026 the president extended the ceasefire indefinitely, saying it would hold until Iran submitted a proposal for talks. House of Commons Library, “US-Iran ceasefire and nuclear talks in 2026,” https://commonslibrary.parliament.uk/research-briefings/cbp-10637/.

[4] By CNN’s count, the president said an Iran deal was imminent more than three dozen times between late March and early June 2026, with no agreement following any of the claims. CNN, “How many times has Trump claimed an Iran deal is around the corner?” (9 June 2026), https://www.cnn.com/2026/06/09/politics/times-trump-iran-deal-close.

[5] The memorandum of understanding announced 14 June and signed 17 June 2026 set a sixty-day window toward a formal end to the war but was a preliminary framework rather than a final agreement, leaving the status of Iran’s nuclear program and enrichment to later negotiations. Britannica, “2026 Iran war,” https://www.britannica.com/event/2026-Iran-war; NPR, “What you need to know about the preliminary U.S.-Iran agreement signed by Trump” (19 June 2026), https://www.npr.org/2026/06/19/nx-s1-5863544/trump-us-iran-agreement.

[6] The first round of technical talks meant to follow the agreement, scheduled for Switzerland, was postponed at short notice; the Swiss Foreign Ministry confirmed the negotiations were called off, and the U.S. vice president’s planned trip was put off (18-19 June 2026). The Washington Times, “First round of U.S.-Iran talks in Switzerland postponed” (19 June 2026), https://www.washingtontimes.com/news/2026/jun/19/first-round-us-iran-talks-switzerland-postponed/; The Times of Israel, “Opening round of US-Iran talks canceled as Tehran said to demand halt to Lebanon fighting” (19 June 2026), https://www.timesofisrael.com/first-round-of-us-iran-talks-in-switzerland-called-off-clouding-prospects-for-lasting-truce/.

The history of political economy is largely a history of what counts as the ground.

Land was the original ground. Whoever owned it received rent. Whoever rented it owed service. The material substrate of production, including soil, water, pasture, forests, and mineral deposits, concentrated wealth and power in the hands of those who held title. The peasant could cultivate. The lord collected. The feudal order was not only an ideology. It was a consequence of what the ground was and who controlled it.

Industrial capitalism shifted the ground. Factories, machines, railways, canals, ports, and coal mines replaced land as the primary substrate of production. Capital became the thing that organized economic life. The worker could sell labor. The capitalist collected surplus. The logic was different from feudalism, but the question remained recognizable.

Who owns the substrate from which value flows?

The AI economy is asking that question again.

Compute is becoming the new ground.

Not merely as metaphor. In political-economic terms, compute is the material substrate of synthetic production. Intelligence does not exist in the abstract. It runs on chips, inside data centers, through energy grids, cooled by water, connected by fiber, financed by long-duration capital, and protected by law.

The model that generates value is not floating freely.

It is sitting on estate.

And estate, as history shows, produces rent.

Compute Is Not a Factory. It Is Estate.

The distinction matters because factories and estates behave differently.

A factory is productive capital. It creates value by transforming inputs into outputs. Its value derives from what it makes. Factories depreciate, become obsolete, require labor, face competition, and can be replicated when the economics justify it. A factory makes its owner a producer.

Estate is different.

Estate is territorial capital. It is productive, but also positional. It does not merely generate output. It controls the conditions under which other actors can produce at all. Its value derives not only from what it makes, but from what others must pay to access it.

Land earns rent not because the landowner works harder than the cultivator, but because the land is where cultivation must happen.

Compute is acquiring the properties of estate, not factory.

A frontier AI system requires advanced semiconductors that only a narrow set of firms can design, fabricate, package, and deploy at scale. It requires data centers whose construction costs rise as facilities are redesigned for AI workloads. It requires enormous and reliable electricity supply. It requires cooling systems, land, fiber connectivity, grid interconnection, water access, regulatory approval, specialized labor, security, and continuous capital expenditure.

These are not ordinary software inputs. They are conditions of existence.

You cannot spin up a competing compute estate the way you spin up a competing website. The barriers are physical before they are financial. They sit in fabs, power grids, data-center campuses, water systems, transmission lines, export-control regimes, procurement contracts, and capital markets.

That is the first property of estate: it cannot be easily replicated.

The second property is positional power.

Estate produces rent because others must pass through it to produce. The AI compute estate produces rent for the same reason. Intelligence can be accessed through APIs, cloud subscriptions, inference services, enterprise licenses, hosted models, and managed platforms. But that access comes at prices, on terms, under policies, and within availability constraints set by the estate owner.

The estate owner may also be productive. Hyperscalers, chipmakers, cloud providers, and model companies are not passive landlords. They build, engineer, maintain, secure, and improve the systems they control.

But their deeper power comes from owning the ground on which others become productive.

The tenant rents access to the ground.

The estate owner controls the lease.

The Physical Ground of Synthetic Production

It is easy to describe the AI economy as software, models, algorithms, and intelligence. That abstraction is partly real. AI does move through digital channels in ways that land, coal, steel, and railways did not.

But the abstraction hides the physical base.

Synthetic intelligence remains anchored to material substrate. That substrate creates scarcity. Scarcity creates concentration. Concentration creates rent.

Start with chips.

Advanced semiconductor manufacturing is one of the most concentrated industrial systems in the world. TSMC accounted for roughly 70 percent of pure-play foundry revenue in 2025, and Counterpoint Research put its share of that market near 72 percent in the third quarter of 2025. Industry estimates also place TSMC’s share of the most advanced process nodes around 90 percent or higher. Those are the nodes that matter most for AI accelerators, frontier models, and high-performance compute. [1]

This concentration did not arise because chips are easy to monopolize in the ordinary sense. It arose because leading-edge manufacturing is extremely hard to reproduce. It requires extreme ultraviolet lithography, specialized equipment, advanced packaging, deep process knowledge, high yields, trusted customer relationships, and decades of accumulated manufacturing discipline.

Building a leading-edge fab takes years and billions of dollars. Even state-backed attempts to rebuild domestic semiconductor capacity move slowly. Intel’s Ohio semiconductor project, originally expected to begin production in 2025, has been pushed toward 2030 to 2031. Micron’s New York megafab has also faced a two-to-three-year construction delay, with the first facility’s operation pushed toward the end of the decade. [2] [3]

Political will can subsidize the chip estate.

It cannot instantly reproduce it.

And reproduction, when it comes, does not break the estate. Land was never scarce because the planet ran short of dirt. There is plenty of dirt. Land was scarce because position was scarce: the field had to be where it was, and acquiring more of it required already holding the means to acquire it. Compute is the same. The capital now pouring into fabs and data centers does not dissolve the estate. It builds more of it. Every increment of leading-edge capacity is raised by the same narrow set of actors who could afford the last one, financed on terms only they can meet, on ground only they can secure. Supply expands. Ownership does not. The buildout is not the estate eroding. It is the estate compounding.

Below chips sits energy.

AI computation does not only require electricity. It requires dense, stable, predictable electricity at locations where data-center capacity can be built and connected. The International Energy Agency projects that global data-center electricity consumption could more than double by 2030, reaching roughly 945 terawatt-hours, comparable to Japan’s current electricity consumption. The same IEA analysis notes that capital expenditure by five major technology companies surged to more than $400 billion in 2025 as they raced to build the infrastructure needed for frontier AI. [4]

This matters because energy is not a marginal input to AI. It is the binding condition of synthetic production.

Intelligence does not run where power cannot reach.

Below energy sit land and water.

AI data centers require large sites near power infrastructure, cooling capacity, fiber networks, and permissive regulatory environments. Cooling systems require water or energy-intensive alternatives. Local governments increasingly face conflicts over grid pressure, land use, water consumption, tax incentives, noise, environmental impact, and public benefit. In 2023, US data centers directly consumed about 17 billion gallons of water, with hyperscale and colocation facilities accounting for most of that direct use. Hyperscale data centers alone are expected to consume between 16 billion and 33 billion gallons of water annually by 2028. [5]

A model may appear weightless to the user.

The estate beneath it is heavy.

And below all of this sits law.

Data centers require permits, zoning approval, environmental review, energy contracts, and regulatory clearance. Export controls govern who can obtain the most advanced chips. Procurement rules shape who can sell AI into governments, hospitals, banks, schools, and corporations. National AI strategies increasingly ask where computation happens, who owns it, under whose jurisdiction it runs, and who can be excluded from it.

Law does not merely regulate compute.

Law helps make compute scarce.

Scarcity is what allows estate to produce rent.

The physical ground is not a metaphor. It is where the intelligence actually lives.

Why Compute Produces Rent

Rent is not the same as profit.

Profit is the return to productive enterprise. It rewards risk, labor, capital, invention, and execution. Rent is the return to positional control over something others must access in order to produce.

Compute produces rent for four reasons.

First, scarcity at the leading edge.

The most capable models require the most advanced chips. The most advanced chips are produced through supply chains that are narrow, specialized, and difficult to replicate. Scarcity at the foundational layer propagates upward through the entire AI stack. Whoever controls access to leading-edge chips does not merely run a factory. They hold a choke point on the envelope within which frontier AI can exist.

Nvidia’s fiscal 2026 full-year revenue reached $215.9 billion, up 65 percent from the prior year, while data-center revenue rose 68 percent to a record $193.7 billion. In the first quarter of fiscal 2027, data-center revenue hit a further record of $75.2 billion, up 92 percent year over year. Gross margins held in the 71 to 75 percent range across the full year. Margins like that, at that scale, in the enabling hardware of a new production system, are not merely evidence of a strong product cycle. They reveal the pricing power that appears when one layer becomes the bottleneck through which the rest of the economy must pass. [7]

Second, scale thresholds.

Training frontier models, running inference for large user populations, supporting enterprise automation, and providing reliable AI services require infrastructure that most actors cannot own. The capital expenditure required exceeds the capacity of almost every startup, small firm, university, city, nonprofit, and mid-sized state. McKinsey estimates that AI-related data-center infrastructure alone could require roughly $5.2 trillion in cumulative capital expenditure by 2030. [6]

At the frontier, compute ownership is measured not in servers, but in campuses, power agreements, supply contracts, specialized teams, and multiyear financing.

This converts many potential competitors into renters before competition begins.

Third, integration costs.

A compute estate is not only hardware. It is the accumulated knowledge of how to operate hardware at scale, integrated with software systems, security architecture, compliance processes, latency optimization, energy management, enterprise support, and trust layers.

This is not a one-time build. It is an ongoing accumulation of integrated capacity.

The deeper the integration, the harder the estate is to reproduce.

Fourth, interface extraction.

Most tenants do not access compute directly. They access it through interfaces: model APIs, cloud contracts, inference endpoints, enterprise AI platforms, managed services, productivity suites, app stores, and procurement-approved vendors.

These interfaces allow the estate owner to price access, enforce usage policies, modify capabilities, deprecate models, bundle services, throttle usage, adjust safety rules, and capture value from the gap between the cost of running intelligence and the tenant’s dependence on the intelligence produced.

The tenant does not see the estate.

She sees the interface.

The gap between what it costs to run intelligence and what the tenant pays for access is where interface extraction lives. The estate owner prices not to the cost of the silicon but to the value of the dependency. A firm that has reorganized its operations around a model API does not face the price of electricity and amortized hardware. It faces the price of its own inability to function without access. That is the structural basis of interface pricing power, and it is what distinguishes the compute estate from ordinary productive capital. A factory charges for what it makes. The estate charges for what you cannot do without it.

The rent is collected before the ground becomes visible.

Open Weights Are Not Open Estate

The strongest objection to the compute-estate argument is open-source AI.

If foundation models can be released with open weights, then perhaps the estate argument fails. The intelligence becomes replicable. Anyone can download a model, fine-tune it, and deploy it without paying rent to a hyperscaler. The tenant condition dissolves when the land is free.

This objection contains a real insight.

It also misses the structural point.

Open weights are not open estate.

A powerful open model on a laptop is not the same as institutional AI capacity. It is not a substitute for a secure, audited, scalable, insured, compliant, supported, enterprise-grade system that a hospital, bank, government agency, defense contractor, school system, or multinational corporation can adopt.

The gap is not only model capability, although capability matters.

The gap is the full stack required to make intelligence economically and institutionally real.

Running an open model at scale still requires compute. The model weights may be free. The inference infrastructure is not. A startup serving millions of users still needs GPU access, hosting, uptime guarantees, security, latency management, observability, compliance, and capital.

Running an open model inside institutions requires trust. Healthcare, finance, government, and enterprise buyers do not adopt AI systems based on model weights alone. They need liability frameworks, audit trails, certifications, service-level agreements, vendor accountability, security reviews, and integration support.

Running an open model over time requires maintenance. Models degrade relative to newer systems. Security vulnerabilities appear. Fine-tuning requires data infrastructure. Evaluation requires human feedback. Deployment requires monitoring. Scaling requires operations. Each layer pushes the actor back toward compute providers, cloud platforms, managed services, and institutional interfaces.

Open-source AI reduces the barrier to intelligence.

It does not eliminate the estate on which intelligence must run when it becomes economically serious.

The tenant condition is not only about model access.

It is about the physical and institutional substrate required to make intelligence usable at scale.

Open source dissolves one gate.

The estate still stands.

The collision between open weights and institutional requirements is not theoretical. It is the operating reality of every regulated sector attempting to deploy AI. In the United States, federal agencies must satisfy FedRAMP authorization requirements for cloud AI procurement. No self-hosted open-weight deployment currently holds FedRAMP High authorization. The path to institutional AI runs through authorized cloud providers: AWS GovCloud, Microsoft Azure Government, and Google Cloud Government. The open model that cannot travel that path does not reach the hospital, the agency, the bank, or the defense contractor regardless of its capability. The compliance architecture is the gate through which the estate extracts rent from institutions that cannot operate outside it. [8]

Compute Produces Sovereignty

Estate is not only economic. It is political.

In agrarian orders, great estates were not merely sources of private wealth. They were the basis of military power, administrative capacity, and political authority. Control of land meant control of food, manpower, taxation, and obligation. Sovereignty was never fully separate from estate.

Compute is acquiring a similar political character.

Not because AI systems already govern directly. The point is more basic. The capacity to operate at the frontier of intelligence, speed, automation, military planning, cyber defense, surveillance, logistics, research, and institutional execution increasingly depends on access to compute.

A state can remain formally sovereign while becoming computationally dependent.

It may have flags, courts, ministries, elections, borders, laws, budgets, police, and diplomatic recognition. But if its actual administrative, military, industrial, and scientific capacity depends on foreign chips, foreign cloud infrastructure, foreign models, foreign cybersecurity tools, and foreign technical expertise, its sovereignty becomes layered.

Legal sovereignty remains.

Operational sovereignty weakens.

The split matters because modern sovereignty was built on the assumption that legal authority could eventually command operational capacity. A state that could legislate, tax, procure, regulate, conscript, and build could translate formal authority into action. Compute weakens that assumption. A government may have the legal right to govern an AI-mediated institution while lacking the chips, cloud capacity, model access, cybersecurity stack, data-center capacity, and technical labor required to execute that authority independently. It can command in law while renting in practice.

That is not the disappearance of sovereignty.

It is the separation of sovereignty into a visible legal layer and a hidden operational layer.

This is why export controls on advanced chips matter. They are not ordinary trade policy. They are attempts to determine who can sit inside the frontier compute estate and who must remain outside it. Restricting access to advanced accelerators is a way of shaping the future boundary of AI capability, military modernization, industrial automation, and state capacity.

The geopolitics of AI is therefore not only a contest over models.

It is a contest over estate.

Who builds the fabs. Who secures the energy. Who controls the grid connections. Who owns the data centers. Who can finance the buildout. Who can sustain the capex cycle. Who can run training at frontier scale. Who controls inference infrastructure. Who gets access during scarcity. Who writes the laws around the estate.

These are not secondary questions that follow from AI capability.

They are the conditions that determine which capabilities can exist.

Singapore makes this condition visible with unusual clarity.

No state has spent more deliberately on the problem of being small. For six decades Singapore built sovereignty through institutional design, financial reserves, diplomatic positioning, legal infrastructure, and military investment calibrated precisely to the vulnerabilities of a city-state with no natural resources, no strategic depth, and no room for error. It is not naive about dependency. It has studied dependency as a condition of national existence.

And even Singapore cannot solve the compute estate problem.

Singapore launched its National AI Strategy 2.0 in December 2023 with over S$1 billion in committed investment across compute, talent, and industry development. The strategy is explicit about ambition: Singapore intends to be a global AI hub, not merely a regional one. Prime Minister Lawrence Wong described AI compute as a strategic necessity in his Budget 2024 speech, committing further investment over five years. [9]

The operational reality behind the strategy is instructive. Singapore’s National Supercomputing Centre runs ASPIRE 2B, its most advanced system, with over 1,500 Nvidia H200 GPUs delivering up to 115 petaFLOPS. At the launch, Singapore’s Minister for Digital Development and Information described the system’s capacity as “nowhere near the cluster sizes available to frontier model developers.” That phrase was not a complaint. It was an honest accounting of where Singapore sits in the compute hierarchy. [10]

Singapore’s government has migrated over 600 digital services and the bulk of its less sensitive government IT systems to commercial cloud infrastructure operated by AWS, Microsoft Azure, and Google Cloud, running through a centralized platform called the Government on Commercial Cloud. More sensitive workloads now run on AWS Dedicated Local Zones, which are physically located in Singapore but fully managed by AWS. The data stays in Singapore. The infrastructure management does not. [11]

AWS has committed over S$23 billion in Singapore cloud infrastructure through 2028. Microsoft has committed S$5.5 billion through 2029. These commitments are genuine investments in Singapore’s digital capacity. They are also the architecture of a dependency that no sovereign strategy can easily undo. Singapore signed the agreements, built the government services on top of them, and trained the workforce around them. The compute estate belongs to the firms that built it. Singapore’s government runs on lease. [12]

Singapore knows this. The updated NAIS priorities released in May 2026 acknowledge directly that “future supply dynamics remain uncertain” on compute access and commit to charting a path toward greater self-sufficiency. But the document also records what the path toward self-sufficiency looks like in practice: partnerships with Nvidia, OpenAI, Google, and Microsoft; agreements to expand data center capacity built and managed by foreign operators; and a national supercomputer that by Singapore’s own government’s account sits below the frontier threshold for the AI work that will define the next decade of institutional and strategic capacity. [9]

This is what the sovereign tenant looks like from the inside: not failure, not weakness, not ignorance. It is the condition of a capable, well-governed state operating at the frontier of institutional sophistication, still unable to own the ground on which its most consequential future capacities will run. If Singapore cannot close the gap between declared AI sovereignty and operational AI dependency, no state below the United States and China in the compute hierarchy can close it either. That is not a forecast. It is the current structure of the estate.

Compute Produces Tenancy

The tenant condition is not only a platform effect.

It is a substrate effect.

When the ground is expensive to own, when ownership requires capital most actors cannot accumulate, when access to the ground is the precondition of productive activity, and when creating competing ground takes years of specialized effort, tenancy becomes the natural result.

Not as a conspiracy.

As a structural consequence of the substrate.

Most firms will not own the compute required to run frontier AI. They will rent access through cloud services, API contracts, enterprise platforms, hosted inference, managed model deployments, and productivity suites.

Most workers will not own the compute that increasingly mediates their productivity. They will use tools selected by employers, priced by vendors, governed by policies they did not write, and hosted on infrastructure they cannot inspect.

Most startups will not train frontier models. They will build on models trained by others, deploy on infrastructure owned by others, distribute through platforms owned by others, and sell into procurement systems that define institutional legibility before the product is judged.

Most states will not own the full stack required for frontier AI sovereignty. They will depend on alliances, vendors, imports, cloud partnerships, chip allocations, foreign expertise, and infrastructure they cannot fully reproduce.

This does not mean no one can build value as a tenant.

Tenants can be creative, profitable, powerful, adaptive, and innovative. They can build companies, products, workflows, media systems, services, and institutions on top of rented intelligence environments.

But they do not own the environment that makes their agency scalable.

That is the class condition emerging inside the AI economy.

The new divide is not only between labor and capital. It is between those who own the environments of synthetic production and those who rent the right to act inside them.

The Landlords of Intelligence

If compute is estate, the owners of compute occupy the structural position of landlords of intelligence.

That claim needs precision.

It does not mean hyperscalers are malicious. It does not mean compute ownership is automatically unjust. It does not mean the relationship is identical to historical landlordism. The major compute owners took enormous capital risk. They built infrastructure that produced real capabilities. They compete with one another. They employ engineers, operate complex systems, and solve difficult technical problems.

The analogy is structural, not moral.

A landlord of intelligence is an actor that owns the environment in which synthetic cognition becomes economically usable.

Cloud providers, AI infrastructure firms, chip designers, foundries, data-center operators, and model-platform companies do not merely sell products. They increasingly own the ground other actors must access to produce, automate, analyze, distribute, and decide.

The API price is not just a fee.

It is a form of rent.

The cloud subscription is not just an operating expense.

It is a lease on synthetic capacity.

The enterprise AI license is not just software procurement.

It is access to the estate where institutional intelligence runs.

The minimum usage fee that converts AI from possibility into operation is rent. The usage policy that determines what can be done with rented intelligence is lease law in platform form. The deprecation notice, throttling rule, safety update, quota limit, pricing tier, and compliance requirement are all part of the legal architecture of the estate.

Nvidia reveals one layer of this structure. It does not primarily rent compute services to end users. It sells the accelerators on which much of the AI compute estate runs. At that layer, the chip is not merely a component. It is a key to the estate.

The obvious objection is that margins like these invite their own destruction. Custom silicon arrives, AMD narrows the gap, inference costs fall every year, and the rent that looks structural today is competed away tomorrow. On that reading the estate is a product cycle wearing the costume of land.

It is the opposite, and the dynamism the objection points to is the proof. Watch who actually contests the frontier. Google fields its own tensor silicon, Amazon builds Trainium, the largest labs negotiate bespoke supply: the challengers to the bottleneck are not tenants breaking in, they are other landlords. The capital, the fabrication access, and the power required to stand up a competing cluster sit at the same order of magnitude that excluded everyone else to begin with, so the competition runs among the few who can own the ground, conducted over the heads of the many who can only rent the right to act on it. Falling unit costs do not change this. Cheaper cognition is met with more demand for it, and the cluster that confers advantage grows rather than shrinks, so the threshold does not disappear. It moves.

Nor does the competition among the owners reach the people renting from them. When Nvidia, AMD, Google, and Amazon contest the frontier, they are bidding against one another for the right to own the ground, not opening that ground to the tenants beneath it. Whatever the rivalry shaves off the price accrues to whichever landlord wins the contract, or to the few renters large enough to negotiate terms. It does not descend to the firm, the worker, or the state that can only take the rate it is offered. The competition is real. It is also sealed. A renter does not become an owner because the owners are fighting.

This is why the landlords may change while the estate logic remains. The rent was never a property of a particular chip or a particular vendor. It is a property of position: the conditions of synthetic production are held by a few and rented by everyone else, and that structure reconstitutes itself each time one owner is displaced by another who can meet its price. A factory can be out-competed into the ground. An estate only changes hands.

Compute does not merely create products.

It creates positions.

The political question is not whether today’s dominant companies remain dominant forever. The deeper question is whether compute itself will remain the kind of ground that produces landlords.

Every structural feature points in that direction: capital intensity, energy dependency, geographic concentration, technical scarcity, legal protection, institutional trust requirements, and scale advantages.

The State Comes Looking for the Ground

States have always tried to tax the ground.

In agrarian economies, land was the ideal fiscal base. It was visible, valuable, immovable, and impossible to hide. A field could not be shifted to a tax haven. Its value was tied directly to the productive capacity of the society around it.

As capital became mobile and intangible, taxation became harder. Profits could be shifted. Intellectual property could be located in favorable jurisdictions. Digital platforms could generate revenue across borders while locating taxable income elsewhere.

The platform era made value harder for states to locate.

Compute begins to reverse that.

The AI economy still produces intangible services, but the capacity to produce them is increasingly anchored in visible, permitted, energy-intensive, immovable sites. The surplus may flow through software, but the ground beneath it has an address.

Compute is heavy. It is energy-intensive. It requires fixed infrastructure. It depends on land, grid access, cooling systems, permits, and local political bargains. A data center cannot be moved to another jurisdiction as easily as an intellectual-property holding company.

This makes compute a future tax base.

Not easily. States will face tradeoffs. Tax too aggressively, and they may drive investment elsewhere. Tax too weakly, and they may subsidize the very estate that later collects rent from their citizens and firms. But structurally, compute looks more like taxable ground than the mobile capital of the platform era.

Data-center taxes, energy surcharges, cloud-service levies, infrastructure fees, compute-based fiscal instruments, and public claims on AI productivity will all become more politically thinkable as the estate becomes more visible.

This is not only a revenue issue.

It is a recognition issue.

The state eventually comes looking for the ground because the ground is where the surplus concentrates.

The compute estate does not only produce rent.

Over time, it invites taxation.

The Land Beneath Synthetic Production

The earlier transformations now acquire their ground.

Production can detach from labor income. Markets can move into interfaces. The wage can weaken as the central grammar of belonging. Independence can become tenancy. But these changes do not float above the world. They require a substrate.

Compute is that substrate.

It is the physical and economic ground that makes rented intelligence durable rather than temporary, systemic rather than incidental. The tenant is not surrounded only because of bad corporate behavior, regulatory failure, or platform ideology. She is surrounded because the ground on which synthetic intelligence runs is expensive to own, slow to replicate, concentrated among a narrow set of actors, and positioned as the precondition for participation in the AI-mediated economy.

That is the compute estate.

It is not ordinary productive capital.

It is territorial capital: positional, rent-producing, sovereignty-generating, and eventually taxable.

The factory produces output.

The estate produces the conditions under which output becomes possible.

In agrarian economies, the political question was who owned the land.

In industrial economies, it was who owned the machines.

In the AI economy, it is who owns the compute, and what everyone else must pay to think.

Compute is the land beneath synthetic production.

And land, once established as the ground, does not need to announce itself to govern.

Sources

[1] Counterpoint Research’s Q3 2025 foundry analysis put TSMC at roughly 72 percent of the pure-play foundry market; TrendForce separately placed TSMC’s pure-play foundry share at about 71 percent in Q3 2025, up from 70.2 percent in Q2 and near 70 percent for the full year. Industry estimates also commonly place TSMC’s share of the most advanced nodes around 90 percent or higher.
https://counterpointresearch.com/en/insights/global-foundry-2.0-market-Q3-2025-revenue

[2] Intel’s Ohio semiconductor project, originally expected to begin production earlier in the decade, has been delayed toward 2030 to 2031.
https://spectrumnews1.com/oh/columbus/news/2025/02/28/intel-delays-ohio-s-chip-plant-to-2030--2031

[3] Micron’s $100 billion New York megafab faced a two-to-three-year construction delay under the revised schedule in its Final Environmental Impact Statement (November 7, 2025); the first fabrication facility’s opening moved from 2028 to 2030, with construction now beginning in 2026 and the full four-fab buildout staggered through 2041.
https://www.constructiondive.com/news/micron-delay-construction-new-york-megafab/805622/

[4] The International Energy Agency projects global data-center electricity demand could more than double to roughly 945 TWh by 2030, comparable to Japan’s current electricity use. Its 2026 analysis also describes surging data-center investment, including more than $400 billion in 2025 capital expenditure by five large technology companies.
https://www.iea.org/reports/key-questions-on-energy-and-ai

[5] Pew Research Center, citing estimates from a 2024 Berkeley Lab report commissioned by the US Department of Energy, reports that US data centers directly consumed about 17 billion gallons of water in 2023, with hyperscale and colocation facilities using 84 percent of that total. It also reports that hyperscale data centers alone are expected to consume between 16 billion and 33 billion gallons annually by 2028.
https://www.pewresearch.org/short-reads/2025/10/24/what-we-know-about-energy-use-at-us-data-centers-amid-the-ai-boom/

[6] McKinsey & Company estimates that data centers equipped for AI processing loads could require roughly $5.2 trillion in capital expenditures by 2030, out of nearly $7 trillion in total data-center infrastructure spending.
https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers

[7] Nvidia reported fiscal 2026 full-year revenue of $215.9 billion, up 65 percent from the prior year, with data-center revenue rising 68 percent to a record $193.7 billion and full-year gross margin of 71.1 to 71.3 percent. In Q1 fiscal 2027 (ended April 26, 2026), data-center revenue reached a record $75.2 billion, up 92 percent year over year, with gross margin of 75 percent.
https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Fourth-Quarter-and-Fiscal-2026/default.aspx

[8] On FedRAMP authorization requirements for federal AI procurement: Office of Management and Budget, FedRAMP program documentation. No self-hosted open-weight AI deployment currently holds FedRAMP High authorization. Federal agencies procuring cloud-based AI must use FedRAMP-authorized providers, which in practice means AWS GovCloud, Microsoft Azure Government, and Google Cloud Government. https://www.fedramp.gov

[9] Singapore Ministry of Digital Development and Information, Update to Singapore’s National AI Strategy, May 2026. The update acknowledges that “future supply dynamics remain uncertain” on compute access and sets out refreshed priorities including expanded data center capacity and deepened partnerships with major AI firms. Prime Minister Lawrence Wong described AI compute as a strategic necessity in Budget 2024. https://www.mddi.gov.sg/newsroom/update-to-singapore-s-national-ai-strategy--refreshed-priorities-to-harness-ai-for-the-public-good-factsheet/

[10] Minister Josephine Teo, Opening Address at NSCC Launch of ASPIRE 2B Supercomputer, Singapore Ministry of Digital Development and Information. The minister stated that ASPIRE 2B’s capacity of up to 115 petaFLOPS with more than 1,500 Nvidia H200 GPUs is “nowhere near the cluster sizes available to frontier model developers.” https://www.mddi.gov.sg/newsroom/opening-address-by-minister-josephine-teo-at-national-supercomputing-centre--nscc--s-launch-of-aspire-2b-supercomputer/

[11] GovInsider, “Key lessons from the Singapore government’s ambitious whole-of-government cloud migration strategy.” GovTech’s Government on Commercial Cloud platform hosts over 600 government digital services running on AWS, Microsoft Azure, and Google Cloud. More sensitive workloads run on AWS Dedicated Local Zones, which are physically located in Singapore but fully managed by AWS. https://govinsider.asia/intl-en/article/key-lessons-from-the-singapore-governments-ambitious-whole-of-government-cloud-migration-strategy

[12] AWS press release: AWS to invest an additional S$12 billion in Singapore by 2028, bringing total planned investment to over S$23 billion. https://press.aboutamazon.com/sg/aws/2024/5/aws-to-invest-an-additional-sg-12-billion-in-singapore-by-2028-and-announces-flagship-ai-programme. Microsoft commitment of S$5.5 billion in Singapore AI and cloud infrastructure through 2029: Computer Weekly, “Microsoft to invest $5.5b in Singapore’s AI and cloud infrastructure,” April 2026. https://www.computerweekly.com/news/366641114/Microsoft-to-invest-55b-in-Singapores-AI-and-cloud-infrastructure