The Revocable License
On the death of American tech rent, the defection of captive allies, and the difference between owning a thing and being permitted to use it.
The Company That Sells You Your Own Sovereignty
There is a particular comedy in watching an arsonist deliver a lecture on fire safety, and something close to that spirit hung over the last day of June 2026, when Palantir Technologies published a nine-point manifesto on the sanctity of AI sovereignty. The company that spent two decades teaching Western governments to pour their most sensitive secrets into a proprietary American platform now wished to warn those same governments about the perils of pouring their most sensitive secrets into proprietary platforms. “Your AI sovereignty dictates your institution’s future,” the first point began. “Relinquishing sovereignty transfers the future choices of your institution to others, who are likely to exploit it for their gain and your loss.” One could only admire the nerve. Here was Palantir, of all the firms on earth, appointing itself the conscience of the sovereignty age, as though the fox had convened a symposium on henhouse architecture and arrived with blueprints.
And yet the thing about a warning is that its source does not determine its truth. A tobacco executive who tells you cigarettes cause cancer is being cynical, opportunistic, and correct all at once. What made the Palantir document worth reading was not its sincerity, of which it had little, but its diagnostic precision, of which it had a great deal. The firm was describing, with the clarity of a company that lives or dies by reading institutional demand, a shift in what its customers now want to buy. That shift, if it is real, unravels one of the two great assumptions on which an alarming portion of the American economy has been quietly resting: the assumption that the companies building frontier artificial intelligence will be able to charge rent for it, rather than watch it dissolve into a commodity like bandwidth, or electricity, or steel.
The commentator Arnaud Bertrand, reading the same document, reached for the old image of the canary in the coal mine, the small bird whose death warns the miners that the air has turned. It is an evocative figure, but it is the wrong bird. A canary’s death is involuntary and, in its accidental way, altruistic; it dies first and by dying protects others, having gained nothing for itself. Palantir is not that. Palantir is the other creature Bertrand invoked almost in passing, the one that gets the metaphor right. “Sometimes it takes a vulture,” he wrote, “to tell you something is dying.” A vulture circles for reasons entirely its own. It has read the situation on the ground more accurately than the animal below, which still believes it is walking. The vulture is not a martyr and not a friend. It is simply an excellent forecaster of death, because death is its business. When Palantir begins to circle the closed-source AI labs, the interesting question is not whether the bird is noble. It obviously is not. The interesting question is what it has seen.
The tell
To understand what Palantir saw, begin with the timing, because the timing gives the game away. The manifesto did not descend from a cloud of disinterested principle. It arrived on June 30, one day after the company announced an expanded partnership with Nvidia to deploy open-weight models, Nvidia’s Nemotron family, inside sovereign, classified, and air-gapped environments. The two firms had already sketched a “Sovereign AI Operating System” reference architecture back in March. The June deal turned the sketch into a product: pair Nvidia’s silicon and open weights with Palantir’s own deployment machinery, its Foundry and its Ontology and its Apollo, and sell the whole assembly to any ministry that wants the power of a large language model without the terror of renting one from a company that might, on a bad Tuesday, decide to switch it off.
So the manifesto is a product launch. Strip away the philosophy and you find a brochure. When Palantir writes that “controlling your weights is controlling your fate,” it is not offering a meditation; it is describing a stock-keeping unit. When Alex Karp went on CNBC the same week to call the frontier labs’ business model “effing insane” and to accuse OpenAI and Anthropic of overselling the dangers of the very models they were happily serving to the entire planet, he was not playing the philosopher-king, though he enjoys the role. He was clearing the field for a competitor. Karp has a gift for delivering a sales pitch in the cadence of a jeremiad, and this was a virtuoso performance: the voice of American business, as he likes to style himself, warning American business about the people trying to sell it the wrong kind of intelligence, shortly before selling it the right kind.
The cleanest statement of the cynicism came from the technology press rather than from any of Palantir’s rivals. For the governments now worried about dependency, one writer observed, true sovereignty means not depending on a single American vendor either, however loudly that vendor preaches independence. Palantir did not invent the anxiety it is monetizing. It is simply very good at selling the cure. Point five of the manifesto insists there is “no contradiction between sovereignty and alpha,” that the ideal architecture lets an institution “own their tribal knowledge, and compound it as alpha.” The unstated codicil, the one the document declines to print, is that you should not hand your tribal knowledge to Palantir either, since a company that spends nine numbered paragraphs explaining why you must never trust a single foreign vendor has an obvious interest in your not applying the lesson to itself.
None of this, to repeat the essential point, makes the diagnosis wrong. It makes the diagnosis motivated, which is a different thing. A motivated diagnosis can be more reliable than a disinterested one, because the party with money at stake has every incentive to see clearly. Palantir has looked at the market it serves, the intelligence agencies and defense ministries and strategic enterprises that constitute the least sentimental buyers on the planet, and concluded that these buyers no longer want to rent intelligence. They want to own it. And a company that concludes its customers want to own the thing rather than rent it is telling you, in the only language it speaks, that the rental business is in trouble.
Why the model is becoming a commodity
The rental business is the AI laboratory. Its dream, the dream underwriting a substantial share of American equity valuations, is that a handful of firms will build models so far ahead of everyone else, and so deeply woven into the fabric of work, that they can charge a premium in perpetuity for access to them. This is the rent thesis. It imagines the frontier lab as a kind of toll booth on a road that everyone must travel, collecting a fee on every token of human and machine cognition that passes through. It is a beautiful dream if you own the toll booth. The trouble is that the road is sprouting detours.
Consider where the tokens actually go. By early 2026, data from OpenRouter, the marketplace that routes developer traffic across competing models, showed that Chinese models accounted for sixty-one percent of total token consumption among the ten most-used models, some 5.3 trillion tokens out of 8.7 trillion. Four of the five most heavily used models were Chinese. Meta’s Llama, the great Western hope for open weights, had fallen out of the rankings entirely. The router’s own operators noted that Chinese open-weight models were disproportionately favored in the agentic workflows that American firms themselves were building, which is to say that when a Silicon Valley startup wired an autonomous system together and cared about the bill, it frequently reached for a model made in Hangzhou.
The reason is price, and the price gap is not marginal. DeepSeek’s latest release was described at launch as an eighty-three percent undercut, listing roughly twelve times cheaper than the comparable American frontier model at similar measured intelligence, seven times cheaper on inputs and seventeen times cheaper on outputs than a leading closed model. Across the board, capable Chinese models were running somewhere between five and thirty times cheaper than their Western counterparts. For that discount you might expect a steep fall in quality, and a few years ago you would have gotten one. You no longer do. The best open-weight Chinese model trailed the top closed Western system by perhaps nine points on a composite benchmark, close enough that on specific tasks the ranking inverted: one Chinese model beat the leading American model on agentic web browsing by a wide margin; another landed within a whisker of the Western frontier on verified software-engineering problems. When the gap between the premium product and the commodity product narrows to single digits, and the commodity costs a twentieth as much, the premium has to justify itself on something other than raw capability. Often it cannot.
The behavior of sophisticated buyers confirms what the benchmarks imply. Airbnb leaned heavily on Alibaba’s Qwen. The coding tool Cursor built a marquee feature on a model from Moonshot. By the middle of 2026, a Chinese model reportedly topped American enterprise adoption in at least one closely watched survey. These are not ideological choices, and in the current political climate they are made in spite of the politics rather than because of it. They are the choices of engineers looking at a bill. Uber, by one account, burned through its annual token budget in four months. When the meter runs that fast, the price of the meter becomes the whole question, and a five-to-thirty-times discount stops being a curiosity and becomes a mandate from the finance department.
There is a further property of open weights that no closed model can match, and it is the property Palantir has built its pivot around. An open-weight model, once downloaded, is irrevocable. The company that released it cannot reach across the network and disable it. It sits on your servers, behind your perimeter, indifferent to the moods of its creator or the instructions of any government the creator answers to. For a residential user this is an abstraction. For a defense ministry it is the entire point, and in mid-2026 it stopped being an abstraction for anyone, for reasons we will come to in the second part of this essay. It is enough for now to note that “you can never be switched off” is a feature only the commodity can offer, and that the premium product’s very architecture, the remote-served closed model, is the source of the vulnerability the market has begun to price.
The arithmetic of the toll booth
Set beside the commoditization of the product is the economics of the firms betting against it, and the economics are vertiginous. According to a review of internal documents, OpenAI’s losses were on course to roughly triple to fourteen billion dollars in 2026, with genuine profitability not expected until the end of the decade, around a hundred billion dollars in revenue, after cumulative burn approaching a hundred and fifteen billion. Only about one in twenty of its weekly users pays anything at all. The company began serving advertisements in February, the traditional recourse of a business that cannot charge enough for the thing itself. Microsoft’s revenue-share disclosures implied that OpenAI was spending something like two dollars in cost for every dollar of inference revenue it collected. That is not a toll booth. That is a toll booth paying drivers to use the road.
Anthropic, its nearest rival, tells a more encouraging story, and honesty requires telling it. Its revenue run rate had reportedly crossed thirty billion dollars, up from nine billion at the close of 2025, with more than a thousand business customers spending over a million dollars a year each, a figure that had doubled in months. Roughly eighty-five percent of its revenue came from enterprises rather than consumers, and it projected positive free cash flow by 2027 or 2028, filing confidentially for a public offering in the middle of 2026. This is a real business with a real path, and any account of a dying rent model has to reckon with the firm that appears to be building a living one.
But even Anthropic’s success sits inside a macro picture that should trouble anyone counting on durable rent. Data-center capital expenditure was projected to approach one trillion dollars in 2026, the first trillion-dollar year in the industry’s history, with Nvidia’s chief executive musing aloud that the figure was heading toward three or four trillion. A striking share of the hyperscalers’ backlog, on the order of a trillion dollars, was reportedly tied to two startups that were, at that moment, both losing money. The industry was building the most expensive infrastructure in the history of private enterprise on the assumption that the intelligence it produced would command a premium, at the precise moment when the price of comparable intelligence was collapsing toward the cost of the electricity required to generate it. You do not need to be a vulture to notice the tension between those two facts. You only need to be able to read a balance sheet.
The case for the moat
An essay that stopped here would be propaganda, and the subject deserves better, because the counterargument is serious and its proponents are not fools. The strongest version runs as follows. The frontier model is not where the durable value lives, and never was; the mistake is to look for the moat in the model at all. The right analogy is web search around 2002, when a hundred commentators declared that search was a commodity and that the various engines were interchangeable, shortly before one of them became one of the most profitable enterprises ever assembled. Value did not accrue to the algorithm. It accrued to the substrate beneath it and the distribution around it. By this reading, foundation models are the loss leaders of the AI economy, and the money will settle one layer down, in the chips and the data centers and the infrastructure, which is why Nvidia was worth north of five trillion dollars while the labs bled, and one layer up, in the software that wraps itself so tightly around a customer’s workflow that leaving becomes unthinkable.
The evidence for stickiness is not trivial. Anthropic’s coding product went from nothing to billions in annual recurring revenue in under a year, not because its underlying model was categorically unmatchable but because it was reliable, integrated, and trusted for work where a wrong answer is expensive. Enterprises do not swap the system their engineers depend on every time a cheaper model tops a leaderboard, any more than a company rips out its accounting software to save a few points on licensing. Switching costs are real. Compliance and safety guarantees are real. The accumulated context that a deployed system builds up about a particular organization, its data, its idioms, its ten thousand small conventions, is a genuine and compounding asset, and it looks a great deal like the “tribal knowledge” Palantir’s manifesto was so eager to name.
The most intellectually honest synthesis is neither triumph nor collapse but bifurcation. The market splits. A durable premium tier persists for the mission-critical work where determinism and accountability outweigh price, where an institution will gladly pay ten or twenty times more to be certain. Beneath it, the vast and rapidly growing background layer of routine machine cognition, the agentic plumbing that runs continuously and cares only about the invoice, commoditizes to open weights and races to the floor. On this reading the labs most exposed are the ones whose revenue leans hardest on the commodity API, the metered stream of undifferentiated tokens, while those insulated by software and enterprise depth ride out the compression. Rent does not die. It retreats to the citadel and abandons the countryside.
What the vulture actually sells
Which returns us, finally, to Palantir, because Palantir’s own commercial design is the single most eloquent piece of evidence in this entire debate, and it points toward the commodity thesis with an eloquence its manifesto never quite matches. Look closely at what the Nvidia deal actually does. Palantir does not build the model. It takes an open-weight model it did not make, Nvidia’s Nemotron, fine-tunes it on the customer’s own data, lets the customer keep the resulting weights, and charges for everything except the model: the ontology that gives the data meaning, the deployment layer that puts it behind the customer’s perimeter, the operational scaffolding that makes the whole thing usable inside a classified environment. Palantir has positioned itself to capture value above the model and below the model and pointedly not in the model. It treats the frontier model the way a razor company treats the razor: as the cheap thing you give away to sell the blades.
A company arranges its business around the parts of the value chain it expects to be durable. When a firm as ruthlessly commercial as Palantir arranges itself to earn nothing from the model and everything from the layers surrounding it, it is placing a very large bet that the model itself is becoming the free part, the razor, the commodity, the thing whose price falls to its cost of production while the margin migrates elsewhere. This is why the manifesto, for all its opportunism, is worth taking seriously as forecast. It is the rare corporate document whose author has staked real money on the truth of its central claim. Karp said the quiet part with unusual candor when he explained the appeal of the arrangement: institutions want to know, he said, that they own the means of production. It is a phrase with a distinguished pedigree, and its reappearance in the mouth of Silicon Valley’s most enthusiastic militarist is one of the stranger ironies of the moment. The Palantir thesis, reduced to its core, is that in the AI economy the means of production is not the model. The means of production is ownership itself, and ownership is precisely what the rentiers cannot sell you, because selling it would end the rent.
So the first of the two great assumptions is under visible strain. The premise that frontier intelligence would be a toll road, that a few American firms would meter the cognition of the world and grow rich on the passage, is being tested against a reality in which the road keeps sprouting free lanes, the tolls barely cover the asphalt, and the cleverest operator in the business has quietly repositioned itself to sell everything but the passage. The rent is not dead. But it is being negotiated down, in public, by the market, and a company whose entire existence is a wager on ownership has begun to circle the companies whose existence is a wager on access.
That is the first assumption. The second is stranger and, for the architecture of American power, more dangerous, because it concerns not the price of the intelligence but the leash attached to it, and the growing suspicion among America’s allies that the leash runs in only one direction. It concerns what happens when a sovereign state discovers that the tap it depends on can be turned off by a hand it does not control, and that the hand has recently shown itself willing. To that discovery, and the summer it stopped being hypothetical, we turn next.
Capable of Turning Off the Tap
Readers of this publication have already watched, in an earlier essay, the American state govern its laboratories by exception. We traced the Friday-evening order that went out from Washington in the middle of June 2026 and left the two most capable models on earth dark for a fortnight, and we asked the domestic question it raised: who rules the laboratory, and by what warrant. That essay looked inward, at the relationship between a state and the private powers it has half-absorbed. This one follows the same act of power outward. The order was issued at home, but it was witnessed abroad, and the capitals that witnessed it were not primarily interested in the constitutional niceties of American administrative discretion. They were interested in a simpler and more alarming fact: the hand that had reached into those servers could reach into theirs, and they had just been shown that it would. Four days after the models went dark, one of those capitals announced what it had concluded.
The capital was Paris, and the voice was that of Sébastien Lecornu, the French prime minister, who on June 16 declared that the Direction générale de la sécurité intérieure, the domestic intelligence service that had run on Palantir’s software since the years after the 2015 Paris attacks, would replace it with the French firm ChapsVision. His reasoning did not hide behind procurement language. France, he said, must use its own models and could not accept new strategic dependencies in the digital sphere. Then he named the thing that had concentrated his government’s mind, the event whose forensic detail this publication has already given its due. France could not depend, he said, on the goodwill of partners who had shown themselves capable, as everyone had just seen in recent days, of cutting off access to Anthropic’s models. He paired the announcement with six hundred and fifty-five million euros for artificial intelligence and a domestic chatbot for a million French civil servants, and the migration itself would run gradually to 2028. Palantir replied, accurately and beside the point, that its contract remained fully in force. That the contract remained in force was exactly Lecornu’s argument. The paperwork was intact and the dependency was total, and neither fact protected France from the hand on the valve.
The oldest question
The prior essay asked who decides the exception, and rested a good deal of weight on the German jurist who framed that question in the last century. This essay asks something blunter and older, the question posed in 1576 by Jean Bodin, who gave the modern idea of sovereignty its first full statement and who has the advantage, for our purposes, of having thought about the problem before anyone had a state to defend in quite the current sense.
Bodin’s sovereignty had one indispensable property. It could not be divided, and it could not be held on loan. It was, in his formula, not limited in power, in charge, or in time. A prince who ruled only so long as some other party permitted it did not possess sovereignty in a reduced or partial degree; he did not possess it at all. What he held was a license, and the true sovereign was whoever could revoke the license. Bodin was writing about kings and estates and the relations between them, but the logic does not care what runs through the network. Lay it over a data center. If a French intelligence service can analyze its own most sensitive material only so long as an American company, bound by an American export regime, chooses not to cut it off, then the sovereignty over that function is not divided between Paris and Washington in some genteel condominium. It sits, entire and indivisible, wherever the power to cut off resides. The contract is the license. The export order is the revocation. And the revocation had just been demonstrated, which is why a prime minister who understood Bodin, whether or not he had read him, moved within the week.
This is the test the whole sovereignty debate reduces to, and it is worth stating plainly because so much marketing exists to obscure it. Do not ask who owns the contract, or who sits on the board, or how many reassuring clauses about partnership the agreement contains. Ask who can turn it off. Whoever can turn it off holds the sovereignty over the thing that can be turned off, and everyone else, however wealthy, however allied, however solemnly assured of their special relationship, holds a license. For two decades the democratic world had been signing licenses in the confident belief that they were something more, and in the middle of June it discovered, in the space of a fortnight, that they were exactly what Bodin said they were.
France was not alone
Germany had begun reaching the same conclusion even before the cutoff sharpened it. Its domestic intelligence service had already chosen ChapsVision over Palantir, and the Bundeswehr went further, excluding the company from its military cloud altogether. The reasoning offered by Vice Admiral Thomas Daum was not the reasoning of a politician performing independence for a domestic audience. It was an architect’s objection, cool and specific: the armed forces would not permit the employees of a private American company to reach German national data, and that was the end of the matter. The shortlist that replaced Palantir held two German firms and, once again, ChapsVision of Paris. Alex Karp, who does not lose gracefully, compared the German posture to a conversation about witchcraft and reminded the world that his software served in Ukraine as an operating system for war. The comparison was vivid and it changed nothing, because Daum’s objection was not superstition. It was an accurate description of where the off-switch lived.
Spain became the third European government in roughly a month to arrive at the verdict, instructing its state-backed companies across communications, defense, and public infrastructure to sign no fresh contracts with Palantir, over the fear that classified material might be exposed, while carving out an exemption for military work already under way. Even Britain, the most reliably Atlanticist government on the continent and the one least inclined to embarrass Washington, found the question forced on it from several directions at once. The National Health Service’s data platform, a contract worth three hundred and thirty million pounds, went under formal review, with a break clause allowing exit in early 2027 openly on the table. The Mayor of London killed a fifty-million-pound arrangement between Palantir and the Metropolitan Police, and the company threatened to sue. A committee of the House of Commons described reliance on Palantir as an unacceptable point of weakness and warned, in a phrase that could have served as an epigraph for the whole season, against vendor lock-in and debilitating dependencies. Four governments, in a single month, using very nearly the same words, had begun to treat a dependency they had lived inside comfortably for years as a wound that needed closing.
The mechanism beneath the intuition
The ministers were reaching, in the imprecise idiom of press conferences, toward something the academic literature had already named with precision, and the name is worth having because a good frame does more than describe a moment. It predicts the next one. The political scientists Henry Farrell and Abraham Newman supplied it in their work on what they called weaponized interdependence, and its central insight is that the networks binding the modern world together are not flat plains across which everyone moves freely. They are shaped like wheels. They have hubs through which a disproportionate share of the traffic is forced to pass, and the state with jurisdiction over a hub acquires two distinct powers over everyone attached to it as a spoke. The first they named the panopticon effect: the hub-state can see what flows through, can watch and gather and know. The second they named the chokepoint effect: the hub-state can deny passage, can seal the artery, can cut a spoke off from the network entirely.
Europe had spent a decade painfully aware of the first of these and comfortably persuaded it was safe from the second. The surveillance revelations of the previous era had taught the continent that its data ran through American systems that American agencies could read, and the panopticon had become a familiar anxiety, the subject of a hundred earnest conferences. The chokepoint felt different, remote, the kind of weapon a hegemon might turn on an enemy but would surely never aim at a friend. This was the comfortable belief the middle of June annihilated. The cutoff was the chokepoint effect materializing against allies, the proof that denial of access was not a sanction reserved for adversaries but an instrument available, on an ordinary Friday, against anyone downstream. And here the frame earns its keep by telling you what happens next, because the chokepoint has a peculiar property. It is most powerful while it remains merely potential. Once a hub-state actually closes the passage, even briefly, even by accident, even against its own commercial interest, every spoke that witnessed it has learned something it cannot unlearn. It has seen the artery seal. A spoke that has seen the artery seal begins, immediately and permanently, to build itself another route. That is not a moral reaction. It is a structural one, as automatic as a body routing blood around a clot, and it is precisely what France and Germany and Spain set about doing within days.
The sharper edge
If the model cutoff gave the theory a demonstration, the matter of Starlink gave it something closer to a premonition, because connectivity sits underneath everything else and its control had already been concentrated, in public and in one pair of hands. In June 2026, SpaceX went public in the largest offering in the history of capitalism, priced near a valuation of one and three-quarter trillion dollars and raising some seventy-five billion, a figure that made the previous record holder, Saudi Aramco’s listing a few years before, look almost quaint at under thirty billion. Starlink alone had produced over eleven billion dollars of revenue the prior year at margins a software company would envy, its constellation approaching ten thousand satellites and its subscribers past ten million across a hundred countries. It was among the most valuable pieces of infrastructure ever assembled, and its founder retained around eighty-five percent of its voting power.
The question the offering could not escape was Lecornu’s question, carried into a domain where dependence runs deeper than the analysis of files. If a French intelligence service concluded it could not rely on a company that merely read its data, what was any government to conclude about a company that owned its connectivity outright? And unlike the AI case, the Starlink case arrived with the demonstration already on the record. In 2023 the company’s own president had confirmed that Starlink was never meant to be weaponized and had restricted its use for certain Ukrainian operations. In the spring of 2025 Musk stated plainly that Ukraine’s entire front line would collapse if he switched the system off, calling it the backbone of the Ukrainian army. Whatever one thinks of the man, the sentence was an exact description of a sovereignty relationship: a nation’s capacity to defend itself running through a switch owned by a private foreigner, and the foreigner narrating the fact aloud.
Fairness requires conceding that the sharpest form of the charge, the picture of a continent’s connectivity hostage to a single temperament, is overdrawn. Operational control over the relevant capability was afterward shifted toward the Pentagon in meaningful respects, ordinary European service continued without interruption, and at least one major European carrier went on partnering with Starlink commercially. The whim is fenced by institutions and by the plain interest of a company that would rather keep selling than switch anyone off. But fenced is not the same as absent, and the governments watching drew the structural lesson rather than the sentimental one. Europe accelerated its hunt for alternatives, weighing Eutelsat and OneWeb, whose terminals cost something like eighteen times the price of Starlink’s, and placing its longer hopes on IRIS², a constellation backed by more than ten billion euros that is not expected to fly in earnest until the decade turns. Italy had already suffered its own rehearsal of this reckoning in early 2025, when the Meloni government’s flirtation with a large Starlink contract for secure communications dissolved into a sovereignty revolt, one opposition figure asking how the prime minister could hand the keys to Italy’s national security to Musk, and the defense minister eventually admitting that the whole thing had come to a standstill. China, for its part, was building not one rival constellation but two, each planned in the tens of thousands of craft. The single planetary network was splitting into an American incumbent, a European aspiration, and a Chinese alternative, which is what sovereignty anxiety looks like once it hardens into orbital hardware.
The weight the argument has to carry
An essay that stopped on that triumphant note would be cheating, because the most important fact in this entire story is the one the celebratory versions omit. Palantir, the company being escorted out of one European intelligence service after another, was at that very moment enjoying the finest financial performance of its existence. Its revenue in the first quarter of 2026 grew eighty-five percent, its American commercial business faster still, its government revenue climbing steadily toward records. The same Britain whose health service was reviewing one Palantir contract had, only months earlier, handed the company a defense contract worth two hundred and forty million pounds with no competitive tender at all, part of a wider partnership that promised more than a billion pounds of Palantir investment and installed London as the firm’s European defense headquarters. The defections were genuine, and they were argued explicitly in the language of sovereignty, and they mattered enormously as signal. But they were concentrated among a particular class of customer, the sovereign and intelligence buyers for whom the chokepoint is not a seminar topic but an operational emergency, and they had not yet made a dent in the commercial whole. The tap can be turned off. That is now settled beyond argument. But the alternatives to the tap are years away and many times more costly, and a government that grasped its dependence in a single clarifying fortnight cannot walk out of it by the following month.
This is the interval that defines the entire moment, the gap between the discovery of a dependence and the capacity to end it. France can announce a migration that concludes in 2028; it cannot conclude one in a week. Europe can finance a satellite constellation; it cannot launch one before the decade runs out. The open-weight model, as the first part of this essay argued, is irrevocable the instant it is downloaded, but the ontology and the deployment layer and the thousand quiet integrations still have to be rebuilt, and while they are being rebuilt the old dependence keeps humming, invisible again until the next time the valve turns. Sovereignty, it emerges, is lost in an afternoon and recovered over years, and the powers that hold the hubs understand this asymmetry perfectly. It is the true source of their composure. They can afford to let the dependents grumble, because grumbling is slow and the hardware is slower, and the license, though everyone now knows it is only a license, remains the only thing plugged in.
All of which sets up the question that governs the final part of this essay, and it is a question about temperament as much as power. A hegemon that understood the asymmetry, that knew how much its chokepoints were worth and how many years its dependents would need to escape them, would guard that advantage the way a wise creditor guards a debtor’s goodwill. It would reassure, it would soothe, it would keep its hand near the valve and its remarks about the valve to itself, and it would let the sheer convenience of dependence do the quiet work of keeping everyone attached. That is what a patient empire would do. It is almost exactly the opposite of what Washington chose to do. At the precise moment its allies were learning, in real time, what their dependence cost, the United States government elected to inform them, in its own official voice and under its own seal, that their wish to build something of their own was not a difficulty to be respected but a delusion to be corrected, a march in perfect formation into the past. To that remarkable decision, and to the self-defeating logic that produced it, we turn last.
Marching in Perfect Formation into the Past
Imagine you are a European minister in the last week of June 2026. Eleven days earlier you watched a decision taken in Washington reach into systems your government depended on and switch them off, allies and adversaries alike, for reasons no one consulted you about. You have spent the intervening days in emergency meetings about how quickly you can build something you can turn on and off yourself. And then, into this atmosphere, arrives an essay from a senior official of the very government that flipped the switch, explaining to you that your desire to build your own is a delusion. Not a difficulty, not a trade-off, a delusion. Your pursuit of sovereignty, the official writes, will deliver you not independence but a kind of synchronized mediocrity, a planet of subscale clones each heroically reconstructing last year’s breakthrough while the breakthrough itself moves on without them. You are, in his memorable phrase, marching in perfect formation into the past.
The essay was real. Jacob Helberg, the United States Under Secretary of State for Economic Affairs and a former adviser to Palantir, published “The Digital Sovereignty Trap” on the State Department’s own channels on June 23. A nation is not digitally sovereign, he argued, because it can reproduce yesterday’s breakthroughs, since half the world can do that. It is sovereign because it can contribute to tomorrow’s, a condition he named innovation sovereignty, and which, in his account, ran only through American technology. While others rebuilt the present, American firms would be inventing the future. There is no polite way to describe the timing except to call it a study in the tone-deafness of the powerful. A government had just given its allies the single most vivid demonstration in years of why they might want their own capability, and eleven days later it told them that wanting their own capability was a march into obsolescence.
The argument worth taking seriously
It would be easy, and lazy, to treat Helberg’s essay as mere arrogance, and the temptation should be resisted, because underneath the condescension sits a genuinely strong argument that deserves its strongest statement. The case is this. Sovereignty pursued as autarky is a trap in the plain economic sense. If France spends billions to build a domestic model that lands a year behind the frontier, it has not purchased independence; it has purchased a worse product at a higher price and called the difference patriotism. The global diffusion of capable models, including open-weight ones, genuinely does let a smaller nation stand on the frontier without reconstructing it from scratch, in the same way that no sensible country smelts its own silicon from sand in the name of self-reliance. A world in which every state insists on its own subscale everything is a world of duplicated cost and diminished capability, and the label “sovereignty” does not make the arithmetic any kinder. Helberg was describing a real failure mode. Nations do sometimes spend enormous sums to reinvent, badly and slowly, a wheel that was available to license, and they do sometimes dress the waste in the language of pride.
The trouble is not that the argument is wrong. The trouble is that it answers a question no one was asking, while pointedly ignoring the question everyone was. An official of the European Union put the objection with lethal economy to Foreign Policy. The critical element missing from Helberg’s piece, this official said, was dependency and trust, the problem of a country using its technological advantage as a weapon by denying or switching off access. That was what was spurring diversification. It was not that Europe doubted American technology was better. It was the willingness to use the power that bothered them. Read the two positions side by side and the tragedy comes into focus. Helberg says: do not waste money reconstructing the present. Europe answers: you have just proven you will weaponize the present, so reconstructing it is not waste, it is insurance. Each is correct on its own terms, and the terms do not meet. The most wounding fact of all is that Helberg’s own government had, eleven days before he published, supplied the definitive refutation of his thesis. The Anthropic cutoff was the demonstration of what sovereignty currently means in application, and no essay about synchronized mediocrity could argue it away, because the ministers had watched it happen.
A government arguing with itself
Behind this collision lies something stranger than a disagreement between Washington and its allies. It is a disagreement within Washington, a government whose several hands were pursuing incompatible theories of the same problem at the same moment, apparently without noticing.
Three doctrines were operating at once. One, associated with the White House’s science and technology office, held to a vision of managed interdependence, in which American technology diffuses to allies under American terms and everyone benefits from the arrangement. A second, Helberg’s, preached innovation sovereignty and warned allies against the autarkic trap, insisting the future ran through American invention and that resisting this was self-defeating. And a third, lodged in the Commerce Department, simply reached out on a Friday evening and demonstrated the chokepoint, cutting off the models to make an export-control point. The first doctrine promises partnership. The second scolds allies for wanting independence from that partnership. The third proves, in a single afternoon, exactly why the independence might be worth having. A dependent power watching this performance does not see a strategy. It sees a government that cannot decide whether it wants to be trusted or feared, and has settled on doing both simultaneously, which produces neither. The reassurance is undone by the scolding, and the scolding is undone by the switch, and the net message received in Paris and Berlin and Madrid is that American technology is indispensable, that America knows it, and that America will say so to your face while keeping its hand near the valve.
The peace of the dominant
The affirmative vision knitting these threads together travels under a revealing name. In December 2025 the administration launched what it called Pax Silica, an economic-security compact that began with seven signatories and expanded, by a summit the following June, to twenty-four countries, spanning critical minerals and energy and semiconductors and artificial-intelligence infrastructure. Helberg described it as an operational document for a new economic security consensus. The White House’s AI Action Plan called the export of American models to allies an imperative, warning that any failure to supply them would simply push nations toward America’s rivals, meaning China. The strategy, in outline, is coherent and even shrewd: bind the democratic world into a single technological bloc anchored in American capability, offer the diffusion of frontier models as the carrot, and let the gravitational pull of the best tools do the work of keeping allies close.
But the name gives away more than its authors may have intended. Every historical pax, from the Roman to the British to the American, has described the same thing: peace under a dominant power, a peace whose defining feature is that the subordinate does not resist. The pax is never a peace among equals. It is the tranquility of an order in which the question of who decides has already been settled, and settled in one direction, so thoroughly that the subordinate parties have stopped asking. Even sympathetic analysts noticed the irony of marketing dependence as peace at the precise moment the dependents had begun, audibly, to resist. A pax announces confidence. The need to announce it announces the opposite.
The case that the leash holds
Intellectual honesty requires the strongest possible statement of the view that all of this ferment amounts to very little, that the sovereignty rebellion is theater, and that the American position is far more durable than the defection narrative suggests. That case is formidable, and anyone predicting the imminent end of American technological hegemony should be made to answer it.
Begin with the sheer reach of the existing lock-in. Pax Silica spans two dozen countries and a supply chain running from the mine to the model. The advanced chips on which every frontier system depends flow through a single American company whose market value exceeds five trillion dollars. The cloud infrastructure of the Western world runs, in overwhelming proportion, through three American providers. Defense integration binds even deeper: the alliance that guarantees European security runs American software on its most sensitive programs, and Palantir’s combat systems function, in Karp’s own boast, as an operating system for war in Ukraine. The alternatives that Europe reaches for remain years from maturity. The satellite constellation meant to free the continent from Starlink is not expected to fly in earnest before the decade’s end, and its terminals cost many times the price of the American service they would replace. The French migration away from Palantir runs to 2028, and 2028 is a long time to remain dependent on a vendor you have publicly renounced.
Above all, the counter-evidence sits in Palantir’s own accounts, which is a delicious complication for anyone building the collapse narrative on Palantir’s manifesto. The company being shown the door across European intelligence services is simultaneously enjoying the best financial performance of its existence, its revenue growing at a pace that decouples entirely from the sovereignty drama, its American government and commercial businesses booming, its British defense contracts expanding even as its British health contract wobbles. Sovereignty rhetoric, in short, has badly outrun sovereign capability. It is one thing to declare independence in a press conference and quite another to smelt the silicon, launch the satellites, train the models, and rebuild the integration layer that a generation of dependence quietly accumulated. On this reading the vulture is circling an animal that has decades of walking left in it, and the entire drama is the sound of dependents complaining about a dependence they have no near-term means of ending. Defection is real at the margin, among the sovereign and intelligence buyers for whom the chokepoint is an operational emergency. Everywhere else, the leash holds.
The blowback the theory predicts
And yet the theory of weaponized interdependence, the very framework that explains the depth of American leverage, contains within it the reason that leverage decays, and this is the pivot on which the whole assessment turns. Farrell and Newman, when they described the chokepoint effect, also warned about its cost. Every use of the chokepoint teaches the dependents that the chokepoint exists. A hub-state that closes the passage even once converts an invisible dependence into a visible one, and visible dependence is the kind that dependents organize to escape. The panopticon and the chokepoint are powerful precisely so long as they remain latent, so long as the spokes forget they are spokes. The moment the hub demonstrates its power, it begins to spend it, because the demonstration is also an education.
This is the structural fact that the durable-lock-in case, for all its force, cannot fully absorb. The lock-in is real, but it was strongest when it was unremarked, and it is being remarked upon now with unprecedented volume, partly by the hegemon itself. Around forty-six countries were funding on the order of a hundred and thirty-five distinct sovereign-AI projects by 2026. China was launching not one but two rival satellite constellations against Starlink’s near-monopoly, each planned in the tens of thousands of craft, fracturing global connectivity into competing blocs. Chinese open-weight models, as the first part of this essay established, were already carrying the majority of the world’s routed token traffic, and an open-weight model, once downloaded, can never be switched off by anyone. Each of these is a lane being built around the toll road, a passage that will not pass through the American hub. None of them matures tomorrow. All of them mature. The question was never whether the lock-in was strong today, which it plainly is. The question is whether the hegemon’s own conduct is accelerating the construction of the alternatives, and the answer, visible in every emergency budget line from Paris to Beijing, is that it is.
Dependency theory gave this dynamic its oldest name. The relationship between a core and its periphery persists, in that tradition, only so long as the periphery accepts the terms as natural, as simply the way the world is arranged. What ends such relationships is not usually a sudden reversal of power but a shift in consciousness, the moment the periphery stops experiencing its subordination as weather and starts experiencing it as a choice that someone else is making, a choice that could therefore be refused. There is a hard-won piece of political wisdom hiding in all of this, and it is the one the summer of 2026 illustrated most sharply. When a more powerful country’s officials begin telling you, in their own voice and under their own seal, that your sovereignty is a problem to be managed rather than a right to be respected, you are being handed the most reliable signal in the historical record that you are about to need it. Helberg meant “The Digital Sovereignty Trap” as a deterrent. Its likeliest effect was to function as a recruiting poster for the very cause it opposed.
Cold snap or collapse
So where does this leave the two assumptions with which we began, the twin pillars of an American economic order: that frontier intelligence would command durable rent, and that dependent allies would accept their dependence without complaint?
The temptation is to declare both dead and be done with it, and the temptation should be refused, because it proves too much and the evidence will not bear it. The rent is not dead. It has retreated toward a defensible premium tier, in the mission-critical work where reliability outweighs price, while surrendering the vast commodity plain to open weights that grow cheaper and nearly as good by the season. The captive allies are not free. They remain bound by chips they cannot make, clouds they cannot replace, and satellites they cannot yet launch, and they will remain bound for years by the brute arithmetic of the interval between discovering a dependence and escaping it. Anyone forecasting the sudden collapse of American technological power in 2026 is forecasting the wrong thing, and the vulture, whatever its instincts, is not circling a corpse.
But something has changed that no recovery of quarterly earnings can reverse, and it is more fundamental than any contract. What is dying is not American supremacy. What is dying is the frictionlessness of it, the settlement in which dependence was permanent, unpriced, and above all unnoticed. Both original assumptions were, at bottom, assumptions about unconsciousness: that the world would keep paying rent without asking whether the thing had become a commodity, and that allies would keep depending without asking whose hand was on the valve. That unconsciousness was the actual asset. It was worth more than any moat, because a dependence no one perceives requires no enforcement and provokes no resistance. And it is precisely that unconsciousness that the events of 2026 destroyed. The buyers have done the arithmetic on the models. The ministers have seen the valve turn. You cannot unsee a valve. Once a dependence has been named out loud, priced, and demonstrated, it becomes a contested and self-conscious dependence, which is a weaker and more expensive thing than the invisible dependence it replaces, even when the underlying power has not shifted by a single chip.
Here the deepest irony of the whole affair comes to rest, and it is an irony about who broke the spell. An adversary spends years and fortunes trying to persuade a hegemon’s allies that their dependence is dangerous, usually in vain, because the allies experience the dependence as comfort and the adversary as propaganda. What no adversary could accomplish, Washington accomplished for itself, in a single summer, at no charge. It switched off the models and proved the chokepoint was real. It launched the largest offering in history for a company whose founder had already said aloud that he held a nation’s front line in his hands. And then it published an essay instructing its allies that wanting to escape all this was a march into the past. A patient empire, understanding the asymmetry between how fast dependence is lost and how slowly it is recovered, would have said nothing, reassured everyone, and kept the valve well hidden. This empire narrated its own leverage in the imperative mood and taunted the dependents for noticing. It broke the unconsciousness that was its greatest possession, and it did so believing it was projecting strength.
That is what the vulture saw, and it is why the cynical bird was, in the end, the honest one. Palantir circled the closed-source labs because it had read, more clearly than they had, that the market would soon pay for ownership rather than access. But the larger carcass, the one the whole flock has begun to sense, is not a company or even an industry. It is a way of holding power that depended on no one looking at it directly. Sometimes it takes a vulture to tell you something is dying, and sometimes the thing that is dying is not the body at all, but the spell that kept everyone from seeing the body was mortal. The spell is broken now. The dependence remains, for a while. But it will never again be invisible, and an empire whose leverage required invisibility has, without quite meaning to, announced the beginning of its own long negotiation with a world that has finally been told, by the empire’s own officials, to look.
not the first time corporate (-America) has pronounced it’s ambiguous status: “It may be dangerous to be America’s enemy, but to be America’s friend is fatal.”
Dependence or trust?