In June 2026, a familiar instrument of U.S. economic statecraft crossed a line it had not crossed before. An export-control order no longer targeted only the chips, servers, or physical inputs that make advanced artificial intelligence possible. It targeted access to the model itself. Days after Anthropic released Fable 5 and Mythos 5, Washington instructed the company to suspend access by any foreign national, including foreign-national employees of Anthropic, after officials said they had become aware of a possible method for bypassing safeguards in Fable 5. Anthropic’s response was immediate and disruptive: it disabled both models for all customers in order to comply with the directive (Anthropic 2026; Time 2026; Al Jazeera 2026a). That sequence matters because it unsettles the ordinary political story told about export controls. The United States has long used export restrictions to shape access to strategic technologies. In artificial intelligence, those measures have focused mainly on advanced semiconductors, cloud compute, and the firms that supply them. What made the Anthropic episode different was the object of control. The machinery underneath artificial intelligence still mattered, but the order reached beyond the machinery. It turned the model, delivered through accounts, APIs, contracts, and workplace access, into the controlled item.
The order also sat awkwardly with the friend-enemy distinction that usually gives technology controls their public logic. Its sweep reached foreign nationals inside the United States, users abroad, firms in allied economies, and members of Anthropic’s own technical workforce. The government later eased the restriction in part, allowing Mythos 5 to be redeployed to a limited group of approved U.S. organizations, including their foreign-national employees. Yet the period of uncertainty was enough to make a broader point. This was not a narrowly calibrated denial of capability to a named adversary. It was a nationality-based restriction placed on a global cloud service, with consequences that spread across allies, domestic firms, employees, and customers (Reuters 2026a; Reuters 2026b).
This article treats the Anthropic episode less as a turning point than as a revealing stress test. It asks what happens when a chokepoint becomes so politically charged, and so technically hard to administer, that its use no longer separates adversaries from allies, foreign targets from domestic firms, or security risks from ordinary users. Weaponized interdependence has not lost its explanatory power. The case confirms its basic insight: control over a central node can become geopolitical leverage. What may be changing is the condition under which that leverage works. In the Anthropic case, the chokepoint effect appears to enter a more reflexive phase, where the act of denying access also damages the credibility, predictability, and domestic stability of the hub being defended.
Weaponized Interdependence, Revisited
Henry Farrell and Abraham Newman’s theory of weaponized interdependence remains one of the clearest accounts of power in a networked world. Their central insight is direct. Global economic networks do not spread evenly. As firms, users, and institutions cluster around efficient systems, standards, and infrastructures, some nodes become far more important than others. States with jurisdiction over those nodes can convert market centrality into strategic leverage (Farrell and Newman 2019).
Farrell and Newman identify two main mechanisms. The first is the panopticon effect: the ability to gather strategically valuable information from flows that pass through a central node. The second is the chokepoint effect: the ability to deny others access to the network. In both cases, power does not require territorial occupation or direct command. It flows through ordinary dependence on infrastructures that often appear technical, private, and routine. A payment system, a data-routing architecture, a cloud provider, or a model-serving platform can look like commercial plumbing until a state uses its jurisdiction over that plumbing as leverage.
The theory was never only about the United States, but early examples made clear why Washington mattered so much. Dollar clearing, SWIFT, the internet backbone, and large American technology firms all gave the United States unusual reach over global flows. Farrell and Newman’s later book, Underground Empire, broadened the argument into a history of how America weaponized the world economy. Their claim is not that the United States designed every network for coercive use. It is that private systems built for efficiency and profit produced centralized infrastructures that public power could later inhabit (Farrell and Newman 2023).
That generosity to the original framework is important. The Anthropic case is not interesting because it disproves weaponized interdependence. At first glance, it looks almost textbook: a state with jurisdiction over a central technological node uses that authority to restrict access. The difficulty begins once we ask who is being restricted, how the restriction is carried out, and what kind of response it encourages. The chokepoint is clearly present. The harder question is whether the state can use it without weakening the network position that makes the chokepoint powerful in the first place.
Farrell and Newman themselves have begun to describe a world in which these instruments are no longer monopolized by Washington. In their 2025 Foreign Affairs essay, they argue that the age of economic coercion is becoming more reciprocal, with other actors learning to exploit their own forms of leverage (Farrell and Newman 2025). The Anthropic case extends that problem into a more intimate domain. It is not only that other states may turn chokepoints against the United States. Washington’s own chokepoints can expose allies, firms, workers, and users inside the American-centered system to the costs of American control.
When the Chokepoint Stops Discriminating
The political promise of the chokepoint effect is selectivity. A state identifies an actor whose access it wants to restrict, then uses its control over a central node to impose costs. In practice, selectivity is always imperfect. Sanctions spill over. Compliance departments over-block. Firms de-risk beyond what the law strictly requires. Still, the credibility of the instrument depends on a basic claim: the state can distinguish among those it wants to protect, those it wants to discipline, and those it wants to coerce.
The Anthropic order strained that claim. Reporting by Al Jazeera and Reuters, as well as Anthropic’s own statement, indicates that the directive applied to all foreign nationals, whether inside or outside the United States. The company then disabled the two models for all customers in order to ensure compliance. The order did not simply deny access to Chinese military-linked users, Russian intelligence services, or a list of sanctioned entities. It also reached users in allied countries, foreign workers in the United States, and Anthropic personnel. National security supplied the justification, but the instrument worked through a blunt nationality screen layered onto a global cloud service (Anthropic 2026; Al Jazeera 2026a; Reuters 2026a).
This matters for theory because the case changes the direction of the chokepoint. In the classic image, the controlling state stands at the hub and projects coercion outward. The target is external: an adversary, a rival government, a sanctioned firm, or a foreign bank. Here, the denial moved outward, sideways, and inward at the same time. It disrupted allied access. It interfered with a domestic company’s commercial rollout. It created uncertainty for foreign-national employees. It also made customers ask whether reliance on a U.S.-based frontier model meant reliance not only on an American firm, but on sudden American administrative discretion.
The reaction from allies made that anxiety visible. At the G7, French President Emmanuel Macron criticized the American response in nationalist terms, while European officials warned that AI security should not be handled in a discriminatory way against partners. Canadian Prime Minister Mark Carney drew a broader lesson about over-reliance and diversification (Bloomberg 2026). These were not statements from strategic adversaries. They came from governments that are normally treated as part of the trusted circle in U.S.-led technology governance (Al Jazeera 2026b).
The contrast with the earlier small yard, high fence approach is instructive. That formula promised a narrow perimeter around the most sensitive technologies and a high barrier against adversarial access. It was controversial, but it at least claimed to preserve allied participation while denying specific capabilities to China and Russia. The Anthropic episode suggested a harsher possibility. Once the sensitive object is not a chip but a model accessed through cloud accounts, APIs, enterprise contracts, research partnerships, and multinational workforces, the yard becomes difficult to keep small. The fence may still rise, but it can rise around everyone.
There is also a domestic dimension. Time and Reuters both situated the episode amid friction between Anthropic and the Trump administration over the military use of AI, including Anthropic’s opposition to deployment in fully autonomous weapons systems. That context does not prove that the export-control order was punitive. It does show, however, that the chokepoint was not operating in a cleanly external field of geopolitics. It was entangled with the state’s effort to govern a domestic private actor that controlled a strategically valuable capability (Time 2026; Reuters 2026a).
For this reason, the case should not be filed away as just another example of American coercive reach. It is more precise to call it a limit case. If the usefulness of a chokepoint depends partly on the ability to distinguish friend from foe, then a chokepoint that cannot discriminate begins to test its own condition of possibility. The weapon still works in the narrow sense that access is denied. Yet it may work by making the hub less trusted, less predictable, and less attractive to the very users whose dependence sustains its centrality.
Diffusion as Counter-Chokepoint
The systemic response to a chokepoint is rarely passive acceptance. Farrell and Newman’s work has always left room for workarounds. If actors believe that dependence on a hub exposes them to coercion, they search for alternative routes, even when those routes are less efficient at first. The history of sanctions, payment systems, semiconductor restrictions, and platform governance shows the same pattern. Coercion produces compliance, but it also produces adaptation. In AI, the most important adaptation may not be another closed frontier model. It may be diffusion. Chinese firms, especially Alibaba’s Qwen ecosystem, DeepSeek, Moonshot, MiniMax, Zhipu, and others, have pushed open-weight models into global developer markets with a strategic logic different from the American closed-model frontier.
These models do not need to lead every benchmark in order to become default tools for a large share of ordinary use. They need to be capable enough, cheap enough, adaptable enough, and available enough. OpenRouter traffic is not a census of global AI use. It overrepresents developers, router-mediated inference, and price-sensitive experimentation, while missing large portions of consumer traffic and many enterprise deployments. Precisely for that reason, however, it is useful for seeing where switching behavior and developer defaults are forming. Data Gravity’s analysis of OpenRouter traffic found that Chinese open-weight models, after being marginal in late 2024, accounted for roughly 61 percent of routed tokens by May 2026, with four of the five most-used models on the platform coming from China (Data Gravity 2026).
Other indicators point in the same direction. Hugging Face’s spring 2026 review found that China had surpassed the United States in monthly and overall model downloads, with Chinese models accounting for a plurality of downloads on the platform. The U.S.-China Economic and Security Review Commission’s report Two Loops concluded that China has gone all in on open-source AI, with most major Chinese labs publishing source code or weights and charging far less for high-end products than global competitors. The report also stressed the feedback loop created by this openness: adoption drives iteration, and iteration drives further adoption (Hugging Face 2026; U.S.-China Economic and Security Review Commission 2026).
None of this is technological philanthropy. It is dependence built by a different method. The American model of control is strongest where a foreign user needs continuing access to a U.S.-based service, data center, account, license, or enterprise platform. The Chinese open-weight strategy works differently. It places usable model weights, tooling, and derivatives into the world, lowering the cost of adoption and letting downstream users localize, fine-tune, quantize, and build around them. Dependence then shifts from access to a remote service toward reliance on an ecosystem of weights, documentation, developer communities, cloud credits, hardware adaptations, and update cycles.
Stanford HAI’s work on China’s open-weight ecosystem captures the geopolitical stakes of that shift. The issue is less that Chinese models have improved than what their diffusion may do to global patterns of technology access and reliance. If a government, startup, university, or mid-sized firm worries that a closed American model can be switched off overnight because of a political decision in Washington, an open-weight alternative gains strategic value even when it performs somewhat below the frontier. Availability becomes a geopolitical feature. Reliability no longer means only uptime; it also means political predictability (Stanford HAI 2025).
The result is a kind of counter-chokepoint. It does not mirror the American chokepoint by giving Beijing the same switch over a centralized global service. Instead, it reduces the value of the switch. The more users can move critical workloads to open models that they can host, adapt, and combine with local infrastructure, the less credible denial of access to a U.S. frontier service becomes as a universal instrument. This does not mean the United States loses the AI race, or that China wins it. The United States may continue to define the quality ceiling and capture much of the revenue. Yet the volume layer, where habits, tools, defaults, and developer ecosystems form, can still migrate.
That migration is exactly what aggressive chokepoint use can accelerate. A carefully targeted control may preserve centrality by reassuring allies that the hub is secure and governed. A poorly discriminating control sends a different signal. It tells users that the hub is powerful, but also that access to it is politically revocable. In a world where usable alternatives exist, that signal can move experimentation elsewhere. The strategic irony is sharp: the more Washington treats advanced AI models as assets to be fenced off through broad administrative discretion, the more valuable open, cheap, and politically less exposed alternatives become.
The Question the Anthropic Case Leaves Behind
The Anthropic episode does not require us to abandon weaponized interdependence. It asks us to follow the theory into a less comfortable space. The chokepoint effect still explains why the United States could act. U.S. jurisdiction over an American AI company, combined with the centrality of that company’s models to cyber, enterprise, and research users, gave the government a lever. What the case reveals is that having a lever and using it effectively are different things. The core tension is discrimination. When a chokepoint can no longer separate adversaries from allies, foreign targets from domestic firms, or national-security risks from ordinary commercial users, it begins to look less like a precise instrument and more like a systemic hazard. It may still deny access. It may even be justified in a narrow security sense. But it also teaches others that dependence on the hub is a vulnerability, not merely a convenience.
This is where Farrell and Newman’s more recent concern with a weaponized world economy becomes especially relevant. The United States is no longer operating in a world where its own chokepoints are uncontested, invisible, or costless. Other actors are learning to build around them. Allies are asking whether reliance on American technology can survive sudden political turns. Private firms are being pulled deeper into national security governance. The Anthropic case adds one more twist: the weapon can now turn inward, not necessarily by design, but through the practical difficulty of governing a global service with national categories.
The open question is therefore not whether the chokepoint effect exists. It plainly does. The question is whether a state can overuse, misapply, or broaden a chokepoint in ways that erode the network centrality on which the instrument depends. If the answer is yes, mapping new hubs will not be enough for the next stage of the theory. It will also need to explain when the exercise of hub power begins to produce exit, diffusion, and distrust. The Anthropic ban suggests that the most important limit of weaponized interdependence may not be the inability to coerce. It may be the cost of coercing too broadly.
Further Reading on E-International Relations
