I. Introduction

President Trump posted those words on 14 September 2026, two days after the heads of Anthropic, OpenAI, and xAI called for the frontier to be slowed, adding that his administration already holds “tremendous CRIMINAL and REGULATORY power” over AI companies. The rhetoric tracks written policy, since the White House’s AI Action Plan declares that the United States (“US”) “is in a race to achieve global dominance in artificial intelligence”, treating frontier Artificial Intelligence (“AI”) less as a product to regulate than as an instrument of national power.

In an earlier piece, I examined the June directive suspending foreign nationals’ access to Anthropic’s Fable 5 and Mythos 5 models, which Anthropic switched off worldwide until the controls on Fable 5 were lifted on 30 June. I argued that export control law was the wrong instrument for a safety problem, because it sorts users by nationality while the risk lies in the model, and that both its authority and its procedure were doubtful. That piece asked whether the tool fitted the problem, whereas this one asks what the episode revealed about who holds power over frontier intelligence.

This piece argues that the US does not monopolise AI but holds jurisdiction over several chokepoints through which frontier capability is built and supplied, and that this creates a dilemma, because Washington wants the world to run on the American AI stack while each visible use of that dependence as leverage gives other states a reason to reduce it. The legal defects of June belong to the same problem, since an unpublished, standardless access power is what foreign governments find hardest to accept.

II. From Instrument to Power

Alex Zhao’s analysis in the Harvard Law Review Blog sharpens the doctrinal half of my earlier argument and supplies the bridge to this one. He shows that the confidential “is informed” letter rested on § 744.22(b) of the Export Administration Regulations (“EAR”), which targets military-intelligence end uses in designated countries such as China, yet purported to reach every foreign national. He also doubts that API access is a “release” of technology, because a user “never sees the model’s weights, its architecture, or its source code”, and because the deemed-export rule disregards object code, from which source code could only laboriously be recovered.

Our routes differ, since my piece relied on the published-information exclusion and the major questions doctrine while Zhao works from the definition of “release”, but they converge on the directive’s fragile footing, which Legion LegalTech v. United States, filed in Washington on 23 June, now tests.

Both analyses, however, examine the directive from inside American law, whereas from Paris, Delhi, or Brussels the same fragility means that access can vanish through a confidential letter and a contested reading of a regulation, at a moment nobody abroad can predict. What looked in Washington like a problem of instrument looks from outside like a problem of sovereignty.

III. Primacy, Not Monopoly

None of this means that the US monopolises AI, since the 2026 AI Index finds only a 2.7 per cent gap between the leading American and Chinese models, Chinese laboratories lead the open-weight field, and the hardware chain runs through Taiwan’s TSMC, the Netherlands’ ASML, and South Korean memory makers. What the US holds is concentration at critical nodes, with Nvidia’s chips supplying over 60 per cent of global AI compute, Epoch AI placing about three-quarters of GPU cluster performance in the US, and three American firms holding roughly 70 per cent of Europe’s cloud market.

Farrell and Newman show how jurisdiction over such hubs lets a state “weaponize” interdependence, and the foreign direct product rules carry that jurisdiction to chips made abroad with American technology, so the accurate description is chokepoint power rather than monopoly.

IV. Sovereignty Beyond Regulation

Much of the debate about “AI sovereignty” runs together capacities that should be kept apart, because regulatory sovereignty, the power to set the rules AI must obey within a jurisdiction, differs from operational autonomy, the ability to keep using strategically important AI without a foreign veto, and both differ from denial power, the ability to cut off another’s access, which is power over other states’ autonomy.

My earlier piece held up the European Union’s (“EU”) AI Act as a better-designed safety regime, and procedurally it is, yet in June it could not keep Fable 5 available to a single European user. Since a Brookings study finds full-stack sovereignty “structurally infeasible for almost any country”, the practical test of sovereignty is substitutability, the capacity to switch providers when access is withdrawn.

V. The Hegemon’s Dilemma

The phrase, borrowed from Arthur Stein’s 1984 account of trade hegemons that “need followers”, fits American AI policy, which pursues two goals at once. Executive Order 14320 promotes exports of “full-stack” American AI, and the Action Plan warns that unmet demand would lead “these countries to turn to our rivals”, while chip controls and now model-access controls restrict the same stack. The order also aims to “decrease international dependence on AI technologies developed by our adversaries”, a logic other capitals are free to apply to America.

This resolves an apparent tension with my earlier piece, which called the directive “futile as strategic denial” because rival models offered the same capability. That remains true against adversaries, who have substitutes, but for dependent partners, whose hospitals, firms, and researchers lost the leading model overnight, the directive was anything but futile. Chokepoint power is weakest against the states at which it is aimed and strongest against those who comply with it.

Hostile intent was not needed to provoke that reaction, for even if Théodore Christakis is right that the directive was a security measure rather than a European “kill switch”, since Americans lost access too, French presidential candidate Bruno Retailleau warned that a dependent nation “can be unplugged overnight”, Mohandas Pai urged a sovereign AI fund for India, and Brussels’ tech sovereignty package of 3 June suddenly looked prescient. As Farrell and Newman predicted, exploiting privilege invites efforts to “undermine” the system, and Pablo Chavez observes that “every visible cutoff and every compelled disclosure gives partners one more reason to replace American technology”.

VI. The Strongest Objections

The most serious objection is that chip controls work against China, which the dilemma accepts, since its cost falls on partners. Escape is also expensive, and even India relies heavily on American cloud providers, which is why this piece confines its claim to cheaper hedges such as open-weight models, inference, and cloud. Allies may prefer American dependence to Chinese, as the Gulf firms G42 and Humain showed in accepting “rigorous security and reporting requirements” for American chips, yet a dependence that persists can still grow costlier. Security risks may also justify coordinated denial, as my earlier piece accepted, but the difficulty lies in discretion rather than control, and ECRA itself demands an export control system that is “transparent, predictable, and timely”.

VII. When Sovereignties Collide

The EU asserts its own claim through its market, since Article 2(1)(a) of the AI Act reaches providers of general-purpose models placed on the EU market wherever they are established, and since 2 August 2026 the Commission can request access to a model, “including source code”, under Article 92(3) and fine a refusal at up to 3 per cent of worldwide turnover under Article 101.

Suppose, hypothetically, that Washington again requires a licence before any foreign person may access a model while Brussels demands access for its own evaluators, leaving the developer facing a licence requirement in one jurisdiction and a fine in the other. If Zhao is right that API access is no release, the conflict dissolves at the API but not for source code, which the EAR plainly treats as releasable. No such clash has yet occurred, since in June the models were withdrawn everywhere, but the structure pits a sovereign claiming the developer against one claiming the market.

Data law eased a similar clash through the comity procedure of the CLOUD Act, but frontier AI has no equivalent, and international law does not fill the gap, since the Delhi Declaration is non-binding and the Council of Europe’s AI Convention exempts national security activities, where export controls operate.

VIII. Conclusion

The claim that “whoever wins AI, wins” treats victory as a moment, whereas hegemony is a relationship that lasts only while others find dependence tolerable. My earlier piece argued that safety problems need law built for safety, and the same holds between states, where chokepoint power used at will teaches customers to build exits. The answer is to bind that power rather than abandon it, through published criteria for restricting models, allied consultation before any cutoff, continuity arrangements for critical users, and a comity mechanism for clashes with foreign regulators.

In Ovid’s telling, Daedalus, imprisoned on Crete, reasoned that “Minos rules everything but he does not rule the heavens”, and we remember Icarus as a warning about ambition while forgetting why the wings were built.

Manraj Singh Chandpuri is a third-year B.A. LL.B. student at the University Institute of Legal Studies, Panjab University