On Saturday, September 19, Donald Trump announced on his social media platform the creation of an “AI Force,” which he likened to the Space Force of his first term, along with the future appointment of an AI czar. In the same post, he dismissed calls to slow down AI deployment, which had come from lawmakers in both parties and from the frontier labs themselves, as a politically motivated hoax. He said his administration would not obstruct the industry’s growth and argued that the existing criminal and civil justice system is enough to deter abuses. The White House did not clarify whether the new body will be military or civilian.

Three months ago, the same administration ordered Anthropic to cut off access to its Fable 5 and Mythos 5 models for any foreign national, inside or outside the United States, including the company’s own employees without US citizenship. The stated justification was national security, with no public detail. The restriction was lifted only after negotiations that produced new safety classifiers and an agreement with the Commerce Department.

The two episodes reflect a coherent policy. The risk of artificial intelligence is treated as a hoax when it would constrain the growth of the US industry and as a concrete threat when it determines who, outside the United States, gets access to the technology. The same danger is weighed on two different scales, and what decides which scale is used is the nationality of whoever would be affected.

At Home

Domestically, Washington’s policy is one of declared acceleration. Trump promised that the government would nurture the industry rather than obstruct it, predicted that AI could reach 25% of US GDP and claimed that the United States is ahead of China and the rest of the world. He called public opposition to data centers “crazed,” on an issue that has become a midterm campaign theme and divides his own party, where Republican lawmakers echo voters’ concerns about electricity bills and disruption to their neighborhoods.

The enforcement apparatus he announced operates after the harm is done. Relying on existing criminal and civil law shifts the burden of proving the injury onto the injured party, in a field where the information asymmetry between developer and victim is the core problem. No prior obligation to test, document or audit follows from Saturday’s announcement.

The House of Representatives went into recess on September 17 and returns on November 9, absent for the entire decisive week. David Sacks, who stepped down as AI czar in March and now co-chairs the President’s Council of Advisors on Science and Technology, has criticized both the labs’ June documents and their September essay. And a lawsuit filed on September 18 in federal court in California accuses AI companies of violating antitrust law by coordinating the slowdown. That hands the acceleration camp a legal weapon against the camp calling for brakes, even though the plaintiffs say they do not object to companies asking the government to regulate them or to grant an antitrust exemption.

Beyond the Borders

Abroad, the same technology is treated as a strategic asset under control. The June episode proved that remote shutdown exists, works globally and depends on an administrative decision by a single government. It was triggered without public explanation and reversed after weeks of negotiation between the company, the White House and the Commerce Department. Five days later, French President Emmanuel Macron warned, at the G7 lunch in Évian, that the US ability to flip the switch at any moment erodes trust in American companies themselves. The warning came after the button had already been tested.

An entire architecture has been built around that mechanism. Pax Silica brought together 35 countries under commitments to pro-innovation regulation and alignment on critical minerals, energy and semiconductors, with a State Department letter barring dual membership in China’s WAICO. Clubs granting access to advanced models, from Glasswing to the White House’s unpublished framework, decide who gets what and on what terms. India, a Pax Silica signatory, had to ask for assurances that technology already granted would not be switched off. The two doctrines underpinning this exclusion come from opposite political origins and reach the same verdict.

The external design works like a valve, calibrated country by country, and the same Bureau of Industry and Security that halted Fable maintains a permissive policy on exporting cutting-edge semiconductors to aligned countries. Pax Silica even provides for signatories to jointly export the American technology stack, which turns some allies into resellers. The heaviest weight falls on those outside the club, which is the case for most countries in the Global South.

A Dividing Line

Dario Amodei’s September 12 essay made this logic explicit on one specific point. The room to slow down exists, he argues, because the United States holds a lead over China, which is why the essay calls for tougher chip export controls combined with a crackdown on unauthorized distillation.

Distillation is the process by which a developer trains a smaller model on the outputs of one already released, capturing part of its capability at a fraction of the original training cost. For anyone without tens of billions of dollars in compute, it is the economically viable route to building models specialized in their own language, case law, school curricula or clinical protocols. Combined with open weights, it is the most accessible path to capability for middle-income countries.

That path has become more attractive on cost. The per-task cost comparisons currently available favor Chinese open-weight models over their Western counterparts, which turns adopting open weights into a budget decision. The open layer on which much of Washington’s diffusion strategy rests was produced by Chinese state industrial policy, a fact American analysts acknowledge without drawing the consequences.

The contradiction emerges when the whole package is considered. The US federal framework exempted open weights from obligations, the result of a campaign organized since February by the companies themselves, while a frontier lab calls for a global crackdown on the very mechanism that makes those weights useful outside the core. Free circulation for the artifact, control over who can turn it into capability. Restricting distillation closes the cheapest door still open and turns a slowdown proposed in the name of humanity into a barrier to entry paid for by latecomers.

The Missing Floor

Since June, the frontier labs have asked for two floors of the same building. One is a mandatory third-party testing regime with legal authority to block releases. The other is an industry-funded operational body under state supervision. The disagreement between them was never resolved and centered on how much of the state to let into the room.

The AI Force answers that question in a way none of them requested. It could establish a state presence without prior obligations, rejects the brake and reserves for the White House the power to police conduct under US law applied to infrastructure used around the world. The labs asked for a brake and may end up with reins.

A Decisive Week

The coming week concentrates the decision. On Wednesday, the US government is convening a high-level AI event on the sidelines of the UN General Assembly, with several countries invited, on the same day the Security Council meets on AI and security, with OpenAI’s Sam Altman expected to brief members. On Thursday, Xi Jinping arrives in Washington for the first state visit by a Chinese president to the US capital in more than a decade, with advanced chips on the agenda alongside Taiwan and the future of the Busan trade truce. The White House confirmed that AI safety is on the agenda. That evening, Altman, Nvidia’s Jensen Huang and Google’s Sundar Pichai will attend the dinner in the visitor’s honor.

If the US lead over China is traded away on Thursday, the premise underpinning Amodei’s entire braking architecture disappears before it is ever tested. If restrictions are tightened, the same architecture gains political backing. In either scenario, the pace of the AI that Brazil and other countries outside these rooms will import will have been set in a room with no Global South representation.

What Needs No Invitation

Any country with administrative capacity can adopt three instruments, detailed in the conclusion of my series on frontier labs (in Portuguese). The first is a regulatory and contractual ban on the unilateral interruption of access to services already contracted by administrative act of a foreign government, with continuity clauses and contingency plans enforceable in public procurement. The second is a requirement to run models locally in health, finance, defense and public security, where sensitive data cannot depend on infrastructure subject to external decisions. The third is independent auditing of the capability claims of frontier models sold within the national territory, backed by in-house technical expertise and incident response.

Implementation has a known benchmark, and it comes from Spain. Madrid cut Palantir out of Telefónica, Indra, Navantia and Correos, and killed both a nearly finalized deal with the military shipbuilder and a project negotiated with the Guardia Civil that the Interior Minister vetoed. All of this came with a €115 million investment in the Catalan company Openchip, part of a €5 billion gigafactory project financed mostly by the state holding company. The investment is what makes the veto hold. The case also teaches through its limits, since the contract that survived intact was precisely the €16.5 million deal with the Armed Forces Intelligence Center, signed in 2023, whose renewal the Army and Navy chiefs are pushing for. In Spain, the civilian periphery proved severable while the defense core held out. That core is where dependence takes root.

New Arrangement

The comparison Trump chose has institutional consequences. The Space Force was created in 2019 as a branch of the armed forces, with its own budget, chain of command and operational domain. Applying that template to AI means treating the layer that answers students, hospitals, courts and public agencies in dozens of countries as a theater of operations. As of this writing, the White House had not clarified the point.

The ambiguity reflects an ongoing dispute within the administration itself. The Commerce Department imposed export controls on an American company’s models a few days after their release, while an executive order issued the previous month to keep regulatory requirements from slowing the sector was already in force. The Pentagon designated Anthropic a supply chain risk, a category historically reserved for foreign adversaries, was sued by the company in March and saw the designation struck down in court, while the National Security Agency kept using the Mythos model. On May 1, eight major tech companies signed agreements to operate on classified networks under the “all lawful purposes” clause, with the Defense Department positioned to write the general AI procurement rules for the entire federal government and $54 billion set aside in the 2026 budget for AI and autonomy. On the other side of the same administration are the venture capitalists and lawmakers who treat any prior obligation as ceding ground to China.

The AI Force could land on either side, and the consequences for the rest of the world diverge. A civilian law enforcement body preserves the current asymmetry, with US rules reaching global infrastructure through contracts. A military branch turns the same infrastructure into a command domain, with its own doctrine and chain of command. That would make the line between commercial service and strategic asset a case-by-case decision in Washington. Countries that run public health, tax collection and defense on foreign models would come to depend on an arrangement whose legal nature has not even been announced.

The double standard holds only as long as the countries receiving the product accept someone else’s scales. Defining in law what is acceptable in the AI operating within their own territory, with enforceable obligations and in-house expertise to verify them, is the one response that requires no seat at anyone’s table. Until that is done, what applies on Brazilian soil will continue to be decided in Washington.