The Careless Birth of an AI Umbrella
An Anxious World Witnesses a New Form of American Power
At 5:21 on a Friday evening in June, the United States discovered a new instrument of national power and promptly turned it against itself. A jailbreak in Anthropic’s powerful Claude Fable model spooked the Trump administration—which had just spent 18 months evangelizing the unfettered export of American technology—into rapidly shuttering its access. For two weeks, the Commerce Department ordered Anthropic to suspend its best-in-class AI models for any foreign national, anywhere in the world.
That was enough time for global publics to absorb a familiar lesson: access to frontier technology is a luxury, not a right. It was the same lesson learned when the United States detonated a nuclear weapon in 1945, and when the Soviet Union launched an artificial satellite in 1957. The unceremonious withdrawal of frontier AI has instilled a frenzy among policy elites in every country—who envy an awesome capability they do not have, and fear what it might mean to go without it.
Granting access to frontier AI is poised to become one of the most potent sources of American power since the dawn of extended nuclear deterrence. The United States now holds a currency that every government on the planet wants—one that China is struggling to mint, and which appreciates with each passing month.
So far, Washington has spent the first half of 2026 dispensing that currency through improvisation and revoking it by tantrum.
A better investment is possible. It requires moving from a posture of denial to one of guarantees.
The End of AI Abundance
The weeks since Anthropic first unveiled Claude Mythos have felt eerily similar to the years that followed 1945 and 1957. Britain, France, and India declared they will build their own sovereign AI programs, while particularly close American allies—Japan and some of the Five Eyes—negotiated privileged access for a handful of their institutions. Smaller states are alternating between balancing and bandwagoning, privately begging for access to the American frontier while publicly aligning with the U.S. vision for AI or demanding that larger states make AI more “inclusive.”
For most of the past three years, the working assumption in allied capitals had been that frontier AI would diffuse like any other software. That assumption seemed reasonable in early 2025, and many free marketeers still cling to the dream of unfettered AI proliferation. It is dead now—killed by a combination of legitimate concerns about AI’s growing capability, the waking leviathan of statist intervention on national security grounds, and the punishing economics of running inference compute at scale.
Critics warn overregulation is poisonous to American innovation and risks pushing the rest of the world toward Chinese alternatives. The reality is that the U.S. government is already erecting a wall around frontier AI—a wall that is here to stay. But Washington has not yet decided who should have access to this special technology, on what merits, granted through what process, or toward what end.
For those who still think it possible to simply wish this wall away, it is worth briefly examining the events of 2026. In March, the Pentagon designated Anthropic a “supply chain risk” after the company refused contract terms permitting domestic surveillance and autonomous weapons. In April, Anthropic launched a rationing system for Claude Mythos. In May, Treasury Secretary Scott Bessent personally granted access to Japan’s three megabanks—the first foreign institutions eligible to preview the American frontier. When European finance ministers demanded the same right at Ecofin, the White House refused. Then in early June, Anthropic unilaterally extended Mythos access to 150 organizations in fifteen countries before making it commercially available to the public under the trade name “Fable.” That lasted for just 48 hours before the White House completely suspended access to Fable, pressed Meta to enter a binding commitment to offer privileged access to its frontier models, and requested a user-by-user review plan for the release of OpenAI’s ChatGPT-5.6. While access to Fable has since been restored, the White House has issued no published standard for what has changed, nor what might trigger the next model’s suspension.
The June 2 Executive Order on “Promoting Advanced Artificial Intelligence Innovation and Security” might have imposed some method on the security-minded madness currently engulfing American AI policy. Instead, it codified the ambiguity by creating a mechanism for identifying “trusted partners” and then—as the lawyers parsing it have noted—providing no public criteria whatsoever for who should qualify.
Each of these decisions may be individually justifiable. Together, they describe a system in which frontier AI access is dispensed principal-to-principal, contested agency by agency, and revocable with no notice. The difference between a partner receiving the most powerful defensive tool on Earth and receiving nothing turns on an inscrutable dime.
As things stand, this is a system no foreign partner can depend on. And so each is in the process of making alternative arrangements.
The Case for an AI Umbrella
The U.S. government has made a decision to leash AI’s diffusion. However costly to industry and unwelcome to champions of American innovation, that decision need not come at the expense of America’s standing in the world. Managed deliberately, custody over a scarce AI frontier may yet be transformed into an enduring source of American power.
The United States is the undisputed leader of an emergent AI club. Frontier AI systems are expensive to build and even more demanding to run at scale. Few economies are capable of standing up end-to-end supply chains to manufacture large quantities of inference compute, and fewer still can match San Francisco’s pull on the world’s top AI talent. Together these assets form a soft moat around America’s lead in building frontier AI systems.
But every technological moat in history has eventually been bridged. The atomic bomb took the Soviets four years to match, the satellite took America one, and China’s open-weight AI models now trail the U.S. frontier by a matter of months.
For America to benefit from the AI umbrella it now unwittingly holds, it must build a geopolitical order that outlasts its temporary technical lead. Britain’s naval supremacy is long gone, survived by the global trading system it once anchored. America’s nuclear monopoly lasted barely half a decade, while the alliance system erected atop it is approaching its ninth.
There are compelling reasons to believe the most economically and militarily consequential technology of this century will be—at least for the next few years—scarce, controlled, and overwhelmingly American. In any other domain we would recognize this condition instantly for what it is: a source of advantage ripe for exploitation. Because it commands dollar clearing, the United States built a sanctions regime that reaches into every bank on the planet. Because it pledged to uphold the security of sea lanes, it secured eighty years of basing rights and alliance cohesion in Europe and Asia.
The task now before Washington is to convert its depreciating technical advantage into appreciating structural power—by engineering a world whose critical infrastructure runs on American rails and answers to American standards.
What the World Wants
Most governments today understand that their access to American AI is conditional, and have made their peace with it. They are not demanding best-in-class capabilities as a birthright. Many states can weather Washington’s intrusive vetting standards or requests for lopsided investment. Indeed, the AI vig is ripe for Washington’s taking—if it can assemble an acceptable term sheet.
What foreign partners cannot abide is American unpredictability. The United States cannot reasonably ask a partner to commit tens of billions of dollars to build data centers, expand grid capacity, and integrate U.S. AI models into its financial system while access to the American frontier can vanish on a Friday evening because two executives had a disagreement that the White House chose to referee by export control.
To construct a durable and prosperous AI umbrella, the United States must transform access to frontier AI from a favor into a guarantee. Favors are cheap to extend and cheap to revoke, and so purchase nothing durable. Guarantees are expensive—they bind the guarantor, and that expense is what makes them valuable enough for other nations to reorganize their affairs around.
Serious students of geopolitics understand that the American alliance system was not built on charity. The countries that accepted the American security umbrella spent decades willfully building a dependence on Washington for their defense postures, industrial bases, and foreign policies—a dependence that now forms the bedrock of America’s global power. In 2026, the question is whether Washington is capable of building a new system to reliably distribute AI’s strategic and economic rewards, or whether it will treat AI access as a favor to be dispensed and withdrawn at whim.
The Fable episode gave the world a preview of AI-as-a-favor. Allied governments learned, as Anton Leicht has put it, that they can be cut off from frontier capability as collateral damage from a purely domestic American dispute. Their access was never the product of any decision in Washington, but suffered from a complete lack of consideration. Governments respond to lessons like that. Some will pour billions into sovereign projects that will quickly prove obsolete, wasting resources that could have strengthened an American-led coalition. Others will diversify toward Chinese open-weight systems that trail the frontier by months but at least come without a cord for the White House to yank. Both of these responses corrode American power—which is why it is in the United States’ interest to provision frontier AI access liberally and predictably, as it has other global public goods.
A Mechanism for Provisioning the American Frontier
What might it look like for America to become the intentional provider of a global AI umbrella? Technological frontiers are notoriously jagged and difficult to define.
The nuclear umbrella was born without a blueprint. Washington’s first instinct in 1946 was to cut off atomic cooperation with everyone—including the British partners who had helped build the bomb—a reflex that promptly spawned competing weapons programs in London and Paris. Truman then invented a new form of American influence by improvisation: the bombers sent to England during the Berlin blockade could not actually carry atomic weapons, and the treaty that founded NATO never mentioned them. A decade passed before formal doctrine hardened into the system scholars would retroactively name “extended deterrence.”
The provisioning of American AI is sure to unfold in a manner no less haphazard. But if it can pause long enough to comprehend its own gravity, Washington may yet lurch toward a posture that resembles Eisenhower’s “Atoms for Peace”—the willful, reliable democratization of frontier AI for legitimate use, in exchange for heading off would-be competitors to its production. Such a system will have four recognizable elements:
1. A formal government-to-government access channel. Scott Bessent’s visit to Tokyo, in which Washington blessed three Japanese banks as eligible to receive early access to Claude Mythos, is the right instinct executed through highly informal plumbing.
If it is to be limited at all, then foreign access to American frontier AI should flow between governments as a matter of design rather than personality. A government-to-government AI partnership can take many forms. In one version, Washington could define some allocation—say, enterprise AI licenses for a specified number of entities; or a fixed amount of AI computing capacity—and a partner government would nominate specific institutions to hold it, be they central banks, grid operators, port authorities, or systemically important firms.
No matter the precise form it might take, the important thing is that a government commands the legal and moral responsibility for the conduct of actors it nominates for frontier AI access. This is where concerns about liability and counterintelligence are anchored. If the United States discovers that a nominated institution presents a security risk, or violates the broader terms of a government-to-government partnership, there is an Embassy to call and a relationship to suspend. This degree of formality, while unappealing to private industry, is preferred by foreign governments—because none relishes a world in which the U.S. Treasury negotiates directly with their domestic banks over their heads.
2. A tiered lattice with published logic. Not every partner will receive unfettered access to the American frontier. It is one thing for Washington to share frontier model access with London or Tokyo, and quite another to consider extending such awesome capability to Cairo or Jakarta. The difficult, highly politicized question then becomes: Who should receive access, to which capabilities, and on what timeframe? A workable mechanism for provisioning AI access has perhaps four rungs:
A) America’s closest security allies could receive frontier AI access approaching that enjoyed by U.S. institutions.
B) Trusted partners might receive privileged access to frontier systems in order to harden their critical infrastructure before capabilities diffuse widely.
C) Broader members of an American-led coalition should be provided some guarantee that their access to standard, commercially-available U.S. AI systems won’t be revoked coercively or arbitrarily.
D) The rest of the world will be served by trailing-edge and open-weight AI systems.
The precise gradation between these rungs matters less than there being some legible and universally applied ladder for aspiring members to climb. The Biden-era “Diffusion Rule” failed not necessarily because partners objected to their placement in a particular access tier, but because the criteria were opaque and applied on a case-by-case basis, rather than against any universal standard. Trump’s June 2 Executive Order repeats this same error by creating a “trusted partner” status with no stated qualifications.
For global publics to depend on American AI, the United States must be willing to publish a set of accessibility criteria, review it systematically, and let partners see for themselves what conduct moves them up the ladder and what moves them down. For most AI have-nots, certainty about the rules will be worth more than generosity within them.
3. A clear price of admission. Access to the frontier is valuable, and the United States is in a position to charge for it. The currency Washington should demand is not cash, but a set of contributions that compound American power and crowd out potential competitors. Compute hosting is the most obvious ask. The United States is struggling to install the computational power produced by its chip design companies, thanks to lengthy permitting fights, delayed interconnection queues, and public backlash against AI data centers. Meanwhile, partners like Singapore and the UAE command spare grid capacity and sovereign capital. A partner that hosts a gigawatt of U.S. inference workloads will have earned its share of America’s AI dividend. This vision of AI as an international megaproject is the logic underpinning Pax Silica, the State Department’s growing “coalition of AI capabilities.” The same logic—the idea that frontier AI is a seller’s market, and the United States can make demands in return for purveying technical capability—motivated the Trump administration’s 2025 technology-sharing partnerships with Saudi Arabia and the UAE.
4. Some guarantee of due process. The value of any guarantee lies in its credibility, and the events of June incinerated a good amount of American credibility in fourteen days. A framework for provisioning frontier AI must specify not only how access is granted but how it may be suspended—with documented evidence, through a defined channel, with notice where possible, and with minimal damage to international commerce.
Government restraint protects Americans as much as allies. The same arbitrary U.S. government machinery that cut off SK Telecom’s Fable access had in spring designated a leading American AI lab as a “supply chain risk” over a contract dispute. American labs, their international partners, and frankly parts of the U.S. government itself are all asking for a set of rules they can plan against.
For its part, Anthropic’s position has been that the U.S. government should indeed hold the power to block genuinely dangerous deployments—but through a process that is transparent, fair, and grounded in technical fact. It says something unflattering about the state of American AI governance that a private company had to articulate this standard because the government would not.
If Washington insists on reserving the right to cut off frontier AI access for anyone, at any time, with no clear process, it risks making the same mistake it at first embarked upon with nuclear weapons—thereby inviting every partner on Earth to make alternative arrangements.
It’s Washington’s Umbrella to Drop or Hold
For eighty years the durable core of American power has been a single trade offered repeatedly in different currencies: security in exchange for alignment. The presence of American warships bought basing rights, extended deterrence bought nonproliferation, and dollar access bought sanctions compliance. Frontier AI is the next currency in that lineage—and it may prove the richest yet, because unlike a carrier strike group, it appreciates.
China cannot make this offer today and will struggle to make it credibly for years, because it cannot produce vast quantities of compute—the substrate of the global intelligence economy—and so cannot host AI inference services at scale. Its open-weight models trail the American frontier, and its previous experiment in technological patronage taught the world through Huawei what dependence on Beijing feels like.
The choice now taking shape is between two futures. In one, access to the American frontier continues to be allocated by ministerial visit and revoked by Friday-evening directive. Meanwhile, America’s allies hedge, foreign moonshots multiply, and China’s free alternatives capture markets abandoned by the United States. In the other, America confidently wields the umbrella it has manifested by accident—by building a durable framework for provisioning frontier AI, with predictable government-to-government partnerships and clear rules of access.
Claiming this source of power starts with answering the question every foreign partner is now asking:
If we buy and build American, will you be there?