America's AI Alliance Is a Subscription, Not a Treaty

In February, India joined Pax Silica, the American-led AI trade alliance, and the price of admission was concrete: tariffs on American goods cut from 25% to 18%, and a halt to Russian crude purchases ^[1]. The reward was Google's $15 billion infrastructure commitment. Four months later, following a US government directive, Anthropic told its Indian customers something that undercut the whole bargain ^[2].

The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. — Anthropic

India had paid for the gate; it did not get the key. Sarvam AI, the Indian startup, put the lesson plainly.

For AI users, it is clear that you should not confuse access with ownership, or adoption itself as an advantage. — Sarvam AI

Membership in the framework had converted market access from a negotiated rate into a revocable subscription. A tariff is a number you haggle over once and then it sits in a schedule. Framework membership is a standing condition: the benefit is real, but the terms can change, and Washington decides when. The same tempo played out elsewhere. In July the United States elevated the United Arab Emirates to A5 export status, a tier normally reserved for NATO members, granting license-free access to advanced AI chips for G42 and other Emirati firms ^[3]. Senator Elizabeth Warren alleged the decision was entangled with a private deal: a UAE royal's secret 49% stake in the president's crypto company, World Liberty Financial ^[3]. Taiwan's tariff relief, signed in January, was conditioned on $250 billion in investment commitments in the United States ^[4]. In each case the benefit was genuine, and in each case the advanced access remained something Washington can revoke at any time, for reasons it sets. But the perimeter this builds is split. The United States controls the hardware — chips, data centers, export licenses — and it has been closing every gap in that regime, from remote-access loopholes in third-country data centers to waivers for allied chipmakers ^[5]. The software is another matter. Open-weight Chinese models circulate freely; the White House exempted them from safety testing, with the national cyber director arguing a strict regime would strangle growth and innovation ^[6]. David Sacks, the White House's own AI adviser, called the push to restrict Chinese models "regulatory capture" by OpenAI and Anthropic to eliminate their open-source competition ^[7]. And the US-China Economic and Security Review Commission warned that 80% of American AI startups are built on Chinese open-source base models ^[8]. The tension runs through the administration itself. In May, at the Beijing summit, Trump said he and Xi had talked about possibly working together on "guardrails" ^[9]. The same month, Anthropic published a paper calling Chinese AI leadership among the greatest threats to humanity ^[9]. One hand extends a framework; the other calls the counterpart an existential risk. The gap is where the hedging happens. Saudi Arabia signed an open-source AI pact with China in November even as it sat inside American initiatives ^[10]. Canada and Germany launched a Sovereign Technology Alliance explicitly to reduce dependence on the dominant technology platforms of both the United States and China ^[11]. The UAE and India, having tasted both the reward and the revocation, are keeping a foot in each bloc. A perimeter that controls the hardware but not the software, and that its own architect calls partly "regulatory capture," is not a wall. It is a turnstile. Countries can walk through it in both directions, and the ones who have already been through once are watching which way the gate swings.