The Either/Or That Builds Its Own Rival

The State Department is drafting a letter that asks 35 countries to do something diplomacy almost never asks: pick one side and only one. Each nation must choose exclusively between Pax Silica, the American-led AI alliance, and WAICO, China's World AI Cooperation Organization. The trigger was Kazakhstan, which signed both. ^[1] It would be easy to read this as a strategy in trouble. It isn't, on paper. Pax Silica launched in June with 35 signatories, including the EU, Germany, India, the UAE, and Armenia. ^[2] The binary choice is not a failing initiative. It is a working one, and the contradiction sits inside it. The strategy has two pillars. The alliance pillar is the diplomacy: the either/or letter, the pitch of AI sovereignty, the promise of trusted supply chains. The restriction pillar is the wall: chip embargoes, model-access cutoffs, blacklists. The problem is that the second pillar keeps producing the outcomes the first pillar is meant to prevent. Start with the embargo paradox. The point of cutting off advanced chips was to keep China from building them. What it did was make China build them. DeepSeek is raising $7 billion to develop custom inference chips specifically to bypass US export controls. ^[3] Huawei's Ascend 910C and SMIC's 7-nanometer production are direct responses to the restrictions. ^[4] The wall built the adversary's ladder. Then the allies. The restriction pillar does not only target China; it reaches into the countries the alliance pillar is courting. When Washington directed Anthropic to suspend advanced model access in India, India's technology secretary warned that the United States cannot impose abrupt cutoffs. ^[2] The EU, meanwhile, voted to reduce its reliance on non-EU digital infrastructure, with Ursula von der Leyen calling it Europe's independence moment. ^[5] The access weapon pushes partners toward the very distance the diplomacy is trying to close. Then the model layer, where the fortress is not sealed at all. Roughly 80% of US AI startups build on Chinese open-source base models, and Alibaba's Qwen has passed Meta's Llama in cumulative downloads. ^[6] Washington is debating a blanket ban on Chinese models, but it runs into First Amendment problems and the simple fact that open-weight models, once released, cannot be recalled. ^[7] The fortress is sealed at the silicon and open at the software. The sharpest version of the contradiction is domestic. The Pentagon blacklisted Anthropic, its own AI champion, as a supply-chain risk after CEO Dario Amodei refused to remove safety guardrails on autonomous weapons and mass surveillance. Yet the military reportedly kept using Claude for intelligence and targeting in Iran and Venezuela. ^[8] The restriction pillar does not even cohere at home: it restricts the very company whose models it depends on. None of this means the strategy is failing on paper. It is not, and some of the infrastructure is on schedule, whatever the satellite data suggests. ^[9] That is what makes the mechanism pointed rather than comforting. The self-defeat runs inside a strategy that is otherwise working. The symmetry at the end is the cruelest part. China's WAICO is framed as open to all sincere and willing participants. ^[4] The rival is positioned to absorb the countries the restriction pillar pushes away. The binary choice rests on a fortress whose own restrictions are filling the exit corridor. The door Washington is trying to lock is the one its own hand keeps pushing people through.