How Washington Built a One-Way AI Valve

In June 2026, the US government forced Anthropic to disable its Fable 5 and Mythos 5 models for all foreign nationals. Overnight, India's two largest IT services firms — TCS and Infosys — lost access to the American AI they had built into their operations. ^[1] The Indian AI startup Sarvam AI distilled the lesson into a single sentence.

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

That is one half of the asymmetry. The United States can reach into any American AI company and sever foreign access with a directive. The power does not run through an export control on software — it runs through the company. Anthropic is a US corporation under US jurisdiction; when Washington tells it to cut off foreign users, it complies. ^[1] The mirror question is the one Washington has no answer for: what tool does it have to do the same thing to a Chinese model? The GAIN AI Act — the sole legislative effort targeting Chinese AI — restricts chip exports to adversary nations. It says nothing about model files, software downloads, or open-weight distribution. ^[2] The FCC has banned imports of foreign-made humanoid robots and power inverters on national security grounds. Hardware, again. ^[3] Even Anthropic, after accusing Chinese firms of industrial-scale IP theft through 24,000 fraudulent accounts and 16 million exchanges, proposed a single remedy.

ongoing efforts to free-ride on the capabilities developed by OpenAI and other US frontier labs. — OpenAI

The US-China Economic and Security Review Commission has now stated the gap explicitly.

US export controls primarily target the digital loop, restricting access to advanced chips used for frontier model training — but are not well suited to addressing the physical loop of deployment-driven data creation and accumulation across China’s manufacturing base. — U.S.-China Economic and Security Review Commission

No existing US regulatory tool addresses model-file distribution or software downloads. The entire apparatus — from legislation to agency rulemaking to the industry's own proposed fixes — is built for a physical supply chain. ^[2][4] Even if a software-level restriction existed, the US industry consensus has already disarmed it. Anthropic CEO Dario Amodei called non-dangerous open-weight models "a public good." ^[5]

Open-weights models that don't have dangerous capabilities are a public good. — Dario Amodei

Meta, Microsoft, Nvidia, and OpenAI have all urged policymakers against restricting open-weight AI. And when Zhipu AI launched GLM-5.2 — a Chinese open-weight model that matches American frontier capability — former Trump AI czar David Sacks responded not with alarm but with a warning against regulation. ^[6]

we cannot afford to do things that slow our companies down. — David Sacks

The one regulatory lever that could close the software channel — restrictions on open-weight model distribution — has been disarmed by the very industry it would protect. China is not merely releasing models. It is pairing them with physical infrastructure and bilateral agreements. At the Ninth Digital China Summit, Beijing showcased AI demonstration centers and data exchange platforms under construction in Vietnam, Indonesia, and Laos, deploying industry-specific models for healthcare and education. ^[7] The open-weight release is the door; the infrastructure is the house. The structural conditions in the developing world make the choice starker. Seventy-one percent of Asian organizations are stuck in the AI "Builder" stage, unable to move pilots into production because they lack the high-density infrastructure and specialized expertise that cloud-based proprietary models require. ^[8] A model that runs locally, on existing hardware, is not a preference — it is the only option that works. Then there is language. Vivek Badrinath of the GSMA put the gap in stark terms. ^[9]

Africa has more than 2,000 languages, yet they make up only 0.02 percent of online content — Vivek Badrinath

American models are optimized for English and do not currently serve low-resource languages. Chinese multi-language open-weight models can fill that gap at low cost. ^[9] The result is a choice that is not really a choice. A government or enterprise in the developing world can adopt an American model — knowing Washington can revoke it at any moment, as India learned in June — or it can adopt a Chinese open-weight model that no government can take away. The US-China Economic and Security Review Commission now reports that approximately 80 percent of US AI startups use Chinese open-source base models. ^[4]

Open model proliferation creates alternative pathways to AI leadership. — U.S.-China Economic and Security Review Commission

That number is not a forecast. It is the measure of a valve that only closes in one direction. The US controls target the physical supply chain — chips, fabrication equipment, the hardware layer — and leave the software layer unaddressed. ^[4] Washington can turn off an American model anywhere in the world. It cannot turn off a Chinese one anywhere. The valve works, but only against itself.