China Bans Software by Name. America Can't Ban the Models.

In January, Beijing did something that took a paragraph. It ordered state agencies and domestic companies to stop using cybersecurity software from Palo Alto Networks, Fortinet, and CrowdStrike, and set a goal of replacing every piece of foreign software in state-owned firms by 2027. ^[1] A list of names, a deadline, done. The United States cannot do the same thing to Chinese AI models. Not won't — can't. The 1965 Supreme Court ruling in Lamont v. Postmaster General protects the right to receive foreign information, and legal experts say a blanket ban on Chinese models would collide with it. ^[2] One side has a prohibition mechanism that works in a paragraph. The other side's Constitution blocks the equivalent move. So Washington has been pushing where the law lets it push: into hardware, and this week deeper into the commodity layer. The administration is pressuring Apple to drop plans to buy memory chips from the Chinese firms ChangXin and Yangtze for the iPhone. Commerce Secretary Howard Lutnick made the position plain.

There have to be other solutions to the memory issue, but it’s not great American companies using Chinese memory. — Howard Lutnick

Apple's own chief operating officer noted the components require little customization — the switch is purely practical, which is exactly why the pressure landed. ^[3] Meanwhile the Bureau of Industry and Security is investigating how Chinese AI firms reach restricted Nvidia chips through cloud agreements in foreign data centers, including allegations that Alibaba accessed chips in Malaysia through a Singaporean shell company. ^[4] Enforcement is no longer just about shipping containers; it is reaching into the memory inside the phone and the cloud rental on the other side of the world. The software side is different. Three doors stay open, and none of them is a gap in the fence — they are choices and constraints stacked on top of each other. The legal door is the Constitution. The administrative door is the government's own hand: the administration exempted open-weight models — including Chinese ones — from safety testing, and its AI adviser David Sacks attacked the push for restrictions as regulatory capture.

the leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open-source competition. — David Sacks

National Cyber Director Sean Cairncross argued a regulatory regime would strangle growth and be obsolete within days of being written. ^[5] The people who would have to write the ban are the ones arguing against writing it. Then there is the commercial door, which is the widest. Microsoft, the company that sells OpenAI's models, is exploring a self-hosted version of DeepSeek-V4 to power a cheaper Copilot tier. ^[6] Roughly 80 percent of US AI startups already build on Chinese base models. ^[7] Pinterest runs DeepSeek in its recommendation engine, and its chief technology officer reports the in-house models are 30 percent more accurate than the leading off-the-shelf models. ^[8] Alibaba has pushed Qwen to 3 billion downloads in six months — more than Alphabet and Meta combined — while its profits are expected to fall by half, a company spending down its earnings to own the developer layer. ^[9][10] The two enforcement mechanisms sit side by side. China's is a list of banned names. America's would be a lawsuit it would lose — and the companies it was meant to protect are already walking through the opening.