AINEWS
A longtime AI critic argues downloadable Chinese models have closed the gap with U.S. frontier systems, and lays out a policy menu from "do nothing" to a public "CERN for AI."
Friday's U.S. stock selloff, which a string of recent Chinese model releases has driven, has reopened an argument being made since the DeepSeek moment: the United States may not be "winning" the AI race at all, but converging to a tie.
Three open-weight Chinese models have effectively closed the gap to the leading U.S. systems: Moonshot's Kimi K3, Z.ai's GLM-5.2, and an upcoming release from Alibaba's Qwen team. "Open weight" means the trained model parameters are downloadable and runnable by anyone, the practical difference from OpenAI or Anthropic, where the weights stay behind a paid API.
GLM-5.2 shows the convergence is real, not narrative. The model ships under an MIT license with a one-million-token context window, on a new architecture called IndexShare. Z.ai's own benchmarks put the model at 81.0 on Terminal-Bench 2.1, up from 63.5 for the prior GLM-5.1, and within roughly 1% of Anthropic's Opus 4.8 on the FrontierSWE coding benchmark. These are the lab's own numbers, not independent verification, but they are the kind of result that would have read as a stretch eighteen months ago.
The American moat in AI software is not as strong as many had hoped, one analyst wrote, reframing the contest as an industrial systems competition to build and install compute, running through TSMC, Nvidia, and the CHIPS Act's fabs—not a secret algorithm.
Seven options are on the table. "Do nothing" is the baseline. Outlawing open source is futile because the weights are already out. A regulatory moat that decides which models Americans can use is corporate communism. Bailing out the leading labs rewards the same strategy that produced the convergence. Banning Chinese models outright is protectionism, not strategy.
Two options remain. The first is a government buyout of OpenAI and Anthropic at a discount, with the labs converted into national-laboratory-style public infrastructure. The second is a CERN for AI: a publicly funded international research effort, ideally including China, that treats frontier capability as a global public good rather than a private race. A recent Xi Jinping speech at a Chinese AI conference is cited as evidence Beijing might be open to such an arrangement, though the reliability of taking Xi at face value is flagged.
Since 2025, the leading U.S. labs have been argued to be unprofitable at their current cost base, Nvidia exposed if frontier training demand plateaus, and the CHIPS Act has subsidized a fab build-out whose returns depend on a software lead that no longer exists.
The American system benefits when there's free flow of capital, ideas, and people, not when Washington tries to win a zero-sum contest.
The next test is whether the U.S. picks an option, or keeps running the 2025 playbook and watches the convergence deepen.