Washington strikes first, Beijing follows suit! The AI "Iron Curtain" is truly descending

Yesterday (July 7), a Reuters report caught many off guard—over the past month, China's Ministry of Commerce has convened top players like Alibaba and Zhipu, discussing restrictions on foreign users accessing China’s most advanced large models. Closed-source models are being controlled, and even those with open weights are now under scrutiny. The potential leakage of AI technology might soon be embedded within the framework of China’s National Security Law, extending oversight to foreign investment in domestic AI startups.

This isn’t speculation—it’s sourced directly from insiders by Reuters.

The U.S. was the first to act

Don’t rush to say “it’s always us closing ranks.” Let’s clarify the timeline: Washington fired the first shot in June.

In June, the U.S. Department of Commerce’s Bureau of Industry and Security (BIS) sent a letter to Anthropic citing the Export Administration Regulations (EAR)—arguing that cutting-edge large models could be misused by adversaries for military intelligence purposes, thus requiring control. As a result, Claude Opus 4.5 (“Fable 5” in informal references) and the specialized Opus 4.1 for cybersecurity training (“Myth 5”) now face restricted access for foreign employees working in the U.S. “Myth 5” remains accessible only to a select few American “insiders,” while Opus 4.5 regained partial export capability only after adding safeguards. This has been reported by Reuters and CNBC—no fabrication.

Thus, the sequence “Washington strikes first, Beijing follows” holds firm. But what Beijing is doing this time goes further than many expect—controlling even “open-weight” models shows that AI is being treated as seriously as strategic goods like chips, HBM, or lithography machines.

Why this can't be dismissed lightly

Three real-world impacts affect ordinary people too:

First, global developers’ budgets will shrink. Consider pricing: GLM-5.2 charges $1.4 per million tokens input and $4.4 output; Claude Opus 4.8 charges $5 and $25 respectively—five to six times higher. Qwen3, DeepSeek R1, and GLM were previously the leading open-source, low-cost champions globally. Now, with restrictions, small teams across Southeast Asia, the Middle East, and Europe will feel the pain first.

Second, the golden age of running 70B-parameter models on laptops may be ending. Reddit’s LocalLLaMA forum has exploded recently—European universities, Latin American developers, and Middle Eastern startups have long relied on China’s open-weight models to survive. If open weights are retroactively regulated, this entire wave of Global South AI innovation could crash again.

Third, two distinct API ecosystems are genuinely taking shape. The American system is expensive, closed-source, and selective; the Chinese one is cheaper—but future models may become inaccessible overseas. Countries like Japan, South Korea, the EU, and the Global South will likely be forced to choose between them—or pay more.

Finally, a candid truth

The term “AI Iron Curtain” sounds dramatic, but reality is far more complex. The U.S. restricts us out of fear that our model weights could be stolen and used to attack U.S. systems. China restricts exports out of concern that adversaries could use our models combined with vulnerability databases to undermine our infrastructure (Zhou Hongyi’s recent warning wasn’t empty). Both sides fear their own tools being weaponized against them—motivations aren’t symmetric, but actions mirror each other.

There’s still a loophole not fully sealed: Reuters reports that restrictions might apply only to future new models. If this gap remains, models like Qwen3 and DeepSeek R1—already released—are safe in the short term. But if retroactive controls kick in, that would mark a true turning point.

Don’t talk about “isolationism”—we’re not there yet. “Mutual regulation + self-contained ecosystems” is a far more accurate description.

As for that promotional line claiming “China’s AI has moved from catching up to leading”—just listen, don’t believe it. Our foundational computing power is still constrained. We may lead in cost-effectiveness and open-source ecosystems, but full-stack dominance is still far off. Keep that clarity.

References: Reuters report dated July 7, 2026; CNBC/BIS letter to Anthropic coverage

Original article: toutiao.com/article/1870115994179596/

Disclaimer: The views expressed in this article are solely those of the author