China Is Building a Silicon Curtain Around Its Most Powerful AI Models

Sources indicate the framework would cover both closed-weight and open-weight models, and that the most powerful frontier models could be limited to domestic deployment . The framework treats model weights — the core intellectual property of an AI system — the same way Washington treats advanced semiconductors: as export-controlled items .

A separate draft rule published on May 28, 2026 by the Cyberspace Administration of China (CAC) opened a 60-day consultation period on generative-AI export rules, requiring security reviews for frontier-model exports to designated jurisdictions .

What Triggered the Recalibration

The proximate trigger was Washington's own escalation. On June 12, 2026, the U.S. Commerce Department sent a letter to Anthropic ordering it to halt foreign access to its most advanced models — Fable 5 and Mythos 5 — citing a "narrow potential jailbreak" and national security risks . Anthropic abruptly disabled both models for all users outside the United States . The Foundation for Defense of Democracies noted that China's move "mirror[s] a move by the U.S. Commerce Department in early June to place controls on Anthropic's latest products due to their hacking capabilities" . The controls were lifted on June 30 after Anthropic addressed U.S. safety concerns , but the precedent was set: Washington had demonstrated it would unilaterally restrict foreign access to frontier AI models.

Broader structural factors also drove Beijing's shift. DeepSeek's surprise efficiency breakthrough in January 2025 showed that China could build competitive models despite U.S. chip sanctions, making its own models a more valuable strategic asset worth protecting . Years of U.S. restrictions on Nvidia chip sales to China created a tit-for-tat dynamic — Beijing now applies software-layer controls to match Washington's hardware-layer controls .

Both Sides Now Treat AI as a Guarded Strategic Asset

Washington's posture: The U.S. has maintained a multi-tier chip export framework since 2022–2025, classifying countries into tiers with different access levels to advanced AI semiconductors . The June 2026 Anthropic order marked the first time Washington directly restricted a software product (an AI model) rather than hardware — a major escalation beyond chips . Both the Biden and Trump administrations have consistently framed frontier AI as a "national security asset" subject to export controls, not a freely tradable technology .

Beijing's posture: MOFCOM's July 2026 consultations signal Beijing is building a "software-layer export control regime that mirrors the hardware controls Washington applied to chips, but targets the open-source distribution channel that drove Chinese AI influence" . The proposed tiered framework would treat model weights like Nvidia chips — an explicit acknowledgment that open AI is no longer free . China's earlier open-source strategy (DeepSeek, Qwen) was a deliberate tool for global influence and adoption; the new posture treats further open distribution as a strategic liability .

The bottom line: The era of freely shared frontier AI is effectively over on both sides of the U.S.-China tech divide. Washington controls hardware and now leading U.S. models; Beijing controls its own model weights and is preparing to lock them down. The global open-source AI ecosystem is being split in two .