Open Weights Debate: US-China AI Restrictions Enter Phase Two

A recurring fallacy in CoreLogic's tech industry reports is that policymakers always assume technology blockades are one-way. But the reality is, when Hugging Face, Meta, Microsoft, Mistral, and Nvidia jointly signed an open letter opposing broad restrictions on open-weight models, the underlying logic of the US AI strategy toward China has already undergone a subtle but critical shift.

This isn't just an "industry appeal," but a concentrated explosion of niche conflicts.

Let's look at the data first. In the 2025 global ranking of open-source large model downloads, Meta's Llama series accounts for 42% of the share, while Chinese open-source models like Alibaba's Qwen and Zhipu's GLM together account for about 18%. But the key variable is this: in the second half of 2025, the weekly active user count for Chinese open-source models on Hugging Face grew by 340% year-over-year, while the growth rate for US open-source models was only 120%. This is a dangerous scissors gap.

If the US implements broad restrictions on open weights, such as requiring all models with over 10 million downloads to obtain export licenses, the ones most severely damaged won't be Chinese developers, but Meta, Nvidia, and Hugging Face themselves. Meta's Llama ecosystem relies on third-party developers for secondary fine-tuning; once restricted, developers will quickly migrate to Chinese or other regional open-source models. This isn't technological substitution, but ecological migration.

Using Porter's Five Forces model to deconstruct this decision dilemma:

• Supplier Bargaining Power: At the GPU and computing power level, Nvidia is a direct beneficiary, but restricting open weights would reduce application demand within its ecosystem, leading to a drop in computing orders. Nvidia signing the joint letter is essentially protecting its downstream market.
• Buyer Bargaining Power: Global developers are the largest buyers. If the US imposes restrictions, they will turn to Chinese open-source models, increasing buyer concentration and raising the bargaining power of Chinese model providers.
• Threat of New Entrants: Open-weight models lower the barrier to AI R&D. Restrictions will only spawn more non-US domestic open-source alternatives, such as Europe's Mistral and China's DeepSeek. The threat actually increases.
• Threat of Substitutes: Closed-source models (like OpenAI, Google) are direct substitutes. But restricting open weights accelerates the monopoly of closed-source models, which is detrimental to the diversity of the US AI ecosystem.
• Industry Competitive Intensity: Internally in the US, competition between Meta and OpenAI is already white-hot. Restricting open weights is equivalent to helping OpenAI eliminate Meta's differentiation advantage.

The core contradiction lies here: The US wants to both curb the rise of Chinese AI and maintain its own global dominance in the open-source ecosystem. These two goals are mutually exclusive regarding open-weight models.

Benchmarking against historical cases: The 2019 US chip ban on Huawei resulted in the birth of China's independent semiconductor supply chain. Now in the AI field, Chinese open-source models already possess L2-level autonomous capabilities—meaning they can iterate independently without relying on US open-source frameworks. If the US tightens restrictions now, China could likely launch the first fully de-Americanized open-source model ecosystem within 12-18 months, at which point the US will lose the right to set open-source standards.

From a strategic consulting perspective, the essence of this joint letter is a stakeholder warning. The letter isn't trying to overturn restrictions, but rather demanding "precise restrictions"—for example, targeting only specific parameter scales (such as models with over 100 billion parameters) or specific uses (such as military applications). But the problem is, the capability boundaries of AI models are fuzzy. A 7-billion-parameter model, after fine-tuning, could completely achieve the performance of a 100-billion-parameter model. Technical definitions will never catch up with policy definitions.

I suggest watching three key variables:

1. Meta's stance: If Meta withdraws Llama's open weights, it will directly cause a split in the US developer community. Meta currently maintains a firm stance, but we need to see how the White House applies pressure.

2. Quality leap of Chinese open-source models: If Qwen 3 or DeepSeek 3 reaches Llama 4 levels in code generation and reasoning capabilities, restrictions become meaningless.

3. Europe's position: Mistral is the representative of European open source. If Europe adopts policies opposite to the US, global AI governance will see a "dual-track system."

Actionable advice for policymakers: Don't make "all-or-nothing" decisions. Focus on controllable risk points—for instance, restrict model weights used for training Chinese military AI systems, but preserve the openness of the developer ecosystem. Meanwhile, the US should accelerate building national-level open-source AI infrastructure, such as providing free computing power and data to developers who comply with US export rules, using positive incentives instead of negative blockades. Otherwise, regarding the ship of open-weight models, the US will kick itself overboard first.

Original link: https://techcrunch.com/2026/07/24/as-us-weighs-response-to-chinese-ai-industry-urges-against-broad-open-weight-restrictions/