Summary of Key Points
Recently, Chinese AI models (such as Zhipu GLM-5.2 and Yuezhi Dianmian Kimi K3) have rapidly caught up with top American proprietary models, intensifying the debate over openness versus propriety between China and the United States. American tech companies (NVIDIA, Microsoft, etc.) are jointly promoting "open-weight models" as part of a strategic plan to compete with China and gain control of the global AI ecosystem. The U.S. is adopting a "dual-track strategy": proprietary models are used to maintain technological leadership and generate commercial revenue, while open-weight models aim to attract developers, establish industry standards, and capture global markets. It's important to note that open-weight models are not the same as traditional open-source software; AI governance needs to be approached in a layered manner and cannot be simply categorized as "open" or "closed" in terms of security.
I. Why Has the U.S. Suddenly "Embraced" Open-Weight Models?
It's not a sudden change in policy but a result of necessity and strategic calculation:
1. NVIDIA's Business Logic: NVIDIA makes money by selling GPUs and computing power. The more open the models are, the more users they will attract (small businesses and developers can download and fine-tune them), increasing demand for GPU-based computing resources. Open models essentially serve as advertising for NVIDIA's chips. In 2026, data center revenue accounted for 90% of NVIDIA's total earnings, highlighting this strategy.
2. Pressure from Chinese Competition: Chinese open models (such as Qwen and GLM) have surpassed American models in terms of downloads on platforms like Hugging Face, with a performance gap reduced to just 2.7%. The U.S. is concerned that Chinese models may become the de facto standard in the global open-source ecosystem and is working to address this threat.
3. National Strategic Needs: The U.S. aims to transform its technological advantage into an ecological one by making its models, tools, and standards widely adopted worldwide. This way, it can control the entire AI industry chain (for example, if companies use American open models, they will have to rely on NVIDIA's chips and interfaces).
II. Open-Weight Models Are Not the Same as Traditional Open-Source Software
Many people mistakenly think that open-weight models are the same as open-source software, but they are not:
- Traditional Open-Source Software: Provides the source code, allowing users to understand, modify, and create similar products (e.g., Linux).
- Open-Weight Models: Only provide pre-trained parameters; users can use them for tasks like chatting or writing articles and may fine-tune them, but they do not have access to the underlying data or training methods and cannot recreate the exact same model.
For example, OpenAI’s GPT-4 is proprietary (accessible only through APIs), while NVIDIA’s Nemotron is an open-weight model that can be downloaded locally, though the training details are kept confidential.
III. The U.S.'s "Dual-Track Strategy": Combining Proprietary and Open Models
The U.S. is no longer relying solely on proprietary models; it is adopting a dual approach:
1. Proprietary Track: Models like OpenAI’s GPT series maintain top performance and generate revenue through APIs, while also allowing for better control over security (e.g., by promptly fixing vulnerabilities).
2. Open Track: Models like NVIDIA’s Nemotron and Meta’s Llama are made available to a wider audience, attracting developers to build applications and gaining market share (especially in cost-sensitive industries and developing countries). This approach helps establish standards that make it difficult for others to bypass U.S. dominance.
IV. AI Governance Cannot Be Simplified to a Binary Choice of "Open" or "Closed"
Security is not determined by whether a model is open-source; rather, several factors need to be considered:
1. Model Capability: Highly capable models (capable of conducting cyberattacks or biological design) require strict regulation, regardless of their source.
2. Degree of Openness: Models that only provide APIs (e.g., GPT) limit developer access to core functionality, while open-weight models pose greater risks due to potential for malicious modifications.
3. Deployment Context: Models used locally (e.g., in corporate offices) are generally safer than those connected to industrial systems.
4. Responsibility: It is crucial to determine who develops, deploys, and is responsible for the security of these models. For example, open-model developers should provide security assessments, and deployers must monitor their use.
In summary, the AI competition between China and the U.S. has evolved from a focus on model performance to control of the entire ecosystem. The U.S.’s dual-track strategy aims to maintain its technological lead with proprietary models while using open models to expand its global influence. For China, the next step is to improve model performance and build a more robust open ecosystem that attracts developers worldwide—this is what true ecological dominance entails.