Kai-Fu Lee recently made a public statement: open source models are the optimal path to achieving "AI sovereignty."
"AI sovereignty" needs unpacking first. Its core meaning: a country, an organization, a developer should have autonomous control over their AI capabilities—not dependent on external companies' APIs, not constrained by others' terms of service, able to modify, deploy, and optimize their own models at any time.
Under this definition, open source does indeed look like the optimal solution. But if you only see the equation "open source = freedom," that is too simplistic.
The Visible Value of Open Source
Lee is right that open source has several hard advantages:
Autonomous control. Download the model weights, run them on your own servers, without going through any third party. No fear of API bans, service discontinuation, or data crossing borders.
Auditable. Open source means anyone can inspect the model's behavior—are there backdoors, biases, or security risks? With closed-source models, you can only trust the vendor's promises.
Customizable. Open source models can be fine-tuned for specific domains, adapted to specific hardware, and stripped of unnecessary features. Closed-source models are black boxes—you can only adjust prompts.
These are real benefits. For government agencies, financial institutions, and healthcare industries, these benefits are almost essential.
But What About the Hidden Costs?
Compute threshold. The "free" in open source only refers to usage rights. Running a 70B parameter model requires a GPU cluster worth hundreds of thousands of dollars. This is not something ordinary developers or small and medium enterprises can afford.
Maintenance cost. Downloading model weights does not mean you are done. You need expertise in model deployment, inference optimization, version management, and security hardening. These professionals are not cheap.
Iteration speed. Closed-source vendors have teams of hundreds of researchers continuously optimizing models. The open source community relies on volunteers and limited commercial companies—iteration speed is usually half a step behind. By the time your open source model catches up to the previous generation of closed-source capability, the closed-source vendor has already released the next generation.
Security updates. When a vulnerability is found in a closed-source model, the vendor fixes it directly. When found in an open source model, you fix it yourself—or wait for the community. This time gap can be fatal.
The Overlooked Third Path
Lee's narrative framework is a binary "open source vs. closed source" opposition. But in reality, there is a third path: open source models + commercial services.
Mistral, Qwen, and Llama—all these open source models are backed by commercial companies. You can use the model for free, or pay for deployment support, optimization services, and security updates. This model combines the freedom of open source with the sustainability of commercial services.
This is the true optimal solution for "AI sovereignty"—not pure open source or pure closed source, but choosing the appropriate level of commercial service based on your needs, on top of an open source model.
My View
Lee's general direction is correct. In the narrative of AI sovereignty, open source is indeed a better starting point than closed source. But treating open source as the endpoint underestimates the real difficulty of "using an open source model well."
For China's AI ecosystem, the value of open source models is not just technological autonomy but industrial infrastructure building. The value of open source models like Qwen, GLM, and DeepSeek lies not only in the models' capabilities themselves but in how they lower the barrier for the entire industry to use AI.
But lowering the barrier does not eliminate it. Open source makes AI "usable," but using it well, using it securely, and using it competitively requires not just open source models, but engineering capabilities, talent reserves, and industrial ecosystems.
These cannot be solved by simply downloading a model weight file.
Primary source: