Open vs. closed AI models

Kai-Fu Lee, Founder & CEO | 01.AI | Chairman | Sinovation Ventures

Dr. Kai-Fu Lee is the founder and CEO of 01.AI, a Beijing-based artificial intelligence company focused on large language model development and enterprise AI transformation, venture-built by his tech venture capital firm Sinovation Ventures. He previously led Google’s operations in China and held senior roles at Microsoft. Dr. Lee is the author of the New York Times and Wall Street Journal bestseller AI Superpowers, which examines the role of the US and China in the future of artificial intelligence, as well as its impact on society. Dr. Lee trained as a computer scientist in the United States and has spent his career at the intersection of American and Chinese technology development.

Understanding AI sovereignty

You have navigated the intersection of technology, culture, and geopolitics across the US and China for decades. How do you define AI sovereignty today, and why does it matter?

Sovereignty means a number of things. First, it means having control over the technologies themselves. Second, it means ensuring that data, which is a critical asset to every company, does not leak. That concern is widespread across most countries, but perhaps less so in the US, given its more advanced acceptance of private cloud. But throughout the world, companies are deeply concerned about confidential data becoming an input that enriches a general model owned by another country.

There is also the question of suitability. A model trained in one country tends to embed the values of that country, and those values may not be applicable elsewhere. Having control over the model allows a company to fine-tune it, to make it fit the organisation. Hosting the model in someone else’s cloud limits that ability significantly.

Throughout the world, companies are deeply concerned about confidential data becoming an input that enriches a general model owned by another country.

When Beijing, Brussels, and Washington each talk about AI sovereignty, are they talking about the same thing?

Not entirely. There are layers. The first is suitability: every country wants a model that reflects its language, culture, and legal norms. In some Islamic countries, questions around same-sex marriage, alcohol, or religious guidance are handled very differently than they would be in the US. Ensuring compliance with local law, religion, and culture is a legitimate and universal concern.

The second layer is competitive ambition. Some countries feel they need to build their own model in order to have a chance at a meaningful position in this technology race. China clearly has that ambition. Japan, Singapore, Saudi Arabia, and India have all declared theirs. That is admirable. But it is sometimes overly ambitious. Training a frontier general-purpose model from scratch is not feasible for every country. The US has enormous resources. China has substantial resources. Most others do not.

The practical alternative, and I think the right one for almost all countries outside the US and China, is to take a leading open-source model and continue training it for the country’s specific language, values, and regulations. Continuing training is different from fine-tuning. It treats an open-source model as a half-baked product. Think of it as buying a frozen pizza and then adding your own ingredients and baking it further per your preference. The result is something adapted to your context, whether an Indian pizza or a Japanese pizza. And, critically, the cost is only a few percent of training from scratch: a few million dollars rather than a few hundred million.

The practical alternative, and I think the right one for almost all countries outside the US and China, is to take a leading open-source model and continue training it for the country’s specific language, values, and regulations.

Open-source models and the business of AI dominance

What has been the actual effect of US export controls on China’s AI development?

It is a double-edged sword, and the jury is still out. The GPU export restrictions are only one side of the equation. The other side is that Chinese companies are extremely practical. If you survey them, the great majority are not trying to beat Anthropic or OpenAI. The Chinese companies know they cannot spend fifty billion dollars on supersized model training. So, with the GPUs available to them, some domestically produced and not as scalable, they ask: what can we do excellently within these constraints?

The answer has been rigorous engineering efficiency. DeepSeek is the most well-known example. When people are constrained but motivated, they innovate around the constraint. They optimise architectures, move away from CUDA[1] dependencies, and target performance that is very good rather than definitively the best. Chinese companies are spending less than ten percent of what the top American companies spend on training, yet are producing models that are ninety to ninety-five percent as capable, running six to nine months behind their American counterparts.

There is also a structural reason Chinese companies have embraced open source. They cannot win on a closed-source basis, so they share. It functions like a study group; not an explicit collaboration where people sit down together, but a dynamic where everyone builds on each other’s public releases and the collective competency rises. The American companies behave more like individual researchers, each convinced they will win the Nobel Prise so keeping their know-how secretive. The Chinese companies behave like a group that knows it will not win the Nobel Prise individually, but can get closer collectively by pooling their work.

The deeper strategic question is whether the first company to reach AGI [artificial general intelligence] will dominate everything, or whether four American companies and two Chinese companies will all get there within months of one another, each serving a distinct ecosystem. If you look at the historical trajectory, open-source models have trailed closed-source models by six to nine months. Chinese models have trailed American models by about the same margin. If that pattern holds, then the race may not end with one winner squashing everyone else. But that remains to be seen.

Chinese companies are spending less than ten percent of what the top American companies spend on training, yet are producing models that are ninety to ninety-five percent as capable, running six to nine months behind their American counterparts.

What will ultimately determine the dominant model? Does being late to AI or AI sovereignty necessarily put countries or companies at a disadvantage?

The “AGI-squashes-everyone” idea is unlikely, though not impossible. If American companies end up becoming dominant, my bet is that it will be because of the business ecosystems they build – not because they crossed a technological threshold that rendered everyone else irrelevant.

The first analogy is Windows. Windows did not win because no other company could have built a similar operating system. It won because applications were built on it, enterprises relied on it, and the revenue this generated funded further development, solidifying the position. That virtuous cycle is real and powerful. You see it already with something like Project Glasswing from Anthropic, which only offers exclusive access to top American tech firms to the powerful Mythos model release.  Claude Code is also transforming how software is written. Companies will pay significantly for that. The revenue funds more compute, which funds better models.

Being late to AI or to AI sovereignty does not automatically put countries or companies at a disadvantage. In fact, open‑source models create a meaningful late‑mover advantage.

The second analogy is iPhone versus Android. The tension emerges when AI reaches the mass market globally. American models are priced for wealthy markets. When you run advanced tools on closed infrastructure, the costs are significant. Chinese models, by contrast, can cost a sixth to a tenth as much. So, when the technology reaches populations in China, India, and other emerging markets, the American closed‑source model may simply be priced out of much of that demand.

Open source may win on reach and adoption, while closed source wins on margin and ecosystem control.

The outcome may look very similar to iPhone versus Android. Apple captures extraordinary profit, perhaps more than all Android manufacturers combined, but Android has a far larger global footprint in user base. Open source may win on reach and adoption, while closed source wins on margin and ecosystem control. In that world, being late does not mean being irrelevant. It means choosing the right path based on your resources, objectives, and markets.

Should companies and countries be worried that a small number of platforms will dominate everything?

I do not think we will end up with just one choice, and that is largely because open source is now unstoppable. History suggests we will see one or two dominant closed-source platforms alongside a large-footprint open-source ecosystem. Linux versus Windows. iPhone versus Android. I predict that AI will follow a similar pattern.

What is important to recognise is that this dynamic may play out differently across layers. The model layer is one competition. The agent layer is another. Open ecosystems like OpenClaw are becoming a platform of choice at the agent level, and they fundamentally change the economics of access and adoption.

OpenClaw, combined with open‑source models, allows companies and developers to build powerful AI‑driven workflows without being locked into a single proprietary stack. That is especially important for countries, companies, and developers that cannot afford closed platforms at scale.

I think the likely outcome is not full concentration, but coexistence: one or more closed ecosystems with strong business models, and a large open ecosystem with much wider global reach. The open approach may be less profitable. However, it will be more accessible, making it very difficult to stop. Concentration becomes dangerous only if there is a single choice. Open ecosystems like OpenClaw ensure such a monopoly does not happen.

Open ecosystems like OpenClaw are becoming a platform of choice at the agent level

Physical AI and China’s emerging edge

You have argued that China has structural advantages in physical AI, devices, and manufacturing. How durable are those advantages?

Hardware is a completely different game from software. When the iPhone launched, American companies held nearly every advantage: design, software engineering, market reach. China’s role was centred on manufacturing. Foxconn made money, but only a fraction of what Apple made.

That is no longer the situation. China now has the supply chain and ecosystem to build virtually any hardware product. China also builds at substantially lower cost than the US – from phones and PCs to robots, manufacturing devices, and new form factors such as glasses, watches, and wearable recorders. Huawei, Xiaomi, and ByteDance are all building devices. The barrier that once protected American hardware dominance no longer exists in the same form.

There is a related data point worth making. We are entering an era where AI will increasingly be speech-driven. The power of LLMs is that we can interact through language, and speech recognition is approaching perfection. Devices that record ambient interactions, workplace conversations, and meetings generate a fundamentally new category of data. I use a ByteDance recorder daily. Every meeting in my company is captured, transcribed, and analyzable by a large language model. I cannot attend every meeting, but I can virtually attend every meeting through that AI analysis. That is a profound operational insight tool.

The broader point is that China has a cultural advantage in data generation at scale. In AI 1.0, China had a data advantage from its vast population and usage. In AI 2.0, driven by large language models, the advantage shifted because the training data was drawn largely from the public internet, where English-language data was of particularly high quality. But now, with new data being generated through audio, video, and physical world interactions, China’s advantage returns – partly from scale, and partly because there is less cultural resistance to recording and data capture than you would find in Europe or the US.

Add to that a government mandate focused explicitly on using AI to transform manufacturing. China’s economy remains fundamentally manufacturing-led, rather than knowledge-based. The State Council of China’s public priority is deploying AI in factories and industrial settings. That mandate creates demand, focus, and funding alignment that does not exist to the same degree elsewhere. And Chinese VCs are more willing to fund hardware ventures precisely because they believe China has the edge there, the cost of building hardware is lower, the required capital is therefore smaller, and speed is much faster.

Chinese VCs are more willing to fund hardware ventures precisely because they believe China has the edge there.

Three options – not two

What are the most consequential mistakes you see countries make in how they think about AI sovereignty?

For countries, the clearest error is believing that the only two choices are to accept an American model or to build one from scratch. Those are not the right two options. The real choices are: first, to adopt a leading closed-source American platform, with full awareness of the sovereignty trade-offs involved; second, to take an open-source model and continue training it for your language, values, and regulatory requirements – at a fraction of the cost and with a realistic chance of success. Building from scratch to compete with the US and China is a third path, but it requires going in with eyes fully open to the fact that, for most countries, it is unlikely to succeed.

[1] CUDA (Compute Unified Device Architecture) is a proprietary parallel computing platform and application programme interface that lets companies use graphics processors to massively speed up data-heavy computing tasks.