As the U.S. and China battle for artificial intelligence supremacy, a fundamental divide in engineering philosophy could determine the winner.
While American pioneers like OpenAI and Anthropic favor a closed-source approach — keeping their proprietary model code locked behind a commercial interface — Chinese AI labs are aggressively releasing open-source models. This strategy allows developers to download, inspect, and deeply customize the underlying code for free, rather than remaining dependent on an American tech giant’s ecosystem. Tiezhen Wang, former head of the Asia-Pacific ecosystem at AI community collaboration platform Hugging Face, has observed this trend over the years. Before he moved on from Hugging Face in May, he assisted AI labs in the region in launching open-source models, and helped researchers make their models easier for developers to use.
At a Rest of World virtual event, Wang spoke about the history of open-source models, how Chinese AI labs monetize despite not charging for their models, and the debate over model distillation and intellectual property.
The conversation has been edited for length and clarity. A recording of the live Rest of World event is available here.
What is the role of open-source models in the China-U.S. AI competition? #
A lot of people talk about the AI competition. But in the open-source world, we hold a collaborative mindset.
Many open-source releases from the Chinese labs are helping U.S. labs. For example, the reinforcement learning training algorithm from DeepSeek is becoming the default setting for many U.S. research labs. Many Chinese open-source weights are running on U.S. hardware.
It is like helping each other, not competing like a zero-sum game. We can both be winners if we are growing the pie together.
OpenAI and Anthropic have accused Chinese AI companies of distilling their models. How do you view this as an open-source advocate? Is this a challenge for Chinese AI labs? #
Distillation is a neutral word in the research world. It’s like I’ve read a book, and I’m telling someone what the book is about. The other person also understands what the book is saying. That’s basically distillation.
It has been done a lot in research labs. I don’t think distillation by itself is anything wrong, and I believe people in the U.S. are distilling each other as well. Recently, we have seen Elon Musk admit that xAI distilled from OpenAI.
It is well known that Anthropic and ChatGPT are crawling the internet, getting all kinds of information. So it’s interesting to see that those who did not generate knowledge are trying to stop others from reusing that knowledge.
All AI-generated content should have zero copyright, otherwise people who have compute can abuse and generate all kinds of combinations and copyright everything.
How does monetization differ for open-source and closed-source models? #
If you open-source your model, it’s hard to make money directly. But you can still make money.
For example, China’s [Kimi released their model for free](https://www.scmp.com/tech/big-tech/article/3350887/moonshot-ai-releases-flagship-model-open-source-push-continues), but their application programming interface and subscription are still in huge demand because they have the best infrastructure support.
Once a model is released as open-source, people have to spend engineering hours to run it. So, on day zero, the research lab that launched the model has an advantage. That’s one way they can make money.
Another way is, for example, Kimi could release a fine-tuned model as open-source, but keep the base models to themselves, which they can sell.
If we want to think about why a research lab would open-source its models, it is so that they can build their branding. When you first start a lab, why would researchers want to work for you? It’s very hard to acquire top talent. But if you have good open-source models, everyone knows you did some great work.
## Why are some Chinese AI labs [pulling back from open-source releases](https://www.reuters.com/commentary/breakingviews/china-ai-labs-face-growing-open-source-dilemma-2026-03-25/)?
Some models are changing their licenses. For example, Minimax changed the license to basically say that if you use this model to make money, you have to pay. This is a very common practice in the open-source world, especially if you want to prevent free riders, like cloud users.
Cloud providers could run an open-source model for free and generate profit without needing to share profit with the research lab. This is unfair.
So basically, the trend of changes is that if you are an individual user, you can use my model for free forever, but if you are a cloud provider generating revenue by serving my model, you have to give me a portion of your profit. I think this is fair and actually a sustained way of supporting open-source.
Do you think more Chinese AI labs will go closed-source because of monetization problems? #
I’m concerned. If they do not find a way to monetize their research, this is a real risk. If they can have money and can keep doing open-source, it’ll be great for everyone.
I think the capital market is coming to help. If you look at the stock price of China’s Zhipu, it is at 10 times growth already. This will definitely help them acquire more compute, talent and data, and generate better models. These investments can keep these labs on the table for a longer time.
If you were advising a U.S. startup today, would you tell them to build on American models, Chinese models, or whichever open-source model performs best? #
It’s a well-known strategy that nowadays in the U.S., if you want to launch a startup, the first step is to find a model that has the best product-market fit. That is more important than your technical decisions. If you don’t have a product that deeply satisfies strong market demand, your startup will fail. These models would be closed-source.
But once you have this closed-source model in place, you start to have the first group of users, and you are accumulating data. You know how your product is being used. You can later use the data you generate to train your model. Then companies will start to consider switching to an open-source model. Eventually, you will save maybe a hundred times on tokens.
You just returned from a trip to China. Do you have any new observations about the country’s AI development? #
I had a very interesting trip talking to a few research labs and companies generating tokens. I do feel that the whole Chinese market is quickly maturing.
The U.S. is definitely doing a great job with exploring what AI can do, but the token in their way is more expensive even for big companies to afford. For example, Uber burned their entire year’s tokens within four months. Microsoft also said it felt the tokens were more expensive than they’d expected.
In China, it’s kind of the other way around because it has a bunch of open-source models that are not that expensive to use.
As soon as Chinese open-source models cross a usable point, usage will grow exponentially. What I’ve seen is that all Chinese internet companies started tokenmaxxing, which is basically giving their employees unlimited tokens just to see what they can do. I have heard that a few big Chinese tech companies are forcing their employees to be AI-native and stopping them from doing normal things like writing a document.
This kind of approach to AI adoption is way faster than what the U.S. has done. I won’t be surprised if over the next year or two, you see a lot of very interesting AI use and adoption in China.