Summary of Key Points
This article discusses the current global development status of Chinese AI models: Chinese models have surpassed American models in terms of the volume of tasks they handle (token share) among developers worldwide, yet their revenue contribution is extremely low. For example, on Vercel, Chinese models process 29% of the tasks but only receive 4% of the income. DeepSeek’s recent price increase represents an experiment with pricing power—previously, Chinese models entered the market by offering lower prices; now, they are testing whether they can shift from handling a large volume of work to generating higher profits. The article also explores the core challenges Chinese models face: how to convert global usage into direct customers, profits, and company value. It highlights the different business strategies adopted by several Chinese companies (MiniMax, Alibaba, Kimi, DeepSeek), concluding that the key to success in the future lies not in being cheap, but in demonstrating differentiation and irreplaceability.
I. Chinese Models: Doing the Most Work, Getting the Least Money
You can think of AI models as “digital workers”—some perform complex tasks (such as planning and review), while others handle basic tasks (like coding and speech recognition). Currently, Chinese models are primarily responsible for the latter:
- Data speaks for itself: On OpenRouter, Chinese models handle a larger volume of tasks than American models. However, on Vercel, they process 29% of the tasks but only earn less than 4% of the revenue.
- Real-world example: American developer Garcia spends $700 per month, with $500 going to American models for complex tasks and $200 to Chinese models for 90% of the basic tasks.
This is similar to a factory where assembly line workers do most of the work but earn much less than the technical staff; Chinese models currently play the role of “cheap, basic workers” in the AI production process.
II. DeepSeek’s Price Increase: Moving from Price War to Pricing Power
Over the past year, DeepSeek has shocked the industry with its low prices (e.g., generating one million tokens for just 6 yuan). Now, it is raising prices: V4 Pro costs 13.5 yuan during off-peak times and 27 yuan during peak hours. This change signals two things:
1. Pre-IPO test: DeepSeek is preparing to go public and needs to prove its profitability rather than relying on low prices to dominate the market.
2. Pricing power experiment: Developers used Chinese models because of their low cost; with the price increase, will they continue to use them? If they do, it indicates that Chinese models offer more than just affordability; if not, it means pricing power has yet to be established.
This move is crucial as it determines whether Chinese models can transition from being widely used but cheaply priced to being both widely used and profitable.
III. Models Going Global, But Customers May Not Be Yours
While it’s easy for Chinese models to enter international markets, customer relationships and revenue often do not follow:
- Open-source issue: Model code is freely downloadable (open source), allowing users to run them on their own servers or on American cloud platforms (like AWS) without signing contracts with Chinese companies. For example, Airbnb uses Qwen but is not an Alibaba customer.
- Middlemen profiting: Revenue from model calls may first go to cloud platforms (AWS) or routing service providers rather than directly to the Chinese model companies. If users deploy models themselves, Chinese companies might not receive any money at all.
This is like selling seeds and only getting money from the sale of the crops grown by others—Chinese models need to establish strong customer relationships and control the service chain, not just sell the models themselves.
IV. Chinese Companies’ Approaches to Generating Revenue
Several Chinese companies are trying different strategies to address the issue of high usage but low revenue:
1. MiniMax: Attempts to generate revenue overseas directly, but at a high cost. In 2025, overseas revenue accounted for 73% of their total income, despite incurring $250 million in research and development losses—meaning the profits from sales were not enough to cover expenses.
2. Alibaba: Builds value through its ecosystem. The launch of Qwen model has driven growth in Alibaba Cloud and Model Studio’s AI-related revenue (30% of Alibaba Cloud’s total income). The model is free, but the cloud services generate profit.
3. Kimi: Directly sells APIs. Both overseas paid users and API revenue have increased by 400%, with annual revenue exceeding $300 million in June—revenue comes directly from developers.
4. DeepSeek: Tests pricing power by raising prices. They use their high global usage as leverage to see if they can increase prices, which is a direct test of their ability to set prices.
V. Can Chinese Models Break Free from the “Cheap” Label in the Future?
The answer is possibly yes:
- Contrary examples: ByteDance’s Seedance generates 34% of videos but earns 49% of the revenue; Kuaishou and Alibaba’s video models also generate significant income. This shows that price is not determined by nationality but by the ability to provide unique solutions.
- Global adoption beyond low-budget markets: Developers in Germany and Canada use Chinese models at over 50%, and the US accounts for 42.9% as well. Chinese models are being adopted in developed countries, not just in low-income regions.
- What the capital market looks for: It’s not just about the volume of tokens (tasks) but whether spending can keep up with task volume, whether overseas companies sign direct contracts, and whether models can handle more complex tasks. Additionally, whether the gross profit is sufficient to fund future research and development.
In summary, Chinese AI models have shown that intelligence can be affordable, but the real challenge is to prove that they can also create significant value for companies. DeepSeek’s price increase is a test of this transition. While this article does not provide a definitive answer, it clearly indicates that Chinese AI models have reached the global stage, but there is still a long way to go before they become profitable entities.