Chinese AI models power Western workloads while revenue stays in the U.S., Dimension Capital says
Key Takeaways
- •Chinese open-weight models account for more than 60% of token usage on platforms such as OpenRouter, according to Dimension Capital.
- •Eight of the top 10 models on OpenRouter reportedly come from Chinese labs and together generate more than 25% of the platform’s token volume.
- •American inference companies such as Fireworks and Baseten are capturing most of the revenue, while Chinese model builders receive only a small share.
- •Dimension said OpenAI is producing about $40 billion in annual recurring revenue, Anthropic more than $6.5 billion, and Chinese model companies remain below the hundreds of millions.
- •The report argues that U.S.-China AI decoupling is incomplete because American firms are building on Chinese-developed models and the supply chain remains interdependent.
Chinese AI labs have quietly become the engine behind a huge share of Western AI workloads. The money, though, stays in America.
That is the central conclusion of an analysis Dimension Capital shared with limited partners after a research trip to Beijing and Shanghai. The venture firm, which closed an $800 million fund in July 2026, said open-weight models built by Chinese teams now account for more than 60% of token usage on platforms such as OpenRouter, up from essentially nothing not long ago.
The work flows east, the dollars stay west
Dimension described the market as a lopsided symbiosis. U.S. frontier labs such as OpenAI and Anthropic continue to develop leading proprietary models, while Chinese teams including DeepSeek, Alibaba’s Qwen, and ByteDance’s Doubao have produced open-weight alternatives that American application companies then refine and deploy. According to the firm, eight of the top 10 models on OpenRouter now come from Chinese labs and together capture more than 25% of total token volume on the platform.
Token volume, however, is not the same as revenue. The inference providers actually collecting payments from developers and enterprises are overwhelmingly American companies such as Fireworks and Baseten. Chinese model builders, despite powering a substantial share of usage, receive only a small portion of the financial return, underscoring how usage leadership and monetization can diverge in AI.
The revenue gap remains wide. OpenAI is generating roughly $40 billion in annual recurring revenue, while Anthropic has surpassed $6.5 billion. Chinese model companies, by contrast, are said to be below the hundreds of millions.
Open-weight models: free to use, costly to monetize
Open-weight models are designed to be freely usable. Labs release the model weights publicly, allowing anyone to download, fine-tune, and deploy them without licensing fees. For U.S. companies building applications on top of these models, that creates access to advanced AI capabilities at a fraction of the cost of licensing proprietary alternatives, which helps explain why these systems can spread quickly through developer tooling and enterprise workflows.
For the Chinese labs producing the models, the economics are different. Open-weight releases can build reputation, attract talent, and expand ecosystem influence, but they do not generate direct revenue at scale. The economic value is captured further downstream, at the inference layer, where American providers charge customers for compute and API access.
Dimension characterized the setup as a “two-way trans-Pacific flow” of AI technology. Even with the revenue imbalance, the firm said Chinese frontier labs are commanding valuations 5-10x higher than their U.S. counterparts on a revenue-multiple basis.
Decoupling remains incomplete
The Dimension report also complicates the broader narrative of U.S.-China tech decoupling. Washington has imposed sweeping export controls on advanced semiconductors and chipmaking equipment in an effort to slow China’s AI progress by limiting access to cutting-edge hardware. But the integration Dimension observed goes beyond hardware and includes model development, data pipelines, and inference frameworks.
American companies are actively building on Chinese-developed models, while Chinese labs continue to iterate quickly despite hardware constraints. The report said Chinese labs have shown a notable ability to do more with less. Working with tighter resource constraints than their American peers, they have produced models competitive enough to dominate usage rankings. That pattern has been visible since DeepSeek first drew attention for strong benchmark results using fewer high-end chips than analysts expected.
U.S. inference providers, in Dimension’s framing, are effectively the toll collectors on a highway built with Chinese technology. The arrangement is profitable today, but it also means American revenue depends in part on the continued willingness of Chinese labs to release their work openly, a reminder that AI supply chains remain interconnected even as governments push for separation.
