I'd like to point out that OpenRouter reports statistics by model, but all Chinese models mentioned are offered by multiple providers, many of which are US companies.
It's of course possible that the information about data center location is incorrect, but if it's not, most of the tokens OpenRouter attributes to Chinese models likely never pass through a data center in China at all.
The “tokens made in China” framing is powerful, but the more actionable question is where the meter actually runs. Even when Chinese open models dominate OpenRouter usage, a large share of that inference is increasingly hosted and billed through non-China providers—often via U.S. inference platforms or hyperscaler-managed shelves—so the immediate profit pool can sit with delivery and governance layers rather than the model origin.
OpenRouter is still a valuable leading indicator for the developer/agent ecosystem, but enterprise adoption will likely route through managed endpoints (governance + audit + liability) instead of direct cross-border inference. In that sense, the story is less “decoupling vs coupling” and more “model globalization vs delivery capture.”
I’ve tried to map this same dynamic from two angles—where inference gets settled/hosted, and how open-model economics compress closed-model premiums—happy to share those notes if useful.
The 'tokens have no physical form, do not go through customs' line really landed. That framing captures something most of the geopolitics commentary misses. I've been tracking the same dynamic from the US inference provider side; American companies like Fireworks and Together AI are hosting DeepSeek and Qwen, profiting from models built on chips the US tried to restrict. The whole thing is a policy feedback loop. Dug into the economics of it here: https://medium.datadriveninvestor.com/who-profits-when-ai-models-are-free-b71ae03f4167
This discussion about artificial agents and their economic implications really resonates with the evolving nature of human attention in a hyper-automated world. If you're interested in exploring how our attention becomes a scarce resource in this new landscape, you might find my piece on Some Simple Economics of AGI meaningful: https://theuncomfortableidea.substack.com/p/the-scarcest-resource-in-an-automated.
I'd like to point out that OpenRouter reports statistics by model, but all Chinese models mentioned are offered by multiple providers, many of which are US companies.
Minimax itself is listed as a Singaporean company with data centers in the US: https://openrouter.ai/minimax/minimax-m2.5/providers
Moonshot is listed as a Singaporean company with data centers in Singapore: https://openrouter.ai/moonshotai/kimi-k2.5/providers
Z(hipu).ai is listed as a Singaporean company with data centers in Singapore: https://openrouter.ai/z-ai/glm-5/providers
Only DeepSeek is listed as a Chinese company, and is disabled by default, even as a provider for DeepSeek models: https://openrouter.ai/deepseek/deepseek-v3.2/providers
It's of course possible that the information about data center location is incorrect, but if it's not, most of the tokens OpenRouter attributes to Chinese models likely never pass through a data center in China at all.
This is really helpful context -- I will note a correction in next week's issue.
The “tokens made in China” framing is powerful, but the more actionable question is where the meter actually runs. Even when Chinese open models dominate OpenRouter usage, a large share of that inference is increasingly hosted and billed through non-China providers—often via U.S. inference platforms or hyperscaler-managed shelves—so the immediate profit pool can sit with delivery and governance layers rather than the model origin.
OpenRouter is still a valuable leading indicator for the developer/agent ecosystem, but enterprise adoption will likely route through managed endpoints (governance + audit + liability) instead of direct cross-border inference. In that sense, the story is less “decoupling vs coupling” and more “model globalization vs delivery capture.”
I’ve tried to map this same dynamic from two angles—where inference gets settled/hosted, and how open-model economics compress closed-model premiums—happy to share those notes if useful.
https://leonliao.substack.com/p/chinas-token-exports-are-explodingwhat?r=731anr&utm_medium=ios
The 'tokens have no physical form, do not go through customs' line really landed. That framing captures something most of the geopolitics commentary misses. I've been tracking the same dynamic from the US inference provider side; American companies like Fireworks and Together AI are hosting DeepSeek and Qwen, profiting from models built on chips the US tried to restrict. The whole thing is a policy feedback loop. Dug into the economics of it here: https://medium.datadriveninvestor.com/who-profits-when-ai-models-are-free-b71ae03f4167
This discussion about artificial agents and their economic implications really resonates with the evolving nature of human attention in a hyper-automated world. If you're interested in exploring how our attention becomes a scarce resource in this new landscape, you might find my piece on Some Simple Economics of AGI meaningful: https://theuncomfortableidea.substack.com/p/the-scarcest-resource-in-an-automated.