Half your AI stack may be speaking Mandarin. That's not the problem. Not knowing is.

Chinese AI models handle more than a third of the tokens US companies burn each week. Since February 8, every single week. The peak hit 46%. A year ago it sat near 11%. That’s OpenRouter data, and it describes real workloads from real American companies routing to models most of them have never heard of.

Your developers didn’t go rogue. This is arithmetic.

Open-weight models like GLM deliver coding on par with Opus at roughly a sixth of the cost. Check any benchmark site. It’s common knowledge now. When one model does the same job for a sixth of the price, the tokens go there. They always do. Nobody held a meeting about it. The router picked the cheaper model that passed the eval, the bill came in lower, and everyone moved on.

Here’s what the market-share headlines miss. Routing work across many models is standard on every serious agent platform. One model drafts. Another reviews. A third handles the cheap bulk calls. Your SaaS vendors do this inside their product. Your own teams do it inside your walls. So your data is almost certainly touching models you never vetted, running in places you never mapped.

Provenance is a visibility question, not a loyalty test.

You don’t need a China policy. You need to know what’s under your own stack. At Cosmos Holdings we run a 73-agent system in production, and we can name which model touches which workload and where each call runs. That’s the bar. If we can say it, your vendors can say it.

So send one email to your top three AI-embedded vendors. Ask two things. Which models process our data. Where does inference run.

A vendor who knows answers in a day. A vendor who doesn’t just told you something worse than any answer.

Then make it permanent. Add the model manifest to your standard vendor review. Same line item, every time. A named list of which models see your data and where they sit. No manifest, no signature.

This is a new idea. Most vendors have never been asked. That’s exactly why you ask first.

The tokens flowing east aren’t the risk. Not knowing which tokens are yours is.

Send the email today.