CNBC reports that Chinese AI models have carried at least 30 percent of enterprise token traffic on OpenRouter every week since February, peaking at 46 percent. Your stack may have voted already, and nobody brought it to a meeting.
On July 7, CNBC published an investigation with a number that should reorganize a few meeting agendas. Chinese-origin AI models have accounted for at least 30 percent of enterprise token volume on OpenRouter, one of the largest model routing platforms, every single week since February 8. The weekly peak hit 46 percent.
A year of context makes it sharper. Over the prior twelve months the average was 11 percent. In the first half of 2025 it was 4.5 percent, per the same reporting.
DeepSeek alone now routes 17.6 percent of OpenRouter's tokens, the largest share of any single vendor on the platform, Western labs included. Alibaba's Qwen adds another 13.9 percent.
If your reaction is "we would never," keep reading. This did not happen at companies that decided to adopt Chinese models. It happened at companies that never decided anything.
1. How this happened without a decision
Price. CNBC found open-weight Chinese models running 60 to 90 percent cheaper than the flagship offerings from OpenAI and Anthropic. DeepSeek's budget tier was listed at $0.14 per million input tokens against $5.00 for GPT-5.5.
Now put that price gap in front of an engineer with a routing platform and a cost target. Switching models takes one line in a config file. There is no purchase order, no vendor onboarding, no security review triggered, because the spend runs through a router the company already approved. The procurement decision dissolves into a thousand pull requests.
That is the real story here. Model choice became an engineering variable while your governance still treats it as a procurement event.
2. June poured fuel on it
Two accelerants showed up in the reporting. First, a US export-control action briefly suspended access to Anthropic's newest models in June before access was restored in early July. Every team that lost its primary model for weeks learned the same lesson about single-vendor dependency, and diversification stopped being a slide and became a config change. Second, the practice went public: CNBC noted that companies including Airbnb and Uber have openly acknowledged using Chinese models.
Washington noticed. On July 8, CNBC reported that lawmakers opened a probe into the growing use of Chinese AI models inside US companies. Whatever your firm's position is, you want to have one before someone asks for it.
3. "Chinese model" is three different risk conversations
The label covers three deployment patterns with very different exposure. Panic flattens them into one. Governance separates them.
| How the model reaches you | What it actually is | What to check |
| Self-hosted open weights | Model files running inside your own cloud tenancy | License terms, provenance, alignment behavior, how updates arrive |
| Direct API to the model vendor | Your prompts leave your boundary | Data residency, retention, training use, governing jurisdiction |
| Embedded in a SaaS product | Your vendor's routing decision, invisible to you | Subprocessor disclosures, model-change notice clauses |
The third row is the one most firms miss. Your vendors face the same margin pressure your engineers do. When a SaaS tool quietly swaps its underlying model to protect its unit economics, your data follows that decision without you.
4. What a grown-up response looks like
Audit your token mix. Router dashboards and egress logs will tell you in an afternoon which models your traffic actually touches. Do this before writing any policy, because the policy conversation changes when the number is real.
Put model provenance in vendor vetting. Every AI tool review should answer: which models, hosted where, under whose jurisdiction, with what notice when it changes.
Write an allowed-model list with tiers. Public marketing copy does not need the same model standard as client work product. Tie model tiers to data sensitivity and put it in writing.
Fix the contracts. Ask for notice-of-change language on models and subprocessors in your SaaS agreements. You will not always get it. The asking still tells you who your careful vendors are.
Make the cost trade explicit. Saving 80 percent on tokens is a legitimate business decision for the right workload. Silence is the problem, because silence means nobody weighed the trade.
Final thought
The headline number will keep moving. The governance lesson stands still: your model mix is a decision, whether or not anyone makes it on purpose. Firms that treat it as one will save money in the right places and hold the line in the right places. Firms that don't will find out their position from a router dashboard, or from a regulator.
AI, Properly is the Ayvede field report: AI news read plainly, with a takeaway you can use. Ayvede helps high-trust firms map which AI touches their data, and under what rules. Advise. Vet. Deliver. Ask a question in writing.