Chinese AI models now handle nearly half of U.S. router traffic

Layer3Labs says Chinese-developed models are now processing up to 46% of token traffic on a major U.S.-based AI router, up from about 30% a year ago. The firm warns that many businesses do not know when vendor tools or shadow AI are sending customer data to Chinese endpoints, creating security and compliance risk.

Why it matters: - Chinese-developed AI models are now handling a large share of U.S. business prompts, raising data-sovereignty and compliance concerns for companies that never chose those models directly. - Layer3Labs says the risk is highest for healthcare, legal and finance firms, where customer data can trigger regulatory exposure if it leaves approved channels. - The firm says high shadow AI exposure is linked to an estimated $670,000 breach premium per incident.

What happened: - Layer3Labs says as much as 46% of token traffic on the largest U.S.-based AI model router is now routed to Chinese models. - The company says U.S. model traffic on that router has fallen from roughly 70% a year ago to about 30% today. - Layer3Labs released the warning on July 22, 2026, from Miami. - The firm says many businesses are using AI features without knowing which underlying model processes their data.

The details: - Cost-optimizing routers increasingly send prompts to cheaper Chinese endpoints, where inference is said to cost 60% to 90% less. - Andreessen Horowitz estimates 80% of U.S. startups build on Chinese base models, with named adopters including Airbnb, Uber and Cursor. - Research cited by Layer3Labs says 70% of enterprise AI activity sits outside IT oversight, much of it through Chinese models small enough to run on a laptop. - Layer3Labs says a review of dozens of small- and mid-sized business AI stacks found that 70% contained at least one Chinese model the company had not knowingly approved. - The company is offering a free AI Stack Sovereignty Check to identify which models are processing company data and flag sovereignty and compliance risk. - Layer3Labs says any prompt processed by a Chinese-provider endpoint falls under China’s National Intelligence Law, which can compel data disclosure to the state. - The House Select Committee on the CCP opened a probe in July 2026 into U.S. firms’ use of these models, and the NDAA already bars DeepSeek from Defense Department systems.

Between the lines: - The shift is happening through vendors, routers and white-labeled AI products, so the model behind a business-facing feature can be different from the brand on the label. - The core exposure is visibility, not just model choice, because companies cannot secure or govern AI use they do not know exists. - Layer3Labs is positioning model provenance checks as an emerging vendor-risk control, not just a technical audit.

What’s next: - Layer3Labs is urging companies to run a provenance check on every AI tool that touches customer data. - The warning is aimed especially at regulated industries that may need to map where prompts go, which models handle them and which jurisdictions apply. - The company’s stance suggests more buyers may start asking vendors to disclose model routing and data-handling paths before approving AI features.

The bottom line: - The biggest AI risk for many businesses may not be the chatbot they picked, but the model they never knew was in the middle.