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

A Miami AI firm says Chinese-developed models are processing as much as 46% of token traffic on a major U.S.-based AI router, with most businesses unaware of where their prompts are going. Layer3Labs says the shift raises data-sovereignty, compliance and security risks for companies that rely on vendor-built or unsanctioned AI tools.

Why it matters: - Chinese-developed AI models are now handling a large share of U.S. business AI traffic, creating data-sovereignty and compliance exposure for companies that may not know where customer information is being processed. - Layer3Labs says the risk is especially high in regulated industries, where prompt data can touch foreign endpoints without a company’s direct approval. - Businesses with heavy shadow AI exposure face an estimated $670,000 breach premium per incident.

What happened: - Layer3Labs, an AI implementation firm based in Miami, says as much as 46% of token traffic on the largest U.S.-based AI model router is now routed to Chinese models. - A year ago, U.S. models handled roughly 70% of that traffic. Today, the share is closer to 30%. - The firm says most business owners do not choose the model directly. They buy an AI feature, and the underlying model can be routed, resold or white-labeled without notice. - Layer3Labs says it reviewed dozens of small- and mid-sized business AI stacks and found 70% contained at least one Chinese model the company had not knowingly approved.

The details: - Cost-optimizing routers can send prompts to the cheapest available provider, which Layer3Labs says is increasingly a Chinese endpoint with inference costs 60% to 90% lower. - Andreessen Horowitz estimates 80% of U.S. startups build on Chinese base models, with Airbnb, Uber and Cursor listed as adopters in cited coverage. - Research cited by the firm puts 70% of enterprise AI activity outside IT oversight, including small models that can run on a laptop and bypass security review. - Any prompt processed by a Chinese-provider endpoint can fall under China’s National Intelligence Law, which can compel disclosure to the state. - In July 2026, the House Select Committee on the CCP opened a probe into U.S. firms’ use of these models. - The NDAA already bars DeepSeek from Defense Department systems.

Between the lines: - The shift is not just about model preference. It reflects how AI is being embedded through vendors, routers and unsanctioned tools that hide the model layer from the business user. - That makes procurement and governance harder than traditional software review, because the company buying the product may not see where the data actually flows. - The issue is becoming a policy and security problem, not just a cost issue.

What's next: - Layer3Labs is urging companies to run a provenance check on every AI tool that touches customer data. - The firm is offering a free AI Stack Sovereignty Check to identify which models are processing a company’s data and to flag sovereignty and compliance risk. - Layer3Labs says healthcare, legal and finance firms should be first in line for review. - Jonathan Teplitsky, Layer3Labs CEO, says business owners should ask whose jurisdiction their customer data now lives under. - Mark Boyhous says the first step is simply finding out what is in the stack.