China's AI Crackdown Hits the Middlemen Routing Users to ChatGPT and Claude

The group-chat message arrived without warning. A Hangzhou-based platform administrator told tens of thousands of subscribers that police had summoned him, ordered the service shut, and that refunds would be issued. The service, which had quietly signed Chinese users up for OpenAI's ChatGPT and Anthropic's Claude, was ending "starting today."

That announcement, screenshotted and shared across Chinese social media this week, is the most concrete sign yet that Beijing is now squeezing a layer of its AI controls it had previously left alone. For years, Chinese regulators blocked ChatGPT and Claude at the network border, requiring users to navigate VPNs, overseas phone numbers, and foreign payment methods. A layer of domestic services grew up to do that work for them: AI relay platforms, Chinese intermediaries that sell subscriptions and API access to overseas models, manage billing in yuan, and translate the experience into a normal local product. They are the on-ramp most Chinese developers, small businesses, and curious users actually used.

Now the state is squeezing the relay layer. The Hangzhou operator is the first publicly identified target. The Ministry of State Security published a public warning in early June criticizing AI relay platforms for lacking proper qualifications, posing cross-border data-transfer risks, and threatening "national security" (Global Times). It urged users to choose "secure and compliant" AI services instead. In the days since, multiple Chinese relay operators have suspended foreign-model offerings or scaled them back, according to industry observers and the Epoch Times report that broke the Hangzhou story (Epoch Times).

The new pressure is not a new law. It is enforcement, not legislation. China already requires generative AI services to register with the Cyberspace Administration of China (CAC) and to filter training data and AI-generated content under the Interim Measures for the Management of Generative Artificial Intelligence Services. By the end of April 2026, 868 domestic generative AI services had registered with the regulator, and another 530 applications and functions that rely on foreign AI models had completed separate filings (Epoch Times). Those numbers describe the existing architecture, not a rebuttal to the crackdown.

The relay layer is small in name but large in footprint. It handles payment localization, account registration, customer support, and often the reverse-engineering of foreign APIs. Shutting it down does not require banning ChatGPT or Claude; they are already blocked at the border. It requires only that the handful of Chinese companies willing to run the pipes stop doing so. The South China Morning Post has documented Chinese authorities warning for years that "relay services" for foreign AI models create leaks and backdoors (SCMP). In Chinese state-media usage the term is less a technical exploit than a regulatory one: a way for foreign models to enter China without going through the registration and content-review process the state built.

For Chinese users and developers, the immediate effect is friction. Some relay platforms had offered ChatGPT and Claude access at near-overseas prices, with Chinese-language interfaces and local payment. Others had built their own products on top of foreign models. The shutdowns remove the cleanest path to those models without going through a domestic Chinese alternative. The state's preferred alternatives, the registered domestic services, are required to comply with content rules, training-data approvals, and government filings. The relay services explicitly did not.

Police summons without a public legal process leave operators and their customers in the dark about what specifically was prohibited. The Ministry of State Security's "national security" framing has been applied to ordinary commercial services. An anonymous current-affairs commentator told the Epoch Times, speaking "out of fear of reprisal," that the ruling Communist Party remains concerned that overseas AI models have not undergone China's registration, content review, and data supervision (Epoch Times). That concern is real, but it is a policy choice, not a fact of the technology.

Outside China, the relay layer is now a named part of the access architecture. Border blocking, the domestic registration regime, and the relay services that connect them have always been the three layers. Until this week, the third was invisible to outside readers.

The Hangzhou platform's refund process is already underway. The Ministry of State Security's warning has not been followed by a named regulation. The next concrete signal will be whether other named relay operators receive the same call, or whether the pressure stays administrative and quiet.