Beijing Watch: China Cracks Down Hard on 'AI Relay Stations'! Operators Detained for 37 Days, Cotton Cloud Cuts Connections — Are ChatGPT and Claude's Backdoor Channels Gone Overnight?

Since OpenAI, Anthropic, and other overseas large language models do not officially serve mainland China, numerous developers, cross-border e-commerce businesses, AI startups, and academic researchers have relied on third-party API relay stations to access GPT, Claude, and similar models for product development, code testing, and overseas operations. These relay services have become a critical infrastructure within China's AI ecosystem.

According to multiple Chinese developers and community sources, public security agencies and cyberspace regulators across mainland China have recently begun investigating API relay platforms that provide access to overseas AI model interfaces. Many active API services have ceased updates or closed new registrations, and some developers have received official notices demanding they stop offering such services. While no unified official statement has been released by Chinese authorities, the developer community widely perceives a significant tightening of regulation.

At the same time, many Chinese developers, small businesses, and researchers who previously depended on API relay stations to access GPT, Claude, and other models for programming, cross-border marketing, and product testing are now seeing these channels rapidly disappear. As a result, companies are being forced to switch to domestic large models or migrate their systems to overseas servers to maintain operations.

From 'de facto tolerated' to 'key target' — enforcement cases have already emerged.

API relay stations have long existed in a legal and regulatory gray zone. These platforms enabled Chinese developers, small companies, and researchers to access overseas models like OpenAI's GPT and Anthropic's Claude at token prices far below official rates, primarily for software development, cross-border marketing, and product testing. Most platforms do not train their own models but act as proxy interfaces: connecting to overseas AI services on one end and providing unified APIs to mainland users on the other, settling payments in RMB.

This model solved practical issues such as payment processing, network connectivity, and account registration, significantly lowering the barrier to entry and rapidly scaling up. For many small and medium-sized enterprises, this approach was even more cost-effective than deploying open-source models directly.

Journalists tested several major overseas API platforms and found that most require users to pay with a Mastercard or Visa issued outside mainland China. While Chinese banks now issue China-specific Mastercards, some payment recipients are flagged as high-risk by domestic banks.

Several individual developers in China reported that a relay station operator in Shanghai was criminally detained for 37 days on suspicion of illegal business operations due to a 'reverse proxy plus account pool' model, later released on bail. In cities like Hangzhou, authorities have directly summoned operators and demanded unlicensed platforms be dismantled within a deadline. Some cloud service providers (e.g., Cotton Cloud) have issued notices explicitly prohibiting the use of domestic nodes to set up any form of global AI relay station, warning of network disconnection, account suspension, and cooperation with investigations for violations.

Earlier, China's Ministry of State Security published an official article directly warning that some relay stations pose risks such as 'data exposure,' 'model degradation,' 'malicious code injection,' and 'uncontrolled data出境 (data exfiltration),' urging users to choose platforms with clear origins, operational licenses, and security safeguards.

Some developers have notified their clients that they will no longer provide access to overseas models.

Domestic AI models face their biggest 'replacement window'

Over the past two years, despite continuous improvements in domestic models like Baidu's ERNIE, Alibaba's Qwen, Tencent's Hunyuan, ByteDance's Doubao, DeepSeek, and Zhipu AI's GLM, many development teams still preferred GPT or Claude as default platforms due to their mature international ecosystems, superior English capabilities, and widespread adoption by overseas clients.

Now, as the cost and difficulty of accessing overseas interfaces rise, companies are reassessing their technical strategies. Many SaaS companies are migrating customer service, office automation, and marketing content generation to domestic models. Some startups are adopting a 'dual-system' approach — using domestic models for domestic operations while relocating servers to Hong Kong, Singapore, Japan, or the U.S. to continue accessing overseas models for international business.

This approach aligns with Chinese government regulations on 'data security,' which prohibit unauthorized cross-border data transfers and require review for inbound foreign data.

Under China's 'Interim Measures for the Management of Generative Artificial Intelligence Services' and related data security laws, providing generative AI services requires filing and approval. Most major overseas models have not completed these procedures in China. Unauthorized value-added telecommunications services and failure to conduct data出境 security assessments may lead to administrative or even criminal liability.

OpenAI's ChatGPT and Anthropic's Claude. (Image source: Storm Media)

China's AI industry enters a new phase prioritizing 'controllability'

Over the past year, China's AI industry has shifted from merely comparing model parameters to competing on ecosystem strength. The ability to offer stable APIs, comprehensive development tools, enterprise deployment solutions, and compliance support will determine long-term partnerships.

For large internet companies, this means more customers entering the domestic ecosystem. But for small and medium-sized development teams, it means real pressures: migration costs, code refactoring, and model re-adaptation.

Many AI applications were originally built around GPT or Claude, with deeply integrated prompt designs, tool calling, function interfaces, and agent workflows. Migrating to domestic models often requires extensive retesting and optimization.

From VPNs and cloud computing to data cross-border flows and generative AI registration, China's regulatory focus has consistently centered on 'self-reliance and controllability' and 'security governance.' For China, AI is no longer just a commercial tool but a critical infrastructure tied to national data security, industrial competitiveness, and digital sovereignty.

This suggests that China's AI development path may increasingly integrate 'domestic models, domestic computing power, and domestic cloud platforms.'

The twilight of relay stations may just be the beginning. As regulatory reach extends to every interface, Chinese developers are no longer just choosing 'which model to use' — but whether they can choose freely at all.