Chinese public security agencies and cyberspace regulators have launched a sweep against third-party API relay services that gave mainland developers backdoor access to ChatGPT and Claude, dismantling what had become quiet but essential infrastructure for thousands of developers locked out of overseas AI platforms.
Shanghai Operator Detained 37 Days on Illegal Business Charges
The sharpest signal of Beijing's intent came in Shanghai, where a relay platform administrator was criminally detained for 37 days on suspicion of illegal business operations before being released on bail. The operator had run a "reverse proxy plus account pool" model — pooling foreign API access and reselling it to Chinese clients at rates well below official prices. In Hangzhou, police summoned operators directly and issued deadlines for unlicensed relay services to shut down.
Infrastructure providers have joined the enforcement chain. Mianhua Cloud issued a notice explicitly banning the use of its mainland nodes for any global AI relay network, warning of service termination, account suspension, and cooperation with investigators. yunwu.ai, operated by Chongqing Yunzhiwu Network Technology Co., announced it had terminated access to all overseas AI models by July 31, 2026.
Relay Operators Ran on Borrowed Time Under China's AI Rules
Relay stations occupied an informal but commercially significant role in China's developer ecosystem. Because neither OpenAI nor Anthropic offers official mainland service, relay operators bridged overseas AI infrastructure with Chinese clients — accepting yuan payments, managing account pools, and solving connectivity problems that would otherwise block access entirely. Cross-border e-commerce firms, AI startups, and university research teams all depended on them.
The legal exposure was always there. Under China's Interim Measures for the Administration of Generative Artificial Intelligence Services, any provider offering generative AI must complete a government registration process that most overseas models never have. Operating unlicensed telecommunications services, or moving data across borders without a security review, can carry criminal liability. China's Ministry of State Security had flagged relay stations in prior official publications, warning of unauthorized data transmission, degraded model output, and malicious code injection.
Domestic AI Giants Stand to Gain as Foreign Access Closes Off
Many Chinese development teams had defaulted to GPT or Claude even as domestic rivals — Baidu's ERNIE Bot, Alibaba's Tongyi Qianwen, Tencent's Hunyuan, ByteDance's Doubao, DeepSeek, and Zhipu's GLM — steadily improved. The foreign preference persisted because of stronger English performance, more mature developer tooling, and alignment with overseas clients' existing tech stacks.
Enforcement is forcing that calculus to change. SaaS firms are migrating customer service, office automation, and marketing generation to domestic models. Some startups have adopted a dual-track strategy: domestic models for mainland operations, with servers relocated to Hong Kong, Singapore, Japan, or the United States for international work — an arrangement that also satisfies China's data localization rules.
Migration Costs Fall Hardest on Smaller Development Teams
Switching is not a simple credential swap. Applications built around GPT or Claude typically have prompt logic, tool-calling functions, and agent workflows calibrated to those models' specific behavior. Moving to a domestic alternative requires full retesting and optimization cycles, a burden large internet companies can absorb far more easily than independent developers.
China's regulatory sequencing — from VPNs and cloud computing to cross-border data flows and now AI interfaces — has followed a consistent logic. Generative AI has been recast in official policy as critical national infrastructure tied to data sovereignty and state security, not a commercial tool. For China's developers, the question is no longer which model to use, but how much freedom remains to choose at all.
