Chinese users are paying third-party brokers to reach ChatGPT, Claude and other American AI models they cannot access directly, with stablecoins becoming part of the payment system behind that trade. Research from blockchain analytics firm Chainalysis describes a commercial market that connects customers to restricted models while keeping their location out of the direct relationship with the provider.
The brokers sell access rather than build the underlying AI. They pass requests to models operated by American companies and return the answers, charging customers through credits, account balances or subscriptions. The arrangement makes geographic restrictions harder to enforce and gives Chinese developers another way to study competing systems.
Stablecoins have become a payment option
The firm examined the market in research published November 10. Its findings describe services that accept Chinese payment methods such as WeChat Pay and Alipay alongside cryptocurrency.
The firm gave a specific measure of crypto's presence in the businesses it examined:
Our survey of 100 proxy services found that 40% also accept crypto as payment, predominately in stablecoins.
Stablecoins are tokens designed to track a reference value, commonly the US dollar. For these brokers, they provide a way to accept payment outside ordinary bank transfers and domestic payment apps.
The money does not always go straight to the company selling AI access. Many brokers use custodial crypto payment processors, which handle funds between customers and service operators. That adds another business to a transaction that already involves an intermediary between the user and the AI provider.
The researchers argue that this structure can help operators avoid maintaining their own crypto payment infrastructure, while creating more distance between buyers and the businesses receiving the money. They also identify potential links to efforts to bypass Chinese capital controls.
Payments rise around new model releases
The wallet analysis found increased activity from suspected China-linked wallets in the days following major American AI launches. Those increases exceeded the pattern across all wallets using the services.
One possible explanation is straightforward: customers want access to new capabilities quickly. For developers trying to understand a rival's latest model, the first days after release are an opportunity to collect responses and test what has changed.
The payment data cannot establish the purpose of each transfer or prove that a particular customer was training a competing model. That limit sits alongside the observed timing, rather than turning the correlation into proof of what every buyer did.
The wallet activity was concentrated in China, with connections also identified in Japan, Hong Kong, Singapore and South Korea. The analysis describes those regional links as consistent with Chinese operators using intermediary routes.
The policy pressure runs in both directions. As previously reported by Apex in the article below, Cursor told House investigators it would explore phasing out Chinese-developed open-weight models because of geopolitical risk. Those findings were released November 20.
OpenAI and Anthropic's rules still exclude China
Mainland China is absent from both companies' current lists of supported locations. Buying access through another business does not make that access part of an authorized service relationship.
OpenAI's policy, covering ChatGPT and its API, warns:
Accessing or offering access to our services outside of the countries and territories listed below may result in your account being blocked or suspended.
Anthropic's rules, effective November 12, go beyond where an individual user is sitting. They cover entities incorporated or headquartered in unsupported regions and businesses majority-owned or controlled by people or entities in those regions, including indirect ownership or control.
The policy applies across its apps, developer platform and Claude models delivered through cloud platforms and authorized resellers. Moving the commercial relationship to a different country therefore does not automatically put an otherwise excluded business within its terms.
These company access policies are central to the market the researchers describe. The brokers are selling a route around restrictions on who can use the systems and where, while the upstream providers retain the ability to suspend accounts.
The bigger concern is copying capabilities at scale
Beyond individual use, American developers are concerned that intermediaries can supply the volume of access needed for model distillation. That technique trains one model using another model's outputs, potentially reducing the work needed to develop similar capabilities.
Distillation is also a legitimate method when carried out with authorization. The dispute concerns access obtained through fraudulent accounts and the use of responses in ways the original provider prohibits.
In a February statement, Anthropic accused DeepSeek, Moonshot and MiniMax of extracting Claude's capabilities through approximately 24,000 fraudulent accounts. It said the campaigns generated more than 16 million exchanges and violated its terms and regional restrictions.
The company described patterns concentrated on reasoning, coding and tool use, along with networks that distributed requests across many accounts.
The financial research describes a similarly divided supply chain: upstream businesses obtain model access, brokers package and resell it, and payment processors move customer funds. That division helps explain why shutting down a single seller may leave other parts of the business running.
For investigators, the stablecoin payments also leave transaction trails to examine. Funds connected to brokers can eventually reach regulated crypto platforms without an obvious direct link to the original service. Mapping those links, the researchers argue, gives enforcement teams a way to look beyond the customer account or website that first attracted attention.