Mankun Legal Education | AI Token Going Global: Three Paths to Selling Chinese Computing Power Worldwide

This isn't about issuing or speculating on cryptocurrencies; it's about turning AI capabilities into a measurable and scalable global service. Many people, upon hearing "token," immediately think of Web3, cryptocurrency issuance, exchanges, and secondary markets. However, the global expansion of AI tokens is different. In Southeast Asia, a child holds a talking AI toy, letting it tell stories, practice English, and answer countless questions. The toy may seem like just a hardware terminal, but its true value lies not in the plastic shell, but in the continuous model reasoning behind it: every wake-up call, every follow-up question, and every voice interaction consumes tokens. These requests originate from overseas terminals, pass through a pilot channel in Shantou, enter the domestic computing center for processing, and then return the results to overseas users. In other words, what's being sold isn't a toy, but a set of AI capabilities that are used on a per-use, per-use basis. What they consume is the new hard currency of the AI era—tokens. One kilowatt-hour of electricity buys in for about 0.5 yuan, converts to tokens, and sells for about 11 yuan, increasing in value more than twenty times. The business we're discussing isn't about "how to sell a token," but rather a positive endeavor: selling domestic AI capabilities to overseas clients through compliant token-based pricing. Why is it worthwhile now? Domestic model capabilities are improving, API prices are decreasing, and overseas demand is real; leading companies are already making a significant portion of their revenue from overseas markets. However, regulations are tightening: AI intermediaries are under criminal investigation, and data security risks are being repeatedly highlighted. In short—AI tokens can go global, but they can't go haphazardly. Going in the wrong direction will turn going global into drowning; choosing the wrong model will turn business into risk. First, define the nature: AI tokens are not Web3 tokens. Before discussing models, we must firmly establish the concepts. The difference between AI tokens and Web3 tokens isn't in the name, but in their function. Web3 tokens are typically issued based on blockchain, allowing on-chain transfers, secondary market access, and market-driven pricing, possessing payment, value storage, and even speculative attributes. AI Tokens are completely different: they lack a blockchain infrastructure, cannot be transferred between users, have no secondary market, and experience no price fluctuations. They are simply an internal unit used by users to measure consumption after purchasing AI services. To illustrate this clearly: a user deposits $100 and receives a quota of 1 million Tokens. This quota cannot be sold, transferred to others, or listed on exchanges; it can only be used within the platform to access the model. This is the prerequisite for the compliant overseas deployment of AI Tokens. The judgment is actually quite simple: a Token can only be used for service consumption; it is a measurement tool. Once it can circulate, be transferred, or be speculated on, it may slide into the regulatory framework for virtual assets.Therefore, the first boundary for going global is—don't turn AI service quotas into financial products. Why now? Three figures bring the timeline closer. In the past year or two, three figures can explain why this business has suddenly become so hot. The first is usage volume. In the week around the Spring Festival this year, on OpenRouter, the world's largest model API aggregation platform, the top ten models in terms of usage consumed a total of about 8.7 trillion tokens, of which domestic models accounted for about 5.3 trillion, or 61%. By the week of early April, six of the top ten were domestic models, with Chinese models processing 12.96 trillion tokens that week, while US models only processed 3.03 trillion. Overseas developers are spending real money to use domestic models. The second is price. The API pricing of domestic models is often only a fraction to a fraction of that of the top US models—the input price difference is about ten to twenty times, and the output price difference is even greater. In the past, AI was mainly used for chatting, with small usage volumes, so the price difference was negligible. However, in the era of Agents, a single task can consume hundreds of thousands or even millions of tokens, drastically amplifying the cost difference, naturally leading developers to choose not to use it. The third factor is revenue. Take MiniMax as an example: its overseas revenue accounted for 73% in 2025, compared to only 19% in 2023. This shows that Chinese AI companies are not just being "used," but are genuinely generating revenue in the global market. Expansion, price reductions, and impressive overseas performance constitute the first half of the statement; the second half is equally crucial—compliance is becoming a key variable determining how far a company can go. The next three paths diverge precisely on this point. Model 1: Direct Official Connection. The most orthodox, clearest, and most compliant path is for model vendors to go global themselves. This involves vendors building their own nodes overseas or relying on international clouds like AWS and Azure to provide official APIs. Data cross-border transactions, model licensing, and local operations all adhere to international cloud rules and target market laws. With a complete authorization chain, clearly defined service providers, and self-responsibility, customers are clear about whose model they are calling, eliminating issues such as unclear sub-authorization in intermediate layers, unknown interface sources, model swapping, and account pool arbitrage. DeepSeek, Zhipu, and Moonlit Dark Side have all launched official APIs; Zhipu, MiniMax, and others also place their models in overseas clouds like AWS, with inference completed on local physical servers, ensuring the data never leaves the region—this precisely addresses the most pressing concern for overseas companies: will my data be transmitted back to China? The logic of this approach is: whoever has the model goes global, and whoever provides the service is responsible. It's suitable for leading vendors with overseas entities and resources; the barrier to entry is high, but once successful, it's the most worry-free. Model Two: Compliance Aggregation/Tools.For small and medium-sized teams lacking the capacity to deploy global nodes independently, a more realistic approach is aggregation and tool development, which is currently the most popular path. The strategy involves aggregating models from multiple authorized providers, offering a unified gateway, a unified API, and a set of developer tools, allowing overseas clients to access multiple models through a single interface. OpenRouter, mentioned earlier, is a global benchmark in this model—one API connects to hundreds of models. Revenue comes from the difference between the purchase and sale price of tokens, gateway and technical service fees, tool subscriptions, and value-added services such as routing, monitoring, billing, and risk control. This is a more complex area with stricter compliance requirements; it's not impossible, but it must be done with specific conditions. The two most crucial questions are model sourcing and data responsibility. Where do the models you aggregate come from? Do they have resale rights? Does the upstream authorization cover overseas markets? Can they be white-labeled and repackaged? To whom is the customer data ultimately routed? Will it be retained or used for training? If these questions are not answered clearly, "compliant aggregation" will become "gray market transit." It's essential to maintain "four closures": closed authorization chain, closed business scope, closed data responsibility, and closed billing evidence. If these four points are not met, the so-called "compliant aggregation" will degenerate into a "gray transit." Model Three: Special Economic Zone Pilot. There is another path with policy dividends, and the most watched example right now is Shantou. In 2025, the state approved the Shantou Overseas Chinese Economic and Cultural Cooperation Pilot Zone to conduct a "data processing" pilot program: overseas data can legally enter the country, be processed within designated areas, and then exported, operating under a unified policy framework without requiring individual application and approval. This is accompanied by a "digital bonded zone" architecture—the zone is physically isolated from the domestic internet; overseas requests are directly connected via international submarine cables, inference is completed within the zone, and results are returned via the same route, with latency to Singapore as low as approximately 32.7 milliseconds. Guangdong Mobile built the Eastern Guangdong Data Center there and also built and operates its own "Token Tongyi" platform, unifying computing power scheduling, token measurement, cross-border matching, and profit sharing, successfully completing the first full-chain closed loop for token export in China. It solves a crucial problem—whether overseas data can be processed domestically. The rule of thumb is: receive data overseas, process it domestically, and return it overseas; if no personal or important domestic data is mixed in during the process, it may be exempt from the three procedures of security assessment, standard contract, and protection certification. However, this path has clear boundaries. Several conditions must be adhered to: the service recipients are overseas users, the data source is overseas, only technical processing is done domestically, the results are returned via the original route, no domestic data is mixed in, physical or logical isolation is maintained between domestic and overseas, and operational traceability is auditable.Once a single system is used both domestically and internationally, data pools are not isolated, or processed results are flowed back to domestic business use, the exemption logic will fail. More importantly, the pilot program is not a nationwide exemption. It relies on specific regions, specific policies, and specific facilities, and cannot be simply replicated as "I can find any data center in China to process overseas data." Moreover, due to local regulations, it is currently mainly aimed at Southeast Asia. Due to strict data localization and export controls in Europe, the United States, and Japan, related services often still need to be deployed locally. It is more like a real but uncopyable testing ground. Third, a negative aspect that must be avoided: Token import. Having discussed the three legitimate paths, we must point out a negative aspect that is often mistakenly regarded as a "fourth way of playing": gray transit, namely, token import. It is the opposite of going global—going global is selling domestic AI capabilities to overseas clients, while import is repackaging overseas models and services that have not yet achieved compliance in China and selling them back to domestic users. The typical approach involves three steps: first, using bulk registration, shared accounts, or even reverse engineering to obtain overseas model interfaces; then, using reverse proxies to package them into Chinese web pages or API recharge platforms; and finally, selling them to the domestic public through social media, e-commerce, and mini-programs as memberships or tokens. The problem isn't a single qualification flaw, but rather the instability of the entire chain. In short: token-based overseas expansion is not a fourth way of going global, but rather a cautionary tale that overseas businesses should avoid at all costs. In conclusion: choosing the right model is crucial for compliance. There's no inherent superiority of one over the other, only suitability. Those with their own models, computing power, brand, and overseas infrastructure should prioritize direct official connections; those with technical integration and customer channels but lacking self-developed large models should realistically opt for compliant aggregation/tools; those needing to leverage domestic computing power to process overseas data while also accessing policy-controlled regions and isolation facilities can explore special economic zone pilot programs. However, packaging overseas interfaces and selling them to domestic users, regardless of whether it's called a transit point, mirror, or gateway, is not going global, it's entering the overseas market. Regardless of the path taken, four lifelines remain unavoidable: model origin, data export, target market, and fund recovery. If the model's origin is unclear, the token isn't an asset, it's a potential liability; if the data flow is unclear, the platform isn't a tool, it's a black box; if the target market is unclear, going global isn't growth, it's intruding into foreign regulatory oversight; if the funding path is unclear, revenue isn't profit, it's an explanation cost that will eventually have to be paid. The key to successful AI token overseas expansion isn't low prices, packaging, or rhetoric, but a business chain that can withstand scrutiny. The token is merely a unit of measurement; what's being sold overseas is always AI capability—where it comes from, who it's sold to, how the data flows, how the money is recovered, and who is responsible if problems arise.If these questions are answered clearly, tokens can truly be a tool for overseas expansion; otherwise, they will become just another gray area. [Mankiw Blockchain Legal Services]