Why AI Token Resale Platforms Matter: Fu Sheng, Justin Sun and the New “Water Business” Around Model Usage

A new kind of platform has become visible at the intersection of AI and Web3: the AI token relay. Here “token” usually does not mean a crypto governance token. It means the units consumed when a user or an AI agent calls a large language model. Instead of connecting separately to OpenAI, Claude, Gemini, DeepSeek, Kimi and other providers, users buy access through one API, one account balance and one billing layer.

That sounds like a thin business. Model API prices are public, token counts are measurable, and a reseller that simply forwards calls should not be able to hide much margin. Yet the current interest from Fu Sheng-related EasyRouter and Justin Sun-advised B.AI suggests a deeper bet: as AI usage grows, the valuable position may be the meter, the pipe and the settlement account, not the individual token itself.

The verified facts first

EasyRouter describes itself as a one-stop AI model gateway. Its public site says it supports more than 40 models, including OpenAI, Claude and Gemini, and offers centralized model management and distribution. Its documentation says the service provides RESTful APIs, including AI model APIs and management APIs, and is compatible with the OpenAI API format. Its public status endpoint also shows a points/quota-based system and exchange-rate fields.

What remains less clear is the operating entity. Some Chinese media have described EasyRouter as a platform launched by Fu Sheng of Cheetah Mobile. But the EasyRouter site and documentation I checked do not directly disclose the legal operating company, nor do they explicitly state that Cheetah Mobile or Fu Sheng controls the platform. Front-end text on the site says online top-ups are collected by a non-mainland China entity, but it does not name that entity.

This distinction matters. Cheetah Mobile’s official materials do show a clear AI turn. In January 2024, Cheetah Mobile announced further investment in Beijing OrionStar through wholly owned subsidiaries. Its 2024 Form 20-F describes an “AI and others” segment and says some enterprise customers are interested in applying LLM technology to improve operational efficiency. Cheetah’s own site lists Orion-14B, ChatMax and multiple AI application components. None of those official disclosures, however, directly identifies EasyRouter as a Cheetah Mobile listed-company business. The careful conclusion is therefore: Fu Sheng and Cheetah Mobile are deeply associated with an AI/LLM strategy, and EasyRouter is discussed by Chinese media in that context, but the platform’s precise corporate ownership still needs clearer disclosure.

Why build it if the resale spread is transparent?

Retail API prices are public, but actual procurement is not just the public price page. Large volume commitments, prepaid credits, regional payment friction, uptime requirements, routing, retries and customer support can all create room for an intermediary.

Justin Sun explained the logic more directly in a TechFlow interview about B.AI, where he is described as the project’s strategic adviser. He said B.AI’s business model is similar to an exchange: it earns mainly from the spread on API-call tokens plus a small fee, currently around 5%. The spread, he said, comes from the platform’s ability to obtain wholesale discounts from model providers because of its throughput. B.AI’s own site describes it as a unified API and settlement network for AI agents, with borderless payment infrastructure.

That is a more ambitious position than a simple proxy. It combines model access, payment, settlement and eventually agent autonomy. EasyRouter’s public shape is closer to a model gateway for developers and enterprises: one API key, one dashboard, unified balance and usage tracking. For developers, it reduces integration work. For enterprises, it turns scattered model accounts, currencies and invoices into a managed budget pool.

Where the money can come from

• Spread.A platform buys or commits to usage at a discount and resells access at retail-like or package prices. The margin may be thin, but high throughput can make it meaningful.
• Fees and payment infrastructure.Cross-border payments, crypto settlement, enterprise top-ups, refunds, risk control and invoice-like documentation can all become paid services. B.AI has publicly mentioned a small fee; EasyRouter’s product flow clearly includes recharge, balance and usage accounting.
• Routing and reliability.Enterprise buyers do not only want cheap tokens. They want availability, fallback models, retry logic, rate-limit management, budget controls, permissions and logs. That is where a reseller can become infrastructure.
• Customer relationship and usage data.The platform that knows which tasks consume which models, how much they cost and where failure happens is well positioned to sell agents, workflow software, governance tools and vertical AI solutions.

What can the “meter” actually see?

The value of the meter should not be described as a platform magically knowing every secret of an enterprise. A more precise description is that a real AI gateway sees operational metadata: which API key, team or application made a request; when it happened; which model and provider were used; how many input and output tokens were consumed; what it cost; how long it took; whether it failed; whether it hit cache; and whether it triggered a security or DLP policy. If the customer attaches metadata such as user ID, project, environment or feature name, the gateway can attribute spend to even finer business units.

This is already how the market is developing. OpenRouter returns token counts, costs, reasoning tokens and cached tokens, and lets users export activity by model, API key or organization member. Cloudflare AI Gateway shows requests, tokens, caching, errors and cost, and its logs can include prompt, response, provider, timestamp, request status, token usage, cost and duration. Portkey, Helicone and LiteLLM all emphasize tracking spend, latency, errors, budgets and usage by user, team, feature or environment. In other words, the “meter” has become an auditable, attributable and governable management system.

Why is that data valuable?

• Cost attribution.A CFO does not only want to know that the company spent $100,000 on AI this month. The useful question is: how much was customer support, sales analysis, coding assistance, document processing or internal agents? Which team is over budget? Which feature has poor unit economics? In FinOps language, this is allocation, showback and chargeback. Without the meter, AI cost is a single bill. With the meter, it becomes operating data.
• Budget control and risk management.AI systems can run away quickly: looping agents, bad retry logic, oversized prompts or leaked API keys can burn budget overnight. A gateway with real-time token, cost, rate and error data can enforce quotas, rate limits, alerts, circuit breakers and anomaly detection. Twilio UsageTriggers, AWS API Gateway usage plans and Portkey budget/rate limits are all examples of this broader pattern.
• Procurement leverage.If an enterprise knows that 60% of usage comes from a certain class of tasks, that 30% can move from an expensive model to a cheaper one, and that batch jobs are not latency-sensitive, it can negotiate better discounts, commit volumes or switch providers. FinOps rate optimization explicitly uses historical and planned usage data to support procurement. A gateway that aggregates many customers can also use its own volume to negotiate wholesale rates upstream.
• Routing and product optimization.The same task can have very different cost, latency, error rate and output quality across models. If the gateway records request type, model, cost, latency, errors and feedback, it can decide which tasks need a frontier model, which can use a cheaper model, which should be cached and which prompt should be redesigned. Helicone calls this understanding unit economics; Portkey frames it as finding patterns and optimizing efficiency.
• Compliance and auditability.Once companies use AI in production, management will ask: did sensitive data go to an external model? Who called what? Which requests were blocked? Can we reconstruct an incident? Cloudflare’s logging and DLP fields, and AWS Bedrock invocation logging, show that large customers need traceability and governance, not just lower prices.

How an entry point becomes infrastructure

The entry point becomes infrastructure only when it grows from a request forwarder into a control plane. Stripe is valuable not because it counts transactions, but because it connects metering, pricing, billing, invoices, collection, credits and revenue analytics. Twilio is valuable not only because it sends messages, but because it turns communications into APIs with usage records, subaccounts, routing and threshold alerts. AWS, Azure, Kong and Cloudflare API gateways are not ordinary reverse proxies; they are enforcement points for authentication, quotas, routing, logging, monitoring, billing and security policy.

An AI token gateway can become infrastructure in the same way. Developers get model access from it. Finance reads the bill from it. Security sets rules on it. Procurement sees provider usage through it. Product teams compare the unit economics of each AI feature through it. At that point, switching away is not just changing an API URL. It means rebuilding budgets, permissions, logs, audit trails, routing rules and reports.

The reverse is also true. If a platform only offers cheap relay access, without enterprise permissions, logs, budgets, routing, compliance and billing workflows, it is unlikely to become infrastructure. Its value remains short-term arbitrage and customer acquisition subsidy, and it will be squeezed when model providers cut prices or improve their own enterprise tooling.

Why Fu Sheng keeps talking about tokens

Fu Sheng’s recent public comments have been centered on AI agents and token consumption. In a Beijing Business Today interview reposted by Tencent News, he described his AI assistant “Sanwan” as consuming roughly 100 to 200 dollars of tokens per day and moving from content work to management tasks such as weekly reports, task breakdowns, group creation and even replies to department heads. He also said token consumption is replacing DAU and MAU as a new evaluation metric, and that the first wave of impact is increased model-token and cloud usage.

A China Entrepreneur article reposted by Sina Finance went further in its framing, saying tokens have become a new hard currency for tech companies and are beginning to appear in compensation and performance incentives. The article quotes Fu Sheng saying the watershed moment for enterprise LLM adoption has arrived, and that tokens will become cheaper over time. It is important to be precise: the “hard currency” wording appears to be the publication’s formulation, not a directly verified Fu Sheng quote. I also did not find a reliable primary source in which Fu Sheng himself says tokens are “like water”; that phrase appears more often as media commentary on how quickly tokens are consumed.

Still, the direction is clear. Fu Sheng treats tokens as a new production input for AI companies. In the mobile internet era, companies tracked daily users, monthly users, time spent and conversion. In the AI-agent era, a more revealing metric may be how many tokens an organization is willing to spend to get real work done. If that view is right, the model gateway is not just a channel. It is a control point for enterprise AI budgets.

What it could mean for Cheetah Mobile

Cheetah Mobile’s old mobile utility story is no longer the center of its growth narrative. Its official disclosures now emphasize AI applications, LLMs, robotics and OrionStar. If a Fu Sheng-related token gateway were ever formally brought into the listed-company perimeter, its importance would probably not be high standalone gross margin. Its importance would be that it gives the company an entry point into model consumption.

Robots, AI assistants, enterprise agents, customer service and office automation all require stable, affordable and governable model calls. Training proprietary models is capital-intensive. Pure API resale is thin. But a combination of model gateway, agent application and enterprise workflow can become a fuller stack: buy model capacity at the bottom, route and bill in the middle, sell productivity and vertical solutions at the top.

That is why a seemingly unglamorous resale platform may matter. It may not be the final product. It may be the account system through which enterprise AI spending becomes visible and controllable. Once that budget sits on the platform, higher-value software and agent services become easier to sell.

Can this become a large platform business?

Only if it moves beyond “cheap relay.” If the business relies solely on arbitrage, model-provider price cuts and better official enterprise plans will compress the margin quickly. If the platform solves multi-model routing, cross-border settlement, compliance documentation, enterprise permissions, logging, autonomous agent purchasing and risk control, the value is much higher.

B.AI is framed more as a financial highway and settlement layer for AI agents. EasyRouter looks more like a model gateway and developer/enterprise access layer. The shared thesis is that model capability will be consumed like electricity, water or cloud computing. The valuable company may not own every drop of water. It may own the meter, the pipe, the valve and the enterprise bill.

So the token resale platform is likely a low-margin business in the short run, but it could become an infrastructure business if AI usage keeps growing. Its ceiling is not “reselling tokens.” Its ceiling is controlling how AI usage is allocated, paid for and governed. That is a large idea, but it still faces two tests: clearer disclosure of legal and financial ownership, and proof that the platform adds reliability, cost control and productivity rather than merely acting as a middleman.

参考资料 / Sources

• EasyRouter
• EasyRouter API docs
• EasyRouter status API
• B.AI
• TechFlow interview on B.AI / Justin Sun
• Cheetah Mobile IR: investment in Beijing OrionStar
• Cheetah Mobile about page
• Cheetah Mobile 2024 Form 20-F
• Beijing Business Today interview reposted by Tencent News
• China Entrepreneur article reposted by Sina Finance
• C114Pro English article on Fu Sheng AI agent product
• NetEase article on token relay platforms
• Yahoo Finance / CoinDesk on SunPump
• SunPump official site
• OpenRouter usage accounting
• OpenRouter activity export
• Cloudflare AI Gateway analytics
• Cloudflare AI Gateway logging
• Cloudflare AI Gateway unified billing
• Portkey analytics
• Portkey metadata
• Helicone cost tracking
• Helicone custom properties
• LiteLLM virtual keys
• Kong AI Gateway
• AWS API Gateway usage plans
• FinOps cost allocation
• FinOps anomaly management
• FinOps rate optimization
• Stripe usage-based billing
• Twilio Usage Records