WAIC 现场中东买家驻足:Token 交易市场这波"出海叙事",模型供应商从卖 API 切换到卖工厂

Instead of presenting another chatbot or promoting a single large model, the section demonstrated how computing resources could be packaged into standardized token services and delivered to overseas markets.

From Selling APIs to Exporting Token Factories: How China Is Building a New Global AI Supply Chain

At the H2 computing-power exhibition area of the 2026 World Artificial Intelligence Conference, or WAIC, an unprecedented scene appeared.

Several exhibitors had created a fully English-language section titled:

“Token Goes Global.”

Instead of presenting another chatbot or promoting a single large model, the section demonstrated how computing resources could be packaged into standardized token services and delivered to overseas markets.

The visitors who remained in front of the displays the longest were not domestic investors. They were developers and technology buyers from the Middle East and Southeast Asia.

Speaking in a mixture of English and hesitant Chinese, many of them asked essentially the same question:

“How do we buy these tokens?”

That question captured one of the most important changes visible at WAIC 2026.

The meaning of “token” is evolving.

It is no longer merely a technical unit used to measure a model’s input and output. It is increasingly being treated as a digital product that can be priced, purchased, scheduled, settled, and delivered across borders.

Throughout the exhibition halls, the change was visible.

Alongside large-model companies and major internet platforms, a new category of suppliers was emerging: companies building token factories.

SenseTime’s large-scale infrastructure reportedly delivers approximately 2.42 trillion tokens per day. PPIO has attracted more than 660,000 registered developers and operates over 5,000 computing nodes across six continents. Lenovo has gone even further in describing the transition:

Move from delivering computing power to delivering tokens.

The message from the exhibition floor was clear:

AI suppliers are moving from selling APIs to selling production capacity.


The Three Essentials: Why China Can Export a Factory, Not Just a Product

At WAIC, Chinese Academy of Engineering academician Zheng Weimin outlined three requirements for exporting large-model services:

“Large models going global require three conditions. First, high-quality large models. Second, token factories capable of producing high-quality tokens. Third, inexpensive electricity. Once all three conditions are present, tokens can be exported.”

These three conditions correspond closely to the structure of a factory.


High-Quality Models Are the Production Line

Without a capable model, there is no valuable token output.

The model determines whether the produced tokens can support:

  • Coding;

  • Translation;

  • Reasoning;

  • Customer service;

  • Document processing;

  • Agentic workflows;

  • Industry-specific applications.

China’s major model companies have widely adopted open-source or open-weight strategies.

DeepSeek, Qwen, Kimi, Zhipu AI, MiniMax, and other model families give overseas developers the option to deploy locally instead of relying entirely on a closed external API.

That flexibility has become increasingly valuable as governments and enterprises pay more attention to:

  • Data sovereignty;

  • Local deployment;

  • Regulatory control;

  • Model customization;

  • Supply-chain independence.

For many international users, an open model is not simply cheaper software.

It provides greater control over where data is processed, how the system is adapted, and whether the service remains available independently of one foreign provider.

This degree of autonomy is difficult for fully closed models to replicate.


Token Factories Are the Workshop

A model alone does not create a stable commercial token service.

The model must be combined with:

  • Computing clusters;

  • Inference engines;

  • Scheduling systems;

  • Caching;

  • Network infrastructure;

  • Monitoring;

  • Security controls;

  • Service-level management.

Together, these components form the token factory.

A PPIO representative summarized the value directly:

“Through domestic and international computing platforms, we can provide a complete token service. Model providers and users do not need to build their own infrastructure overseas in order to produce or consume tokens.”

SenseTime co-founder Yang Fan compared a token factory to a shoe factory.

Making one pair of shoes may be simple. Producing shoes with consistent quality, reliable delivery, efficient supply chains, and controlled costs is much more difficult.

The same is true for tokens.

Generating a model response is only the beginning.

A competitive token factory must consistently deliver:

  • Low time to first token;

  • High generation speed;

  • Stable concurrency;

  • Predictable output quality;

  • Reliable uptime;

  • Competitive cost.

Two companies may operate the same model and own similar hardware, yet produce token services of very different commercial value.

The difference lies in system engineering.


Cheap Electricity Is the Energy Cost

Electricity is one of the most important marginal costs in industrial-scale token production.

A token appears to be a digital product, but its production depends on physical infrastructure:

  • GPUs consume electricity;

  • Data centers require cooling;

  • Network and storage systems run continuously;

  • Redundant clusters increase power usage;

  • Long-context and agentic requests extend inference time.

China’s industrial power system, renewable-energy capacity, and large-scale infrastructure can provide a structural cost advantage in suitable regions.

A model can be deployed domestically while its token output is sold abroad through an API.

Pilot data from the Shantou Overseas Chinese Pilot Zone has been cited as showing deployment costs more than 30% lower than comparable European and American solutions.

In regions such as Southeast Asia and the Middle East, where AI demand is growing faster than local computing infrastructure, this cost difference can become highly competitive.


From Product Exports to Token Exports

When models, token factories, and lower energy costs are combined, China’s AI globalization strategy begins to change.

Traditional AI exports usually revolve around a specific product:

  • A chatbot;

  • A model API;

  • An enterprise application;

  • A cloud platform;

  • An industry solution.

Token exports follow another logic.

The customer may not be purchasing one named application or one fixed model.

Instead, the customer purchases a standardized intelligence service defined by:

  • Token quantity;

  • Model quality;

  • Response latency;

  • Throughput;

  • Reliability;

  • Price;

  • Data-processing requirements.

The export is no longer merely a product.

It is production capacity for intelligence.

This is why the phrase “Token Goes Global” matters.

It signals a shift from exporting individual AI products to exporting an industrial system that can continuously manufacture and deliver tokenized intelligence.


The Numbers Behind the Shift

China’s daily AI token consumption has reportedly increased from around the hundred-billion level in early 2024 to more than 140 trillion by March 2026.

That represents growth of more than 1,400 times in approximately two years.

OpenRouter-related data cited at WAIC indicated that Chinese models accounted for approximately 54.1% of global token consumption, compared with 41.5% for American models.

In coding-related workloads, the share attributed to Chinese models reportedly reached 67.6%.

Exact market shares may continue to change, but the direction is clear:

Chinese models are no longer serving only domestic demand. They are increasingly participating in global inference traffic.


Selling an API and Selling a Factory Are Different Businesses

The difference is not the token itself.

It is the commercial logic behind it.

Selling an API

The traditional API model works like this:

  • The provider owns a model;

  • The user sends requests;

  • The provider charges according to usage;

  • The supplier defines the price;

  • The user accepts the published terms.

Supply is centralized, and pricing is largely static and one-directional.

Selling a Factory

The token-factory model is broader:

  • Models, computing resources, electricity, and scheduling are packaged together;

  • Standardized token capacity is produced continuously;

  • Users purchase capacity according to workload;

  • Different suppliers may contribute to production;

  • Pricing can respond to market conditions;

  • Settlement occurs according to real consumption.

In this structure, the model provider is no longer the entire supply chain.

It becomes one part of the production line.

The token itself becomes the standardized commercial output.

At WAIC’s English-language “Token Goes Global” section, exhibitors were not simply demonstrating an API endpoint.

They were showing the complete chain through which computing capacity could be transformed into standardized token services and delivered internationally.

Overseas buyers were not only asking which model had the highest benchmark score.

They wanted to know:

  • How is the service connected?

  • How is it priced?

  • How is capacity guaranteed?

  • How is it delivered?

  • How is usage settled?

The transition from selling APIs to selling factories is therefore a transition from:

Technical output

to:

Industrial capacity output.


The Token Marketplace Connects the Factory With the Buyer

A factory can be built and production capacity can be increased.

But several commercial questions remain:

  • Who purchases the tokens?

  • How is pricing determined?

  • How is capacity routed?

  • How are suppliers paid?

  • How are overseas users served?

  • How is quality compared across providers?

These are circulation problems rather than production problems.

Without an efficient market connecting supply and demand, a token factory may become nothing more than idle computing capacity.

This is why the token marketplace becomes the missing link.

If tokens are compared with electricity, Zheng Weimin’s three essentials describe the equivalent of:

  • The energy resource;

  • The power plant;

  • The cost of generation.

The token marketplace resembles the electricity-trading market:

It allows production to flow efficiently from the factory to the user who needs it.

The more active the marketplace becomes, the more value the token factory can realize.

Factories need stable demand to absorb production.

Model providers need transparent pricing to plan costs.

Low electricity prices only create a meaningful advantage when fixed infrastructure is utilized at sufficient scale.


Unexhub: Building the Road Between Token Factories and Global Buyers

According to its public positioning, Unexhub is building a computing-power trading center for token aggregation and distribution.

Its intended role is to connect:

  • Model providers;

  • Computing suppliers;

  • Individual developers;

  • Enterprise customers;

  • International markets.

Several mechanisms in its proposed model are particularly relevant to the circulation of token capacity.


Real-Time Settlement: Let Suppliers Recover Capital Faster

Small and medium-sized suppliers usually need to spend money before they can generate revenue.

They may first need to pay for:

  • GPUs;

  • Cloud instances;

  • Model deployment;

  • Electricity;

  • Network bandwidth;

  • Engineering operations;

  • Customer support.

After delivering token capacity, they may still wait through a long billing and settlement cycle.

The more capacity they provide, the more funds can become trapped in accounts receivable.

Unexhub is attempting to shorten this process through real-time or near-real-time settlement based on actual token usage.

The intended model is simple:

Supply capacity, generate usage, and settle the corresponding revenue more quickly.

For suppliers, faster settlement can help:

  • Reduce receivables;

  • Improve cash flow;

  • Reinvest in capacity;

  • Respond to sudden demand;

  • Lower the financial barrier to expansion.

The hardware is not free, and the operating costs do not disappear.

The value lies in reducing the time between delivering a service and recovering the funds needed to provide the next unit of capacity.


Early Zero-Commission Access: Lower the Cost of Market Entry

Unexhub also proposes zero-commission incentives for qualifying early suppliers.

For suppliers entering a new international token market, early uncertainty is high:

  • Will developers use the model?

  • Which regions will create demand?

  • Will the route meet latency requirements?

  • Can the pricing generate a profit?

  • Will enterprise users trust the platform?

  • Can settlement operate consistently?

Zero commission helps reduce the cost of answering those questions.

Early suppliers can:

  • Complete integration at lower cost;

  • Test real demand;

  • Build usage records;

  • Establish market credibility;

  • Validate unit economics;

  • Secure early distribution share.

The benefit is not only the commission saved.

It is the ability to complete the first commercial loop—connection, routing, usage, settlement, and reinvestment—under more favorable conditions.


API Bidding: Let the Market Discover the Price

Unexhub’s proposed API bidding model allows prices to move with supply and demand.

When capacity is abundant:

  • Developers can access lower prices;

  • Suppliers can monetize idle GPUs;

  • Batch tasks can be routed to cheaper services;

  • Utilization improves.

When demand rises:

  • Prices can increase;

  • Suppliers gain an incentive to release more capacity;

  • Urgent enterprise workloads can receive priority;

  • Scarce compute flows toward higher-value use cases.

This creates a more flexible market than one static API price.

The price can reflect:

  • Model quality;

  • Current supply;

  • Regional demand;

  • Latency;

  • Stability;

  • Context length;

  • Peak or off-peak periods;

  • Service-level guarantees.

Pricing becomes a signal that coordinates supply and demand.


Millisecond-Level Routing: Match Each Workload With the Right Model

Unexhub also emphasizes millisecond-level request routing across different models and suppliers.

According to its positioning, the platform’s ecosystem includes access to model families such as:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • Models provided by individual developers and smaller companies.

Different workloads can then be assigned according to their actual requirements.

For example:

  • Lightweight extraction can use a lower-cost model;

  • Complex reasoning can use a stronger model;

  • Coding tasks can be routed to specialist capacity;

  • Flexible batch jobs can wait for cheaper off-peak routes;

  • Time-sensitive enterprise requests can use priority capacity.

This avoids sending every request through the most expensive model.

Developers pay for the capability the task actually requires.


Developer Benefits: Pay by Actual Token Usage

Through a unified token marketplace, developers can pay according to actual usage rather than investing in their own infrastructure.

Depending on the model, route, capacity, and market conditions, some services may be available at prices as low as approximately 30% of standard market rates.

For high-volume developers and enterprise users, this can reduce the cost of:

  • Coding agents;

  • Long-context processing;

  • Document analysis;

  • Customer-service automation;

  • Data extraction;

  • Batch generation;

  • Multi-model testing.

The real value of the discount is not limited to one cheaper API call.

It gives teams more flexibility to:

  • Test additional models;

  • Increase workload scale;

  • Move non-urgent work to lower-cost periods;

  • Reduce dependence on one provider;

  • Control budgets through intelligent routing;

  • Enter production without immediately constructing a private cluster.

A lower price becomes genuinely valuable only when combined with stable service, transparent billing, and predictable quality.


Supplier Benefits: One Integration, Access to Global Demand

For model and computing suppliers, the platform offers a different set of benefits.

A supplier may have:

  • A competitive model;

  • Specialized domain capability;

  • Regional computing resources;

  • Idle GPUs;

  • Low electricity costs.

But it may not have:

  • International sales channels;

  • Overseas developer traffic;

  • Billing infrastructure;

  • Cross-border settlement;

  • Multi-tenant API management;

  • Enterprise customer access.

Through a single integration, the supplier can potentially access a wider market.

Combined with real-time settlement, the path becomes more direct:

Connect capacity, receive routed traffic, and earn revenue based on actual token usage.

For smaller suppliers, this can reduce customer-acquisition and market-entry costs.

For larger suppliers, it can create an additional distribution channel for idle or off-peak capacity.


Heterogeneous Computing Makes the Token-Factory Model More Flexible

At WAIC, Zheng Weimin also mentioned an experimental project in Shanghai combining approximately 9,000 domestic accelerator cards with 1,000 international cards to produce tokens.

This kind of heterogeneous computing system is important because different chips offer different advantages:

  • Cost;

  • Availability;

  • Memory;

  • Energy efficiency;

  • Model compatibility;

  • Performance for specific workloads.

A token factory does not need to run every request on the same hardware.

It can assign tasks according to the most efficient combination.

For example:

  • Lightweight inference can use lower-cost domestic accelerators;

  • Specialized workloads can use international hardware;

  • Long-context tasks can use memory-rich nodes;

  • Batch workloads can use economy capacity;

  • Priority requests can use premium hardware.

The token marketplace can amplify this advantage because the user does not need to purchase access to a particular card.

The user purchases:

  • A model capability;

  • A token volume;

  • A latency objective;

  • A service guarantee.

The factory and marketplace determine which model, chip, region, and route can deliver that result most efficiently.


When the Three Essentials Are Connected, a Complete Export Route Emerges

Returning to Zheng Weimin’s framework:

  • High-quality models are the production technology;

  • Token factories are the manufacturing lines;

  • Low-cost electricity is the energy advantage.

Each is essential.

But the token marketplace is what connects production with demand.

It answers:

  • What should be produced?

  • Who wants to buy it?

  • How should it be priced?

  • Which route should deliver it?

  • How should the supplier be paid?

  • How should the overseas customer receive the service?

Once this circulation layer is added, tokens move from a technical metric to a digital product that can be:

  • Priced;

  • Routed;

  • Scheduled;

  • Settled;

  • Delivered internationally.

The complete route becomes:

Model → Token Factory → Token Marketplace → Global Developer or Enterprise


The Real Export Is Standardized Intelligence Capacity

China’s AI globalization strategy is undergoing a structural change.

The objective is no longer limited to exporting an application or promoting one domestic model.

What can be exported is an entire standardized system for producing and delivering intelligence:

  • Competitive models;

  • Efficient inference infrastructure;

  • Low-cost energy;

  • Distributed scheduling;

  • Token pricing;

  • Cross-border routing;

  • Financial settlement;

  • Enterprise support.

The companies and platforms that successfully close the full loop—production, circulation, and delivery—will gain a stronger position in the next phase of global AI competition.


Tokens May Be the New Oil, but Oil Creates Value in the Market

Tokens are often described as the “new oil” of the AI economy.

The analogy is useful, but it is incomplete.

Oil underground has limited commercial value.

Its value emerges through:

  • Extraction;

  • Refining;

  • Transportation;

  • Pricing;

  • Trading;

  • Distribution.

The same is true of tokens.

A high-quality model creates the potential for intelligence.

A token factory turns that potential into standardized output.

Competitive electricity lowers the production cost.

The token marketplace connects production capacity with real global demand.

The Middle Eastern developers standing in front of WAIC’s English-language “Token Goes Global” exhibition were not merely looking to purchase access to one Chinese model.

They were evaluating the output of a factory capable of producing tokens consistently, affordably, and at scale.

The shift from selling APIs to selling factories is more than a change in business model.

It represents a movement in pricing power—from the model provider toward the broader market.

Once tokens become standardized, tradeable services, the platforms that control routing, settlement, and price discovery will hold an increasingly important position in global AI distribution.

Unexhub’s proposed real-time settlement, early zero-commission access, API bidding, and millisecond-level routing are attempts to build that market layer.

Whether the model succeeds will depend on reliability, security, compliance, and operational execution.

But the direction is increasingly clear:

The next stage of AI globalization will not be defined only by who builds the strongest model. It will also be defined by who can manufacture, price, route, settle, and deliver high-quality tokens to the global market most efficiently.


中文版本

从“卖 API”到“卖 Token 工厂”:中国 AI 出海正在形成一条新的全球产业链

在 2026 年世界人工智能大会的 H2 算力展区,出现了一个前所未有的场景。

多家参展商专门设置了一块全英文板块,标题是:

“Token 出海”。

展区没有继续展示另一款聊天机器人,也没有重点宣传某一个大模型,而是在演示如何将算力封装成标准化 Token 服务,再交付至海外市场。

在展板前驻足最久的,并不是国内投资者,而是一批来自中东与东南亚的开发者和技术买家。

他们用尚不熟练的汉语或英语,反复询问着同一个问题:

“这些 Token 怎么买?”

这个问题,恰好揭示了 WAIC 2026 上最值得关注的变化之一。

“Token”的含义正在改变。

它不再只是衡量模型输入与输出的底层技术单位,而逐渐成为一种可以定价、采购、调度、结算和跨境交付的数字商品。

这一变化在整个展馆中都清晰可见。

除了大模型厂商与互联网平台之外,一批专注于 Token 工厂 的企业开始出现。

商汤大装置的日均 Token 服务量已达到约 2.42 万亿;PPIO 拥有超过 66 万注册开发者,并运营着 5000 余个覆盖全球六大洲的算力节点;联想则更加直接地提出:

从交付算力,升级为交付 Token。

展会现场向行业释放出的信号已经非常清晰:

AI 供应商正在从卖 API,走向卖生产能力。


“三件套”:中国为何能卖“工厂”,而不只是卖产品?

中国工程院院士郑纬民在 WAIC 上,为大模型出海提出了三个必要条件:

“大模型出海必须满足三个条件:第一是高品质的大模型,第二是能生产高品质 Token 的 Token 工厂,第三是电费要便宜。这三个条件具备了,Token 就可以出国。”

这三个条件,恰好对应一座工厂的三层结构。


高品质模型是“生产线”

没有高质量模型,就无法生产有价值的 Token。

模型能力决定了生成的 Token 能否完成:

  • 编程;

  • 翻译;

  • 推理;

  • 客户服务;

  • 文档处理;

  • 智能体工作流;

  • 行业应用。

国内主流模型厂商正在广泛采用开源或开放权重路线。

DeepSeek、Qwen、Kimi、智谱、MiniMax 等模型,使海外开发者有机会进行本地部署,而不必完全依赖封闭的外部 API。

随着各国政府和企业越来越重视以下问题,这种灵活性正在变得更加重要:

  • 数据主权;

  • 本地部署;

  • 监管控制;

  • 模型定制;

  • 供应链独立性。

对许多国际用户而言,开源模型并不只是更便宜的软件。

它还意味着可以更自主地决定数据在哪里处理、模型如何适配,以及服务能否脱离单一海外厂商持续运行。

这种自主可控的价值,是完全闭源模型难以替代的。


Token 工厂是“车间”

一款模型本身,并不能自动形成稳定的商业 Token 服务。

它还必须与以下环节结合:

  • 算力集群;

  • 推理引擎;

  • 调度系统;

  • 缓存;

  • 网络基础设施;

  • 监控;

  • 安全机制;

  • SLA 管理。

这些环节组合起来,才构成 Token 工厂。

PPIO 负责人对此给出了直白解释:

“我们可以通过国内外不同算力平台提供完整的 Token 服务,模型厂商与用户不需要在海外自建基础设施,也能生产和消费 Token。”

商汤联合创始人杨帆则将 Token 工厂比作鞋厂。

做出一双鞋可能并不困难,但要长期保证质量、稳定交付、控制供应链与成本,每个环节都会拉开能力差距。

Token 同样如此。

生成一次模型回复只是起点。

真正有竞争力的 Token 工厂,还必须持续提供:

  • 更低的首字延迟;

  • 更快的生成速度;

  • 稳定的并发能力;

  • 可预测的输出质量;

  • 可靠的服务可用性;

  • 有竞争力的成本。

两家公司即使运行相同模型、拥有相似硬件,生产出的 Token 服务也可能具备完全不同的商业价值。

差距来自系统工程能力。


低价电力是“能源成本”

电力是工业化 Token 生产中最重要的边际成本之一。

Token 看起来是一种数字产品,但其背后的基础设施仍是物理的:

  • GPU 需要持续耗电;

  • 数据中心需要散热;

  • 网络与存储设备需要长期运行;

  • 容灾集群会增加电力消耗;

  • 长上下文与智能体任务会延长推理时间。

中国的工业电力体系、绿电资源与规模化基础设施,在适合的地区可以形成结构性成本优势。

模型在国内部署,再通过 API 将 Token 服务卖向海外,由此形成一种新的跨境结构:

  • 在基础设施成本更有竞争力的地区完成计算;

  • 通过 API 向全球交付 Token;

  • 海外开发者无需自建集群;

  • 国内算力供应商通过国际需求实现变现。

汕头华侨试验区的试点数据也被用来说明这一优势:相关部署成本较欧美方案低 30%以上。

在东南亚、中东等 AI 需求增长快、但本地算力供给仍然不足的市场,这种成本差距具有明显竞争力。


从“产品出海”走向“Token 出海”

当模型、Token 工厂与低成本电力结合起来,中国 AI 全球化的逻辑便开始发生变化。

传统 AI 出海通常围绕某个具体产品展开:

  • 聊天机器人;

  • 大模型 API;

  • 企业应用;

  • 云平台;

  • 行业解决方案。

Token 出海则遵循另一种逻辑。

客户购买的不一定是某个固定应用,也不一定是某一个指定模型。

他们购买的是一种标准化的智能服务,其核心指标包括:

  • Token 数量;

  • 模型质量;

  • 响应延迟;

  • 吞吐能力;

  • 服务稳定性;

  • 调用价格;

  • 数据处理要求。

出海的不再只是产品。

而是 智能生产能力。

这正是“Token 出海”这一概念的重要意义。

它标志着中国 AI 正从出口单个产品,转向出口可以持续生产和交付 Token 化智能的工业体系。


数据背后的产业变化

中国日均 AI Token 用量,据称已经从 2024 年初的千亿级别,增长至 2026 年 3 月的 140 万亿以上。

约两年时间,增长超过 1400 倍。

WAIC 讨论中引用的 OpenRouter 相关数据显示,中国模型的 Token 消耗量约占全球市场的 54.1%,美国模型约为 41.5%。

在编程场景中,中国模型占比据称进一步达到 67.6%。

具体市场份额还会持续变化,但方向已经清晰:

中国模型正在从服务国内用户,逐渐转向参与全球推理流量竞争。


“卖 API”与“卖工厂”,差别不在货,而在商业逻辑

两者销售的最终都可能是 Token。

真正不同的是背后的商业结构。

卖 API 的逻辑

传统 API 模式通常是:

  • 厂商拥有模型;

  • 用户发起请求;

  • 厂商按照用量收费;

  • 供应商制定价格;

  • 用户接受厂商公布的条件。

供给高度集中,价格相对静态,并且主要由供应商单方面决定。

卖 Token 工厂的逻辑

Token 工厂模式更加完整:

  • 将模型、算力、电力与调度能力整合;

  • 持续生产标准化 Token 产能;

  • 用户根据任务需求采购;

  • 多家供应商可以参与生产;

  • 价格可以跟随市场环境变化;

  • 根据真实用量完成结算。

在这一结构中,模型厂商不再等同于整条产业链。

它只是生产线中的一个环节。

Token 本身,成为可以标准化交付的商业产出。

在 WAIC 的“Token 出海”英文展区,参展商展示的并不只是某个 API 地址。

他们展示的是算力如何转化为标准 Token 服务,并跨境交付给海外客户的完整链路。

海外买家也不再只关心哪个模型跑分最高,而更关注:

  • 服务如何接入?

  • 价格如何计算?

  • 算力如何保障?

  • 产品如何交付?

  • 用量如何结算?

从卖 API 到卖工厂,本质上是从:

技术输出

转向:

工业产能输出。


Token 交易市场,是连接工厂与买家的那条路

工厂可以建成,产能也可以提高。

但商业问题依然存在:

  • Token 卖给谁?

  • 价格如何确定?

  • 算力如何路由?

  • 供应商如何回款?

  • 海外用户如何获得服务?

  • 不同供应商的质量如何比较?

这些都属于流通侧问题,而不是生产侧问题。

如果缺少高效的供需市场,Token 工厂可能最终只是一批没有得到充分利用的闲置算力。

这正是 Token 交易市场的重要性。

如果将 Token 比作电力,郑纬民的“三件套”描述的是:

  • 能源来源;

  • 发电厂;

  • 发电成本。

Token 交易市场则类似电力交易系统:

将工厂生产的 Token 高效输送给真正需要它的用户。

市场越活跃,Token 工厂的价值越容易兑现。

工厂需要稳定需求消化产能。

模型厂商需要透明价格规划成本。

低电价只有在规模化运营中,才能真正被摊薄为竞争优势。


Unexhub:搭建 Token 工厂与全球买家之间的交易通道

根据公开定位,Unexhub 希望建设一个 Token 聚合与分发的算力交易中心。

其目标是连接:

  • 模型厂商;

  • 算力供应商;

  • 个人开发者;

  • 企业客户;

  • 海外市场。

平台所提出的多项机制,直接对应 Token 流通和结算环节中的现实难题。


即时结算:让供应商更快收回资金

中小型模型与算力供应商通常需要先投入,才能产生收入。

他们可能需要提前支付:

  • GPU;

  • 云实例;

  • 模型部署;

  • 电力;

  • 网络带宽;

  • 技术运维;

  • 客户支持。

Token 产能交付之后,还可能面临较长的账期。

供应量越大,沉淀在应收账款中的资金也可能越多。

Unexhub 希望按照真实 Token 调用量,提供即时或准即时结算。

理想逻辑是:

供应算力、产生调用,并更快获得相应收益。

更快结算有助于供应商:

  • 减少应收账款;

  • 改善现金流;

  • 将资金重新投入算力;

  • 应对突发需求增长;

  • 降低扩产的资金门槛。

硬件并不会因此免费,运营成本也不会消失。

真正的价值是缩短“已经交付服务”与“获得下一轮扩产资金”之间的时间。


早期零佣金:降低进入市场的成本

Unexhub 还计划为符合条件的早期供应商提供零佣金权益。

供应商进入一个新的国际 Token 市场时,往往面临较高不确定性:

  • 开发者是否会使用相关模型?

  • 哪些地区会产生真实需求?

  • 路由能否满足延迟要求?

  • 当前价格是否可以盈利?

  • 企业客户是否信任平台?

  • 结算系统是否足够稳定?

早期零佣金,可以降低验证这些问题的成本。

供应商可以用更小压力:

  • 完成技术接入;

  • 测试真实需求;

  • 建立调用记录;

  • 积累市场信誉;

  • 验证单位经济模型;

  • 提前获得渠道份额。

其价值并不只是少支付一笔佣金。

而是帮助供应商以更有利的条件,跑通接入、路由、调用、结算与再投入的第一轮商业闭环。


API 竞价:让市场发现 Token 价格

Unexhub 提出的 API 竞价模式,希望让价格根据市场供需动态变化。

当市场算力充足时:

  • 开发者有机会获得更低价格;

  • 供应商可以出售闲置 GPU 产能;

  • 批量任务可以选择经济线路;

  • 集群利用率得到提高。

当市场需求旺盛时:

  • 价格可以适度上浮;

  • 供应商获得增加供给的动力;

  • 紧急企业任务可以获得优先资源;

  • 稀缺算力流向更高价值的需求。

这种机制比单一固定 API 价格更加灵活。

价格可以综合反映:

  • 模型质量;

  • 当前供给;

  • 地区需求;

  • 延迟;

  • 稳定性;

  • 上下文长度;

  • 高峰或低谷时段;

  • SLA 服务要求。

价格由此成为协调供需的信号。


毫秒级路由:让任务匹配合适的模型与算力

Unexhub 同时强调跨模型、跨供应商的毫秒级路由。

根据公开定位,其模型生态覆盖:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • 个人开发者及中小企业提供的其他模型。

不同任务可以根据真实需求分配:

  • 轻量信息抽取使用低成本模型;

  • 复杂推理选择更强模型;

  • 编程任务匹配专业线路;

  • 可延后批量任务等待低价算力;

  • 紧急企业请求选择优先容量。

这可以避免所有请求都经过最昂贵的模型。

开发者只需要为任务真正需要的能力付费。


开发者权益:按真实 Token 用量付费

通过统一 Token 市场,开发者可以按实际用量付费,而不需要先投入大量资金建设自有基础设施。

根据具体模型、线路、可用算力与市场条件,部分服务价格最低可达到常规市场价格的约 3 折。

对于高频开发者和企业团队而言,这可以帮助降低:

  • 编程智能体费用;

  • 长上下文处理成本;

  • 文档分析支出;

  • 客服自动化成本;

  • 数据抽取费用;

  • 批量生成预算;

  • 多模型测试成本。

优惠的真正价值,不只在于某一次 API 调用更加便宜。

它还让团队拥有更大的操作空间:

  • 测试更多模型;

  • 扩大工作负载;

  • 将非紧急任务安排到低价时段;

  • 降低对单一供应商的依赖;

  • 通过动态路由控制预算;

  • 不必在进入生产阶段时立刻自建集群。

只有当低价同时配合稳定服务、透明账单和可预测质量时,折扣才能真正变成长期权益。


供应商权益:一次接入,连接全球需求

对模型与算力供应商而言,平台带来的价值有所不同。

供应商可能拥有:

  • 有竞争力的模型;

  • 专业领域能力;

  • 区域算力资源;

  • 闲置 GPU;

  • 较低电力成本。

但它也可能缺少:

  • 国际销售渠道;

  • 海外开发者流量;

  • 计费基础设施;

  • 跨境结算能力;

  • 多租户 API 管理;

  • 企业客户资源。

通过一次接入,供应商有机会获得更广泛的市场分发。

配合即时结算,可以形成更加直接的路径:

接入算力、接受平台路由,并按照真实 Token 调用获得收益。

对中小供应商而言,这能够降低获客和市场准入成本。

对大型供应商而言,则可以增加一条消化闲置与低谷产能的渠道。


异构算力,让 Token 工厂更加灵活

郑纬民在 WAIC 上还提到,上海正在探索一项异构算力项目,使用约 9000 张国产卡与 1000 张国外卡联合生产 Token。

这种异构部署十分重要,因为不同芯片具备不同优势:

  • 成本;

  • 可获得性;

  • 显存容量;

  • 能耗效率;

  • 模型兼容性;

  • 特定工作负载性能。

Token 工厂没有必要让所有任务都运行在同一种硬件上。

它可以寻找更加高效的组合:

  • 轻量推理使用成本更低的国产加速器;

  • 特殊任务使用国外硬件;

  • 长上下文任务使用大显存节点;

  • 批量任务使用经济型算力;

  • 高优先级请求使用高性能设备。

Token 交易市场可以进一步放大这种优势,因为用户不需要购买某一张具体的卡。

用户购买的是:

  • 某种模型能力;

  • 一定数量的 Token;

  • 一个延迟目标;

  • 一项服务保障。

至于请求由哪种模型、芯片、地区和线路完成,则由 Token 工厂与交易市场寻找更加高效的方案。


串起“三件套”,才会形成完整的出海航路

回到郑纬民提出的框架:

  • 高品质模型是生产技术;

  • Token 工厂是制造体系;

  • 低价电力是能源优势。

三者都非常重要。

但 Token 交易市场,真正将生产连接到了需求。

它需要回答:

  • 应该生产什么?

  • 谁愿意购买?

  • 应该如何定价?

  • 使用哪条线路交付?

  • 供应商如何结算?

  • 海外客户如何获得服务?

当这层流通体系被补齐,Token 才会从一个技术指标,变成可以:

  • 定价;

  • 路由;

  • 调度;

  • 结算;

  • 跨境交付;

的数字商品。

完整路径由此形成:

模型 → Token 工厂 → Token 交易市场 → 全球开发者与企业客户


真正出海的,是标准化智能产能

中国 AI 出海正在发生结构性变化。

目标不再只是把一个应用推向海外,也不只是宣传某一款国产模型。

真正可以出海的,是一整套标准化智能生产与交付体系:

  • 有竞争力的模型;

  • 高效推理基础设施;

  • 低成本能源;

  • 分布式调度;

  • Token 定价;

  • 跨境路由;

  • 金融结算;

  • 企业级服务。

谁能够率先完成“生产—流通—交付”的完整闭环,谁就更有机会在下一阶段全球 AI 竞争中取得关键位置。


Token 或许是新石油,但石油的价值在交易市场

Token 越来越多地被称为 AI 时代的“新石油”。

这个比喻具有一定道理,却并不完整。

石油埋在地下时,商业价值十分有限。

真正创造价值的是:

  • 开采;

  • 炼化;

  • 运输;

  • 定价;

  • 交易;

  • 分发。

Token 同样如此。

高品质模型创造智能潜力。

Token 工厂将潜力变成标准化产出。

低价电力降低生产成本。

Token 交易市场则将生产能力连接至全球真实需求。

WAIC“Token 出海”英文板块前,那些来自中东的开发者,并不只是在询问如何购买某一个中国模型。

他们真正关注的,是一套能够稳定、低成本、规模化生产 Token 的工厂产能。

从卖 API 到卖工厂,不只是一次商业模式切换。

它还意味着 AI 产业链的定价权,正在从单一模型厂商逐渐转向更广泛的市场。

一旦 Token 成为标准化、可交易的服务,能够掌握路由、结算与价格发现的平台,就会在全球 AI 分发体系中占据越来越重要的位置。

Unexhub 提出的即时结算、早期零佣金、API 竞价与毫秒级路由,正是在尝试建设这层市场基础设施。

这套模式最终能否成功,仍取决于稳定性、安全性、合规能力与真实运营表现。

但方向已经越来越清晰:

下一阶段的 AI 全球化竞争,不只取决于谁能造出最强的模型,也取决于谁能以更高效率生产、定价、路由、结算,并向全球市场交付高质量 Token。