Three Cards Played Simultaneously
In July 2026, three developments unfolded almost simultaneously.
On July 9, the first “Guoxin-trained National Model” 10,000-GPU intelligent computing cluster in the Guangdong-Hong Kong-Macao Greater Bay Area was officially launched in Shaoguan, with a total investment of approximately RMB 550 million. One day later, on July 10, the 15th Five-Year Plan for the Information and Communication Industry in Guangdong Province (Draft for Public Comment) was released, explicitly proposing that basic telecommunications companies accelerate the development of ci-yuan (Token) pricing standards.
That same month, Guangzhou Nansha officially unveiled China’s first provincial-level Ci-Yuan Trading and Service Center.
Guangdong was effectively playing three cards at once:
Computing infrastructure. Pricing standards. Trading infrastructure.
Together, they cover the entire emerging value chain—from producing computing power, to measuring and pricing it, and ultimately to trading and settling it.
And Guangdong is not alone.
Looking north to Shaanxi, the National Supercomputing Center in Xi’an completed link scheduling with core nodes of the National Supercomputing Internet in July. Users in Xi’an can now access computing resources equivalent to a 100,000-GPU-scale intelligent computing network without leaving the province. The province’s total intelligent computing capacity has exceeded 3,380P.
A regional structure is gradually taking shape: central Shaanxi as an innovation hub for computing power, southern Shaanxi as a green-computing demonstration zone, and northern Shaanxi as an edge-computing node.
Further north, Inner Mongolia is building another piece of the puzzle.
Horinger New Area has brought together three telecom operators, 16 financial institutions, and 62 computing centers, including infrastructure associated with Huawei, ByteDance, and Kuaishou, representing approximately 150,000P of computing capacity.
On August 17, the “Mobile Engine Confidential Token Service,” jointly developed by China Mobile Intelligent Computing Center and Volcano Engine, was officially launched, creating what was described as a large-scale confidential ci-yuan factory.
Three regions. Three approaches.
But they are all pointing in the same direction:
Computing power is being remeasured.
The Paradigm Shift in Measurement
Changes in measurement units have historically been a precursor to industrial transformation.
During the industrial era, electricity became commercially useful only after society established standardized measurement units such as the kilowatt-hour. Once electricity could be measured consistently, it could be priced, compared, traded, and ultimately transformed into a widely accessible utility.
AI computing power is now undergoing a similar transformation.
On March 24, 2026, Liu Liehong, Director of the National Data Administration, officially introduced “ci-yuan (词元)” as the Chinese term for Token, while announcing that China’s average daily Token consumption had exceeded 140 trillion.
A technical term was becoming a standardized measurement concept.
The significance is not merely linguistic.
Tokens represent a measurable unit of large-model information processing. Whether the underlying service comes from a telecom operator, an AI model provider, or a computing platform, Token provides a common denominator for measuring AI inference consumption.
That creates the foundation for the next step:
standardized pricing.
This is precisely where Guangdong is moving.
Its policy proposal encourages basic telecommunications companies to accelerate the development of ci-yuan pricing standards while cultivating an ecosystem spanning computing power, models, applications, and security.
Measurement and pricing cannot be separated.
Before Token can evolve from a technical concept into a widely traded commodity, the market needs a relatively consistent way to answer a basic question:
How much is one Token worth?
The positioning of the Guangdong Ci-Yuan Trading and Service Center goes even further. It aims to connect the entire chain of:
Computing Power → High-Quality Datasets → Large Models → Ci-Yuan Products → Market Trading
It also addresses supporting infrastructure such as measurement, rights confirmation, compliance, and cross-border circulation.
The larger objective is clear:
Move the digital economy from simply circulating data elements toward settling value through Token consumption.
From “Selling Time” to “Selling Ci-Yuan”
When the unit of measurement changes, the business model inevitably changes with it.
Traditional computing-power leasing follows a simple logic:
Sell computing time.
Rent an H100 eight-GPU server for an hour, a day, or a month. Whether you use 100% of the machine or only 10%, you are still paying for the reserved capacity.
The emerging ci-yuan factory model represents a fundamentally different approach.
Instead of selling access to a physical machine, the system measures the actual computing output in Token units and turns that output into standardized, independently tradable products.
In other words:
Computing power moves from “renting machines” toward “buying Tokens.”
This distinction matters.
For users, there is no need to pay for the idle capacity of an entire server. They purchase the amount of computing consumption they actually need.
For suppliers, computing resources become measurable, standardized, and potentially tradable in a way much closer to conventional commodities.
Horinger has proposed an even more ambitious target: building 10 large-scale standardized ci-yuan factories by 2028, with stable daily production capacity reaching 20 trillion ci-yuan.
Notice the change in language.
The question is no longer simply:
“How many P of computing power do you have?”
It becomes:
“How many Token units can you produce, deliver, and settle?”
Computing “supermarkets” are already being explored in cities such as Horinger, Hangzhou, and Shanghai, while the Computing Bank concept has also entered pilot programs in Xiong'an.
Telecom operators are also experimenting with Token-based packages.
The underlying direction is increasingly clear:
Token is moving toward becoming a common settlement unit connecting computing power, models, applications, and users.
The Missing Piece: Settlement and Clearing
Measurement standards are emerging.
Trading platforms are being built.
Business models are changing.
But one critical layer remains:
Settlement and clearing.
This is where the Unexhub Computing Power Trading Center seeks to establish a differentiated position.
Unexhub is designed as a Token aggregation and distribution computing-power trading center, connecting AI suppliers with developers, enterprises, and other computing consumers.
Its model focuses on three mechanisms.
1. Real-Time Settlement and Clearing
Traditional AI infrastructure can create a significant cash-flow problem for smaller suppliers.
A supplier deploys a model, provides inference capacity, waits for usage to accumulate, and then waits for payment.
Unexhub aims to reverse this cycle:
Supply more → settle more.
Instead of “work first, get paid later,” suppliers can receive corresponding revenue based on actual usage.
For smaller AI providers, this can be particularly valuable because it reduces the amount of working capital tied up in the inference business.
2. Early Zero-Commission Access
Unexhub’s early supplier program is designed around 0% platform commission.
In an ecosystem where routing platforms may charge around 5%–5.5% in service fees, zero commission can allow participating suppliers to retain a larger share of their inference revenue.
That creates room for a second benefit:
more of the supplier-side cost advantage can potentially be passed through to users.
For developers, the value is not simply “a discount.”
It is the combination of more suppliers competing for demand, lower platform friction, and more room for competitive pricing.
3. API Bidding
This is the most important part of the model.
Traditional API pricing is essentially a fixed menu:
Model A = X dollars per million Tokens.
A trading-center model changes the equation.
When supply exceeds demand, suppliers can compete for workloads through lower prices.
When demand spikes and capacity becomes scarce, prices can rise accordingly.
That means developers gain something more important than a fixed discount:
choice.
They can decide whether to prioritize price, latency, model capability, or availability.
Suppliers, meanwhile, gain a mechanism for monetizing otherwise idle capacity.
The result is a transition from a fixed-price catalog toward a market in which supply and demand continuously interact.
From “Exchange Trading” to a Complete Token Market
The Guangdong Ci-Yuan Trading and Service Center is focused on building an ecosystem around market-based Token trading.
Unexhub approaches the problem from another angle:
aggregation + routing + bidding + settlement.
Together, these mechanisms point toward a potentially complete Token circulation infrastructure.
The Requirements for the Operation and Management Capability of Artificial Intelligence Tokens, jointly formulated by the China Academy of Information and Communications Technology and industry participants, has also pushed the industry toward more standardized management across areas such as:
Resource allocation
Service management
Cost management
Operational analysis
Compliance auditing
This is important because a mature Token economy cannot rely solely on model providers publishing individual price lists.
It needs standardized measurement.
It needs transparent pricing.
It needs reliable settlement.
And eventually, it needs credible market data.
That is the infrastructure required to transform Token from a unit used for billing into an actual marketplace commodity.
Why the Economics Matter to Developers
For developers and enterprises, the most immediate value is simple:
you pay for what you actually consume.
Instead of purchasing an entire GPU server, developers can access different AI models through a unified Token-based infrastructure.
Instead of committing to one supplier, they can compare available supply.
Instead of accepting a single fixed API price, they can potentially benefit from competitive market pricing.
And instead of waiting through long settlement cycles on the supply side, participating providers can receive usage-based settlements.
Unexhub is therefore not simply positioning itself as another API gateway.
Its broader proposition is to turn Token consumption into a market-based computing resource.
For developers, that can translate into several practical benefits:
Lower entry costs.
No need to build a GPU cluster before experimenting with AI workloads.
Flexible consumption.
Pay according to actual Token usage instead of paying for idle hardware capacity.
Competitive pricing.
Multiple suppliers can compete for workloads, creating the potential for lower prices.
Millisecond-level routing.
Requests can be routed across available AI providers according to market conditions and service requirements.
Early platform incentives.
Developers can access competitive Token pricing, while early suppliers can participate under the platform’s zero-commission policy.
According to Unexhub’s stated platform model, developers can pay directly by Token, with promotional pricing potentially reaching as low as 30% of prevailing market prices, depending on available supply and bidding conditions.
The important point is that this is not simply a permanent “70% off” coupon.
The real value lies in the market mechanism: when suppliers compete, users can potentially capture part of the efficiency gain.
In Conclusion
Guangdong is building standards and trading infrastructure.
Shaanxi is strengthening supercomputing and integrated scheduling.
Inner Mongolia is building large-scale computing infrastructure and ci-yuan factories.
Three regions.
Three strategies.
One emerging unit:
The Token.
The evolution is becoming increasingly clear:
From technical terminology → standardized measurement
From selling computing time → selling Token consumption
From individual billing → standardized accounting
From fixed pricing → market-based pricing
From isolated computing resources → tradable computing capacity
Once this infrastructure matures, the definition of computing power itself may change.
The industry will no longer focus exclusively on:
“How many P of computing power have been built?”
It may increasingly ask:
“How many Token units can be produced, priced, traded, delivered, and settled?”
That is why ci-yuan has the potential to become the “kilowatt-hour of the AI era.”
The kilowatt-hour transformed electricity from a technical phenomenon into a measurable utility.
Token is attempting to play a similar role for AI intelligence.
And the three cards being played across Guangdong, Shaanxi, and Horinger may ultimately be competing for something even more important than computing capacity:
the right to define how AI intelligence is measured, priced, traded, and settled.
Unexhub is positioning itself within this emerging transition by building a global Token aggregation, distribution, and computing-power trading center.
Rather than relying on a single model provider, the platform brings together AI suppliers across different scales—from major providers such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, to high-quality models supplied by individuals and small and medium-sized AI companies.
The idea is straightforward: aggregate supply, introduce competition, and let enterprises and developers access AI inference at more competitive prices.
For developers, Token-based payment means they only pay for actual consumption. With competitive bidding and multi-provider routing, Unexhub aims to create opportunities to access AI inference at substantially lower prices, with its stated promotional mechanism allowing eligible usage to reach as low as 30% of prevailing market prices.
For suppliers, the value is equally direct: one-click channel access, real-time settlement, and a larger pool of enterprise and developer demand.
The long-term proposition is not simply cheaper APIs.
It is a shift from “buying computing infrastructure” to “trading intelligence consumption.”
And if Token truly becomes the measurement unit of the AI economy, the companies building the infrastructure for its aggregation, distribution, pricing, routing, and settlement may ultimately occupy one of the most important layers of the next-generation computing market.
中文版:三张牌同时打出
2026 年 7 月,三件事情几乎同时发生。
7 月 9 日,粤港澳大湾区首个“国芯训练国家模型”万卡级智算集群在韶关正式启动,总投资约 5.5 亿元人民币。
一天之后,7 月 10 日,《广东省信息通信行业“十五五”规划(征求意见稿)》发布,明确提出鼓励基础电信企业加快完善 “词元(Token)计价标准”。
同月,广州南沙正式揭牌全国首个省级 词元交易服务中心。
广东实际上同时打出了三张牌:
算力基础设施、计价标准、交易基础设施。
这三张牌,覆盖的是一条正在形成的新产业链:
算力生产 → Token 计量与定价 → Token 交易与结算。
而广东并不是唯一一个动作加速的地区。
把视线转向陕西。
7 月,国家超算西安中心与国家超算互联网核心节点完成链路调度。西安用户无需离开陕西,就可以接入相当于 10 万卡级别的超智融合算力资源。
目前,陕西全省智能算力规模已经超过 3,380P。
关中作为算力创新源、陕南作为绿色算力示范区、陕北作为边缘算力节点的格局正在逐步形成。
再往北看,内蒙古也在加速布局。
和林格尔新区已经聚集三大运营商、16 家金融机构以及华为、字节跳动、快手等相关算力中心,共形成约 15 万 P 算力规模。
8 月 17 日,中国移动智算中心与火山引擎联合建设的“移动引擎机密词元服务”正式上线,进一步推动大型 **“词元工厂”**建设。
三个地方。
三种打法。
但指向的是同一个方向:
算力正在被重新计量。
计量单位正在发生范式变化
历史上,每一次计量单位的变化,往往都是产业升级的前奏。
工业时代,电力真正成为一种可以大规模交易的商品,需要先建立统一的计量方式。
“千瓦时”出现之后,电力才能被准确衡量、定价、比较和交易,并最终从一种技术能力变成一种人人可以使用的公共服务。
AI 算力正在经历类似的变化。
2026 年 3 月 24 日,国家数据局局长刘烈宏正式宣布将 Token 命名为“词元”,同时披露中国平均每日词元使用量已经超过 140 万亿。
一个原本属于技术领域的概念,开始进入统一计量体系。
这并不只是名称变化。
Token 本质上提供了一种衡量大模型信息处理量的共同单位。
无论底层服务来自运营商、模型厂商还是算力平台,都可以通过 Token 来衡量 AI 推理消费。
而统一计量之后,下一步自然就是:
统一定价。
这正是广东正在推进的事情。
政策提出鼓励基础电信企业加快完善词元计价标准,同时培育覆盖算力、模型、应用、安全等环节的完整产业生态。
因为没有统一的计量与定价体系,Token 很难从“技术概念”真正变成“可交易商品”。
广东词元交易服务中心的定位更加直接。
它试图打通:
算力 → 高质量数据集 → 大模型 → 词元产品 → 市场交易
同时补齐词元计量、确权、合规、跨境流通等基础能力。
其背后的目标,是推动数字经济从简单的 “数据要素流通”,进一步走向 “词元价值结算”。
从“卖时长”到“卖词元”
当计量单位发生变化,商业模式也必然发生变化。
传统算力租赁的逻辑很简单:
卖算力时间。
租一台 H100 八卡服务器一个小时、一天或者一个月。
无论机器利用率是 100%,还是只有 10%,你支付的都是整台机器的时间成本。
而词元工厂代表的是另一种模式。
它不再单纯出售物理机器的使用权,而是将实际产生的计算结果通过 Token 进行标准化计量,再形成可以独立交易的产品。
换句话说:
算力正在从“租机器”走向“买 Token”。
这个变化非常重要。
对于用户而言,不需要为整台服务器的闲置时间买单,只需要购买自己真正需要的 Token 消耗。
对于供应商而言,算力开始具备更加明显的标准化、可计量、可交易属性。
和林格尔提出的目标更加激进:
到 2028 年建设 10 个大型标准化词元工厂,形成稳定的每日 20 万亿词元产能。
注意这里的语言已经发生变化。
过去问的是:
“你有多少 P 算力?”
未来可能越来越多地问:
“你每天能够生产、交付和结算多少 Token?”
目前,和林格尔、杭州、上海等地已经在探索算力“超市”。
“算力银行”也已经进入雄安试点。
运营商也在不断推出 Token 套餐。
整个行业正在形成一个越来越清晰的方向:
Token 正在成为连接算力、模型、应用和用户的通用结算单位。
最后一公里:结算与清算
计量标准开始建立。
交易平台开始出现。
商业模式正在变化。
但还有一个关键环节:
结算与清算。
这正是 Unexhub 算力交易中心希望切入的位置。
Unexhub 将自身定位为 Token 聚合与分发的算力交易中心,连接 AI 供应商、开发者和企业用户。
其核心机制包括三层。
1. 实时结算与清算
传统 AI 服务模式容易给中小供应商带来现金流压力。
供应商部署模型、提供推理能力,然后等待调用量积累,最终再进入结算周期。
Unexhub 希望改变这一循环:
供多少,结多少。
也就是说,不再是“先干活、后收款”,而是根据实际调用量实时获得相应收益。
对于中小 AI 供应商而言,这能够减少资金长期沉淀在推理服务中的压力。
2. 早期零佣金
Unexhub 的早期供应商机制强调 0 佣金。
如果行业部分路由平台通常收取约 5%-5.5% 的平台服务费,那么早期零佣金意味着供应商可以保留更多推理收入。
更重要的是,这种成本空间有机会进一步传导给终端用户:
供应商成本降低 → 竞争空间扩大 → 用户获得更具竞争力的价格。
因此,对开发者而言,价值并不只是“打折”。
真正重要的是:
更多供应商、更多竞争、更低交易摩擦,以及更大的价格优化空间。
3. API 竞价模式
这是整个模式中最关键的一环。
传统 API 定价是一张固定价目表:
模型 A = 每百万 Token X 美元。
而算力交易中心希望改变这种逻辑。
当供给超过需求时,供应商可以通过更低价格竞争订单。
当需求快速增加、算力供不应求时,价格则可能上涨。
开发者因此获得的不只是一个固定折扣,而是一种更重要的东西:
选择权。
可以根据自己的业务需要,在价格、速度、模型能力和稳定性之间进行选择。
供应商也能够通过竞价,把原本闲置的算力转化为收入。
AI 推理因此从:
固定价格目录
逐步走向:
供需实时匹配的市场。
从“交易中心”走向完整 Token 市场
广东词元交易服务中心正在探索以 市场化交易为核心的 Token 生态。
而 Unexhub 的路径则更加偏向:
聚合 + 路由 + 竞价 + 结算。
两者共同指向的是一套更加完整的 Token 流通基础设施。
由中国信息通信研究院与行业企业共同制定的《人工智能词元运营管理能力要求》已经围绕:
资源配置
服务管理
成本管理
运营分析
合规审计
等方向推动行业标准化。
这意味着,成熟的 Token 经济不能永远停留在“每家模型厂商自己公布一张价格表”。
它需要统一计量。
需要透明定价。
需要可靠结算。
最终还需要可信的市场数据。
只有这些基础设施逐步成熟,Token 才有可能从一个单纯的计费单位,进一步成为真正意义上的:
市场化算力商品。
对开发者而言,真正的价值是什么?
对于开发者和企业来说,最直接的价值其实很简单:
用多少,付多少。
过去,为了使用 AI,企业可能需要采购 GPU、建设机房、部署模型。
现在,开发者可以通过统一的 Token 基础设施直接获得模型推理能力。
不必购买整台 GPU 服务器。
不必把所有需求绑定到单一供应商。
也不必完全接受某一家厂商的固定 API 价格。
当多个供应商同时竞争同一批需求时,用户就可能获得更加有竞争力的价格。
因此,Unexhub 的实际价值可以概括为几个方面:
降低进入门槛
无需先建设 GPU 集群,就可以开始使用 AI 算力。
按需消费
按照实际 Token 使用量付费,而不是为闲置硬件买单。
供应商竞争
多家供应商同时参与,为用户提供更多价格选择。
毫秒级路由
根据可用供应商、市场价格以及业务需求进行快速路由。
早期平台权益
早期供应商可以享受零佣金机制;开发者则能够在多供应商竞争环境中获得更具竞争力的 Token 价格。
按照 Unexhub 当前的商业模式,开发者可以按 Token 实际消费付费,在具体供应与竞价条件下,平台宣传的优惠价格最低可达到 市场价格的 3 折。
这里真正值得关注的,并不是简单的“永久 7 折优惠”。
更重要的是背后的市场机制:
当供应商之间开始竞争,用户可以获得其中一部分效率红利。
结语
广东正在制定标准、建设交易中心。
陕西正在强化超算与融合调度。
和林格尔正在建设大型算力基础设施和词元工厂。
三个地方。
三种策略。
最终都指向同一个新单位:
Token。
这条演进路线正在越来越清晰:
从技术概念 → 标准化计量
从出售算力时长 → 出售 Token 消费
从独立计费 → 统一核算
从固定价格 → 市场化定价
从孤立算力资源 → 可交易算力能力
一旦这套基础设施逐步成熟,算力本身的定义也可能发生变化。
行业关注的重点将不再只是:
“建设了多少 P 算力?”
而可能越来越关注:
“生产、定价、交易、交付和结算了多少 Token?”
这也是为什么:
词元可能成为 AI 时代的“千瓦时”。
千瓦时让电力从一种技术能力变成可以被精确计量、定价和使用的公共资源。
而 Token 正在尝试为 AI 智能完成类似的事情。
广东、陕西和和林格尔同时打出的三张牌,真正争夺的也许并不仅仅是算力规模。
更重要的是:
谁能够定义 AI 智能如何被计量、定价、交易和结算。
而 Unexhub 正是在这一变化中,尝试建立一个 全球 Token 聚合、分发与算力交易中心。
平台并不局限于单一模型供应商,而是汇聚不同规模的 AI 供应商,包括 Gemini、ChatGPT、Claude、DeepSeek、GLM 等大型模型生态,同时也吸纳个人开发者以及中小 AI 企业提供的优质模型。
其核心逻辑是:
聚合供给 → 引入竞争 → 智能路由 → 降低交易成本。
对于开发者而言,可以按照实际 Token 消耗付费,不必提前购买昂贵的 GPU 资产;通过多供应商竞争和 API 竞价,有机会获得更加优惠的调用价格。
按照 Unexhub 当前的优惠机制,在符合具体供应和竞价条件的情况下,开发者最低可享受 市场价格 3 折的 Token 调用权益。
对于供应商而言,则可以获得:
一键接入渠道、实时结算、调用分成以及更广泛的企业与开发者需求。
所以,Unexhub 真正试图推动的并不只是:
“更便宜的 API。”
而是:
从购买算力基础设施,走向交易 AI 智能消费。
如果 Token 最终成为 AI 经济时代的核心计量单位,那么围绕 Token 建立的聚合、分发、定价、路由、竞价与结算基础设施,也可能成为下一代算力市场中最重要的基础设施层之一。