
Four Token Factories Arrived in One Week: AI Compute Has Reached Its “Kilowatt-Hour Moment”
Bilingual Article|English first, followed by the complete Chinese edition below
英中双语内容|英文在前,中文完整版本请继续向下阅读
During the first week of July 2026, China’s AI compute industry crossed an important threshold.
On July 1, the Nanjing Token Collection Store officially launched at the Qilin Science and Technology Innovation Park, becoming the city’s first one-stop service platform for Token-based compute.
On the same day, Guizhou Mobile’s Token Factory reportedly exceeded 10 billion Token calls per day, while officials in Beijing’s Yizhuang district disclosed that the first phase of the Beijing No. 1 Token Factory had reached a daily production capacity of more than 1.4 trillion Tokens.
On July 3, the Shaanxi Silk Road Data Exchange and Inspur Cloud announced plans to jointly build the Silk Road Token Factory. Beijing Data Group also introduced the JingSuan Token Factory, described as Beijing’s first state-owned Token factory.
On July 6, the Beijing No. 1 Token Factory formally entered production.
By July 10, the Wuxiang Cloud Valley Intelligent Computing Center announced production capacity of approximately 200 billion Tokens per hour, becoming one of China’s first commercially operational Token factories.
Within a single week, Nanjing, Beijing, Xi’an and Nanning all advanced major Token-factory projects.
This was not a coincidence.
It marked the moment when the AI industry’s transition from selling GPUs and compute hours to selling standardized Token output began moving from theory into real infrastructure.
From a Hardware Supermarket to a Standardized Output Factory
To understand the significance of the Token factory, it helps to compare it with the traditional intelligent-computing center.
In the past, compute centers operated much like hardware supermarkets.
They supplied raw infrastructure such as:
GPU servers;
Storage systems;
Networking equipment;
Rack capacity;
Reserved compute instances;
GPU-hour leasing.
Enterprises purchased access to the hardware, but still needed to handle much of the difficult work themselves:
Deploying models;
Configuring inference frameworks;
Optimizing performance;
Managing energy consumption;
Scheduling workloads;
Monitoring clusters;
Maintaining service stability.
What customers bought was not a completed AI capability.
They bought the right to use infrastructure for a period of time.
The Token factory changes the delivery model.
Instead of selling access to GPUs, it delivers standardized AI output through APIs. Enterprises can purchase capabilities such as:
Question answering;
Summarization;
Code generation;
Model inference;
Document analysis;
Agent execution;
Multimodal generation.
Settlement is based on actual Token consumption.
The customer no longer needs to know whether the underlying system uses H100s, H200s or B200s, where the data center is located or how its cooling system works.
The questions become simpler:
How much does one million Tokens cost?
How reliable are those Tokens?
How much useful work can they complete?
Compute is shifting from an asset to a service—and from physical hardware to measurable intelligent output.
Wuxiang Cloud Valley Pushes the Utility Model Further
The Wuxiang Cloud Valley project illustrates how far this logic can go.
Customers do not need to lease racks, build clusters or operate inference infrastructure. They access model capabilities through APIs and pay according to actual Token consumption.
The service begins to resemble a public utility:
Connect when needed, consume on demand and pay according to measured usage.
Its infrastructure reportedly combines dedicated 220 kV substations with microchannel liquid cooling, converting electricity into usable AI output as efficiently as possible.
The project has reported cluster utilization of approximately 57%, compared with industry averages that often remain below 30%.
That difference matters.
A compute center with thousands of GPUs but weak utilization is not necessarily producing more value than a smaller, highly optimized cluster.
The Token factory reframes the performance question.
Instead of asking:
How many GPUs do you own?
The market begins asking:
How many reliable Tokens can you produce from each GPU, each watt and each unit of capital?
Why Is This Happening Now?
Two opposing data points explain why Token factories are emerging so quickly.
Demand Is Expanding at Extraordinary Speed
In early 2024, China’s average daily Token consumption was still measured in the billions.
By May 2026, it had reportedly exceeded 170 trillion Tokens per day, representing growth of more than one thousand times in just over two years.
During the first week of July 2026, Chinese model usage reportedly reached approximately 23.45 trillion Tokens, maintaining the country’s position as the world’s largest Token consumer for ten consecutive weeks.
The source of demand has also changed.
AI is no longer used only for isolated chatbot questions. Tokens are now consumed by:
Coding agents;
Enterprise automation;
Intelligent customer service;
Scientific research;
Industrial simulation;
Multimodal generation;
Long-running autonomous workflows.
One Agent task may consume dozens or hundreds of times more Tokens than an ordinary conversation.
Supply Remains Highly Underutilized
At the same time, many domestic intelligent-computing centers report average rack utilization of only 20%–30%.
Some enterprise-oriented centers operate below 10%.
The contradiction is clear:
Model providers report insufficient capacity;
GPU leasing prices remain high;
Chip shortages continue;
Expensive infrastructure still sits idle.
The industry does not simply lack compute.
It lacks an efficient mechanism for turning fragmented, idle hardware into standardized, easily purchased AI output.
Owning compute does not automatically create productive capacity.
Token factories are designed to close that gap.
The JingSuan Token Factory Shows What Optimization Can Change
The JingSuan Token Factory provides a useful example of how software and scheduling can transform the economics of existing hardware.
Through heterogeneous compute pooling and unified scheduling, it reportedly increased resource utilization from an industry average of 30%–40% to more than 80%.
Its prefill–decode disaggregated scheduling architecture reportedly improved throughput by as much as 90%.
For long-context workloads, the cost per Token was reduced by more than 50%.
These improvements do not come from simply purchasing twice as many GPUs.
They come from operating existing resources more efficiently:
Assigning workloads to the most suitable hardware;
Separating different inference stages;
Improving batching;
Managing memory more effectively;
Reducing idle intervals;
Scheduling long and short tasks separately;
Reusing cached context.
When utilization doubles and Token costs fall by half, GPU-hour pricing becomes increasingly inadequate.
A GPU-hour measures how long hardware was occupied.
It does not measure how much useful intelligence the hardware produced.
AI Compute Has Reached Its “Kilowatt-Hour Moment”
The history of electricity offers a useful comparison.
In the early industrial era, many factories generated their own electricity.
Power was tied to local equipment. Businesses needed to purchase generators, operate them and accept the risks of inefficient production.
Electricity became a scalable commodity only after the rise of alternating current, standardized grids and kilowatt-hour pricing.
Consumers no longer cared which generator produced the electricity.
They cared about:
Availability;
Reliability;
Unit price;
Actual consumption.
AI compute is beginning to experience the same transition.
Jensen Huang’s “Token factory” concept reframes data centers as production facilities whose output is not raw FLOPS, but Tokens.
Industry performance therefore moves from peak compute toward metrics such as:
Tokens per second;
Tokens per watt;
Cost per million Tokens;
Successful tasks per unit of compute;
Useful output per dollar.
In China, the standardization of Token terminology as 词元 further supports a common framework for measurement, pricing and settlement.
The Token is evolving from a technical unit into an industrial unit of account.
It is AI compute’s equivalent of the kilowatt-hour.
Four Cities, Four Entry Points, One Direction
The concentrated launch of Token projects across several cities shows that this is no longer one company’s isolated experiment.
Each region is entering from a different part of the value chain.
Nanjing: Aggregation and Enterprise Access
The Nanjing Token Collection Store focuses on helping enterprises, universities and research institutions access standardized compute and model services through one platform.
Its value lies in reducing the need for small and medium-sized organizations to build expensive clusters themselves.
Beijing: Production and Distribution
Beijing’s projects emphasize industrial-scale Token production, heterogeneous compute scheduling and standardized API delivery.
The objective is to turn large pools of infrastructure into continuously available model output.
Shaanxi: Pricing and Settlement Exploration
The Silk Road Token Factory plans to explore standardized Token pricing and transaction settlement, connecting western-region compute resources with broader model demand.
Guangxi: Commercial Operations
Wuxiang Cloud Valley focuses on operating the Token-factory model commercially—transforming power, cooling, servers and model infrastructure into API products sold according to actual Token consumption.
Different cities, different operators and different entry points are converging on the same model:
Compute becomes standardized Token output, and Token output becomes a tradeable service.
From “Factory” to “Market”
Token factories solve the production problem.
They convert fragmented infrastructure into measurable, standardized AI output.
But production alone is not enough.
Tokens also need a market in which they can:
Move between suppliers and users;
Be compared across models;
Respond to supply and demand;
Settle quickly;
Be routed across regions;
Maintain service continuity.
Without a market, every Token factory remains an isolated island.
A developer may still need separate accounts, contracts, APIs and balances for each supplier.
Prices may remain fixed even when large amounts of capacity are idle.
Suppliers may generate Tokens today but wait months for enterprise settlement.
The next stage of AI compute infrastructure is therefore not another factory.
It is a market connecting the factories.
Unexhub: Building the Market Layer Above Token Factories
This is the gap Unexhub aims to address.
Unexhub is building a Token aggregation, distribution and compute-trading center that connects model providers, compute suppliers, enterprises and developers.
Its role is not to replace Token factories.
It is to provide the market infrastructure that allows their output to circulate.
The platform is designed around several core capabilities:
Unified access to multiple models;
Token-based metering;
Millisecond-level intelligent routing;
Dynamic supplier selection;
API bidding;
Real-time clearing and settlement;
Centralized usage and billing.
Unexhub is bringing together model ecosystems such as Gemini, ChatGPT, Claude, DeepSeek and GLM, while also allowing individual developers, smaller companies and specialized model suppliers to contribute high-quality capacity.
The aim is to turn Token supply from scattered endpoints into a connected marketplace.
Real-Time Settlement Changes Supplier Economics
In the traditional API business, suppliers often pay inference costs immediately but collect revenue much later.
Enterprise customers may settle:
Monthly;
Quarterly;
After project acceptance;
Under longer framework agreements.
During that period, suppliers continue paying for GPUs, electricity, networking and operations.
Unexhub’s real-time settlement model aims to shorten this cycle.
When a Token is consumed, the corresponding supplier revenue can enter settlement at the same time.
For suppliers, this can provide several practical benefits:
Faster cash conversion;
Lower working-capital pressure;
Reduced exposure to long payment terms;
Faster monetization of idle compute;
More predictable revenue sharing;
Easier expansion for smaller providers.
The mechanism changes the Token from a usage record settled later into a measurable unit of value that can be cleared as it is consumed.
Each API request becomes both an inference event and a financial event.
API Bidding Lets Prices Reflect Real Supply and Demand
Most model APIs still use fixed price lists.
The same model usually costs the same at midnight and during the busiest hour of the workday.
But compute markets are not static:
Capacity is often underused at night;
Business hours create concentrated demand;
New model releases generate traffic surges;
Some regional routes become congested;
Suppliers may suddenly release additional capacity.
Unexhub is developing an API bidding mechanism that allows prices to move with real-time supply and demand.
The model resembles peak and off-peak electricity pricing.
Developers Can Benefit During Off-Peak Periods
Lower rates can be used for workloads such as:
Batch inference;
Overnight code review;
Data labeling;
Background agents;
Large-scale content generation;
Non-real-time research.
Suppliers Can Capture More Value During Demand Peaks
When high-quality capacity becomes scarce, suppliers can receive returns that better reflect actual market demand.
Price becomes a live signal of:
Capacity;
Latency;
Model quality;
Reliability;
Demand.
Developer Benefits: More Models, Lower Entry Cost and Flexible Routing
Through a unified interface, developers can access multiple model providers without separately maintaining:
API keys;
Billing accounts;
Integration logic;
Usage limits;
Failover rules;
Settlement records.
Unexhub’s routing layer can select resources based on:
Task type;
Model capability;
Price;
Latency;
Available capacity;
Historical success rate;
Context requirements.
For selected models and routes, pricing may be available at approximately 30% of corresponding official or market rates, depending on supply conditions, route quality and commercial arrangements.
The practical value goes beyond a cheaper individual request.
It gives developers and enterprises:
More Agent runs within the same budget;
More room to compare models;
Larger batch-processing capacity;
Lower dependence on a single supplier;
Automatic switching during congestion or limits;
More freedom to schedule work during cheaper periods.
The platform’s goal is not merely to find the lowest Token price.
It is to reduce the total cost of producing a successful result.
Supplier Benefits: One Connection, Broader Distribution
For model and compute suppliers, Unexhub offers a unified distribution channel.
A supplier can connect once, while the platform supports:
Developer and enterprise traffic;
Protocol adaptation;
Usage metering;
Intelligent routing;
Revenue reporting;
Dynamic pricing;
Clearing and settlement.
Suppliers do not need to build a complete global sales, billing and customer-acquisition system from zero.
Smaller providers can turn underused capacity into revenue, while high-quality routes can receive more traffic based on real performance.
Qualifying early suppliers may also receive more favorable onboarding terms, including promotional channel fees as low as zero commission in selected programs.
The benefit is not only lower platform cost.
It allows suppliers to:
Test market demand with less financial pressure;
Build early usage records;
Validate pricing;
Reach enterprise customers;
Reinvest more revenue into model and infrastructure improvements.
The Market Must Optimize Task Cost, Not Only Token Price
A mature Token market cannot treat every Token as identical.
One model may offer a low unit price but require repeated retries.
Another may cost more but complete the task in one pass.
A route may be inexpensive but unstable during peak periods.
Another may be more expensive but provide lower latency and a higher success rate.
Therefore, intelligent routing must consider:
Unit price;
Model quality;
Total Token consumption;
Retry probability;
Latency;
Capacity;
Context length;
Task success rate.
The final objective should be:
The lowest total cost for a task that is actually completed.
This is the difference between a simple API marketplace and a true AI compute exchange.
The Full Closed Loop: Production, Circulation and Settlement
The concentrated arrival of Token factories is only the beginning.
A mature AI compute economy requires a complete loop:
Electricity and hardware produce compute;
Models convert compute into Tokens;
Token factories standardize the output;
Exchanges aggregate and route supply;
Developers consume Tokens through APIs;
Payments settle back to suppliers;
Revenue funds additional compute production.
In this system, every participant can play more than one role.
A model provider may sell Tokens while buying compute.
A compute supplier may sell GPU capacity while purchasing model APIs to improve its own services.
An enterprise may buy external models while supplying its own vertical model or unused inference capacity.
The market becomes a multi-sided network rather than a one-way supply chain.
Conclusion: The Factory Is Built—Now the Market Must Make Tokens Flow
Four Token-factory projects landing within one week is not the destination.
It is evidence that AI compute has reached a new stage of industrialization.
The industry is moving:
From hardware to services;
From GPU-hours to Tokens;
From peak compute to Tokens per watt;
From fragmented clusters to standardized output;
From fixed APIs to dynamic markets.
The Token factory solves the production problem.
The next challenge is circulation.
Tokens need market infrastructure for aggregation, intelligent routing, dynamic pricing and real-time settlement.
That is the role Unexhub aims to play.
By connecting model and compute suppliers, enabling millisecond-level routing, developing API bidding and supporting real-time revenue settlement, Unexhub is attempting to create the market layer above the new generation of Token factories.
For developers, selected routes priced near 30% of comparable market or official rates can create more room for agents, batch workloads and model experimentation.
For suppliers, unified access, real-time revenue sharing and early-stage low-commission programs can shorten the path from idle capacity to sustainable cash flow.
Electricity became a universal commodity when the kilowatt-hour connected power plants to users through a shared grid and market.
AI compute is now approaching the same moment.
Token factories produce standardized intelligence. Compute exchanges allow that intelligence to move.
The true “kilowatt-hour moment” will arrive not merely when Tokens can be produced at scale, but when they can be priced, routed, traded and settled as naturally as electricity itself.
中文版
一周落地四座词元工厂:AI 算力的“度电时刻”已经到来
2026 年 7 月第一周,AI 算力产业跨过了一道重要门槛。
7 月 1 日,南京 Token 集合店在麒麟科技创新园正式上线,成为南京首个算力词元一站式服务载体。
同一天,贵州移动词元工厂日均调用量突破 100 亿词元;北京亦庄披露,北京一号词元工厂一期日产能力已超过 1.4 万亿词元。
7 月 3 日,陕西丝路数据交易中心与浪潮云宣布联合建设丝路词元工厂;北京数据集团也发布了北京首个国资词元工厂——京算词元工厂。
7 月 6 日,北京一号词元工厂正式投入生产。
7 月 10 日,五象云谷智算中心宣布,每小时可生产约 2000 亿词元,成为国内较早进入商业运营阶段的词元工厂之一。
短短一周,南京、北京、西安、南宁四座城市集中推动词元工厂落地。
这并不是偶然。
它标志着 AI 产业从卖显卡、卖卡时走向生产词元、销售标准化智能产出的转型,开始从概念进入真实基础设施。
从“硬件超市”走向“标准件工厂”
要理解词元工厂的意义,首先要看它与传统智算中心有什么不同。
过去的智算中心,更像一座硬件超市。
平台提供的是:
GPU 服务器;
存储设备;
网络设施;
机柜;
云实例;
卡时租赁。
企业采购了这些底层资源后,往往还需要自己完成大量复杂工作:
部署模型;
配置推理框架;
调整性能;
管理能耗;
调度任务;
监控集群;
维护服务稳定性。
企业买到的并不是可以直接使用的 AI 成品能力,而是一段时间内使用硬件的权利。
词元工厂则改变了交付方式。
它不再主要出售 GPU 使用权,而是通过 API 交付标准化智能产出。企业可以按需采购:
智能问答;
文档摘要;
代码生成;
模型推理;
数据分析;
Agent 执行;
多模态生成。
最终按照实际词元消耗量结算。
客户不必再关心底层使用的是 H100、H200 还是 B200,也不必研究机房建在哪里、冷却系统如何运行。
客户关心的问题变成:
每百万词元多少钱?
这些词元质量是否稳定?
它们能完成多少有效工作?
算力正在从固定资产变成服务,也正在从硬件资源变成可以计量的智能产出。
五象云谷把“公用事业模式”推得更远
五象云谷的实践,将这一逻辑进一步推向了商业化。
客户无需租机柜、建设集群,也不必自己维护推理环境,只需要通过 API 获取模型能力,再按照实际消耗的词元付费。
这套服务越来越像水电基础设施:
需要时接入,使用时消耗,最后按表结算。
通过 220 千伏专用变电站、微通道液冷等基础设施,五象云谷尝试将更多电力转化为有效 AI 产出。
其集群利用率据称达到约 57%,而行业平均水平往往不足 30%。
这一差距非常关键。
拥有几千张 GPU、但长期低利用率的智算中心,未必比一座规模较小、调度效率更高的集群创造更多价值。
词元工厂正在重新定义评价方式。
过去人们问:
你拥有多少张 GPU?
未来市场更关心:
每张 GPU、每瓦电力和每一元投入,究竟能够生产多少稳定、可用的词元?
为什么偏偏是现在?
需求端与供给端的两组数据,解释了词元工厂为何会集中出现。
需求正在高速膨胀
2024 年初,中国大模型日均词元消耗量仍以十亿级计算。
到了 2026 年 5 月,据称已突破每日 170 万亿词元,两年多增长超过千倍。
2026 年 7 月第一周,中国 AI 模型调用量约达 23.45 万亿词元,连续十周位居全球首位。
更重要的是,需求结构已经改变。
AI 不再只用于偶尔回答几个问题,而是进入:
编程 Agent;
企业自动化;
智能客服;
科研分析;
工业仿真;
多模态生成;
长时间自主工作流。
一个 Agent 任务消耗的词元,可能是普通对话的几十倍甚至上百倍。
供给端却仍然大量闲置
与此同时,国内不少智算中心的平均机柜利用率仍然只有 20%—30%。
部分企业级中心甚至不足 10%。
于是,产业出现了明显矛盾:
模型厂商频繁表示算力不足;
GPU 租赁价格持续高位;
芯片供给仍然紧张;
大量昂贵设备却在数据中心里闲置。
行业缺少的不只是算力总量。
更缺少一种把分散、闲置硬件转化为标准化、可购买 AI 产出的机制。
有算力,不等于有生产力。
词元工厂正是为了填补这道缺口。
京算词元工厂证明:优化软件也能改变算力经济学
京算词元工厂展示了软件和调度如何重新释放现有硬件价值。
通过异构算力池化和统一调度,其资源利用率据称从行业常见的 30%—40% 提升至 80% 以上。
通过 Prefill–Decode 分离的智能调度架构,算力吞吐最高可提升约 90%。
在长上下文场景中,单词元调用成本降低超过 50%。
这些效率并不是依靠简单采购两倍 GPU 获得的。
它来自更精细的资源运营:
把不同任务分配到适合的硬件;
拆分推理阶段;
优化批处理;
提高显存管理效率;
减少设备空闲;
区分长短任务;
复用缓存上下文。
当利用率翻倍、单词元成本腰斩,“卡时”就越来越无法准确反映真实价值。
卡时只能告诉用户硬件被占用了多久。
它无法告诉用户,这段时间究竟生产了多少有效智能。
AI 算力迎来了自己的“度电时刻”
电力产业的发展历史,为 AI 算力提供了一个非常清晰的参照。
工业化早期,很多工厂需要自己发电。
电力与本地设备深度绑定,企业必须购买发电机、自己维护,并承担低效率生产的风险。
直到交流电、电网和“度电”计价方式出现,电力才真正成为标准化商品。
用户不再关心电由哪一台发电机生产,只关心:
是否随时可用;
是否稳定;
每度电多少钱;
实际用了多少。
AI 算力正经历相似变化。
黄仁勋提出的“Token 工厂”概念,将数据中心重新定义为持续生产词元的工业设施。
产业竞争也从峰值算力,逐渐转向:
每秒词元数;
每瓦词元数;
每百万词元成本;
单位算力完成的成功任务数量;
每元投入获得的有效输出。
国内对 Token 的“词元”标准化命名,也进一步为计量、定价和结算提供了统一基础。
词元正在从技术术语变成产业记账单位。
它就是 AI 算力世界里的“度电”。
四座城市,四种切入方式,同一个方向
多地词元项目集中落地,说明它已经不是某一家企业的孤立实验。
不同城市正从产业链的不同环节进入。
南京:聚合交易与企业接入
南京 Token 集合店重点帮助企业、高校和科研院所,通过一个平台获取标准化算力与模型服务。
中小企业无需自己建设昂贵集群,也可以按需采购词元。
北京:生产与分发
北京相关项目更强调规模化词元生产、异构算力调度和标准 API 成品交付。
目标是把大规模硬件资源持续转化为可随时调用的智能产出。
陕西:标准化定价与交易结算
丝路词元工厂计划探索词元标准化定价和结算机制,将西北地区算力资源与更广泛的模型需求连接起来。
广西:商业化运营
五象云谷重点验证词元工厂的商业模式,将电力、冷却、服务器和模型基础设施转化为按照实际词元消耗出售的 API 产品。
城市不同、运营方不同、切入角度不同,最终却汇聚到同一条路径:
算力被转化为标准化词元,词元再成为可以购买和交易的服务。
从“工厂”走向“市场”
词元工厂解决的是生产问题。
它把分散硬件转化为可计量、标准化的 AI 产出。
但只有生产还不够。
词元还需要一个市场,才能实现:
在供应商与用户之间自由流动;
跨模型比较;
根据供需形成价格;
快速结算;
跨区域路由;
保持业务连续性。
如果缺少市场,每一座词元工厂仍然是一座孤岛。
开发者可能仍要为不同供应商分别注册账户、签订协议、维护 API 和充值余额。
即使闲时存在大量空闲算力,价格仍可能全天固定。
供应商今天生产了词元,却可能要等数月才能收到企业客户款项。
因此,AI 算力基础设施的下一站,不是再建一座工厂。
而是建设一座能够连接所有工厂的市场。
Unexhub:建设词元工厂之上的市场层
这正是 Unexhub 希望填补的空白。
Unexhub 正在建设 Token 聚合、分发与算力交易中心,连接模型厂商、算力供应商、企业与开发者。
它不是要取代词元工厂,而是为工厂生产的词元提供可以流通的市场基础设施。
平台围绕以下能力展开:
多模型统一接入;
按 Token 计量;
毫秒级智能路由;
动态供应商选择;
API 竞价;
实时清结算;
统一调用与账单管理。
Unexhub 正在聚合 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型资源,也为个人开发者、中小企业和垂直模型供应商提供优质能力的接入渠道。
目标是把分散的 Token 供给,从一个个孤立接口,连接成统一交易市场。
实时结算,改变供应商的现金流逻辑
传统 API 业务中,供应商需要先支付推理成本,却往往很晚才能获得收入。
企业客户可能采用:
月结;
季结;
项目验收后付款;
更长周期的框架协议结算。
在等待期间,供应商仍需持续支付 GPU、电力、网络和运维成本。
Unexhub 的实时结算模式,希望缩短这一周期。
当 Token 被消耗时,对应的供应商收益可以同步进入结算流程。
对供应商而言,这能够带来:
更快的资金回笼;
更低的垫资压力;
减少长账期风险;
更快变现闲置算力;
更稳定的调用分成;
让中小供应商更容易扩大业务。
Token 因此不再只是月底统一统计的使用记录,而逐渐变成消费时即可计量和清算的价值单位。
每一次 API 调用,既是一次推理行为,也是一笔资金流转。
API 竞价,让价格真正反映供需
目前大多数模型 API 仍采用固定价格。
同一款模型在深夜和工作日最高峰,通常执行相同报价。
但真实算力市场并不稳定:
夜间可能存在大量空闲资源;
工作时间请求高度集中;
新模型发布可能带来流量高峰;
某些地区线路临时拥堵;
供应商可能突然释放额外容量。
Unexhub 正在筹备 API 竞价机制,让价格随着实时供需变化。
其逻辑类似电力市场的峰谷电价。
开发者可以利用闲时低价
更低的价格适合安排:
批量推理;
夜间代码审查;
数据标注;
后台 Agent;
大规模内容生成;
非实时研究。
供应商可以在高峰期获得更高收益
当优质容量稀缺时,供应商可以获得更符合真实需求的回报。
价格不再只是静态价目表,而成为反映:
容量;
延迟;
模型质量;
稳定性;
市场需求;
的实时信号。
对开发者:多模型、低门槛与灵活路由
开发者通过统一接口,可以调用多家模型,无需分别维护:
API Key;
充值账户;
接入逻辑;
使用限额;
故障切换策略;
多张账单。
Unexhub 的路由系统可以根据以下条件选择资源:
任务类型;
模型能力;
价格;
延迟;
当前容量;
历史成功率;
上下文要求。
部分模型和线路在特定供需、线路质量和合作方案下,价格最低可达到相应市场或官方标准价格的约 30%。
这项权益真正带来的,不只是一次请求更便宜,而是:
相同预算运行更多 Agent;
获得更多模型比较空间;
扩大批量处理规模;
减少对单一供应商的依赖;
在限额或拥堵时自动切换;
将任务安排在价格更低的时段。
平台最终要寻找的,也不只是最低 Token 单价。
而是一个成功任务的最低综合成本。
对供应商:一次接入,获得更广分发
对模型与算力供应商而言,Unexhub 提供的是统一商业分发渠道。
供应商一次接入后,平台可以协助处理:
企业与开发者流量;
API 协议适配;
Token 计量;
智能路由;
收益统计;
动态定价;
财务清结算。
供应商无需从零建设全球销售、计费和客户获取体系。
中小供应商可以把闲置容量转化为收入,优质线路则可以根据真实表现获得更多调用流量。
符合条件的早期供应商,还可能获得更有吸引力的接入政策,部分计划可享受阶段性最低 0 佣金。
它带来的价值不只是节省一笔平台费用,还包括:
以更低成本验证市场需求;
积累早期调用数据;
测试模型定价;
触达企业客户;
将更多收入投入模型与算力优化。
市场最终要优化的,是任务成本而不是 Token 单价
成熟的 Token 市场不能把所有词元视为完全同质的商品。
一款模型可能单价很低,却需要反复重试。
另一款价格更高,却能一次完成。
某条线路闲时便宜,高峰期却不稳定。
另一条线路报价稍高,却拥有更低延迟和更高成功率。
因此,智能路由需要综合考虑:
单价;
模型质量;
总 Token 消耗;
重试概率;
延迟;
容量;
上下文长度;
任务成功率。
最终目标应当是:
以最低综合成本,完成一个真正可验收的任务。
这也是普通 API 商城与真正 AI 算力交易中心之间的区别。
完整闭环:生产、流通与结算
词元工厂集中出现,只是产业变化的起点。
成熟的 AI 算力经济需要形成完整循环:
电力与硬件生产算力;
模型将算力转化为 Token;
词元工厂标准化产出;
交易中心聚合并路由供给;
开发者通过 API 消耗 Token;
资金实时结算给供应商;
收益继续投入新的算力生产。
在这个体系中,每个参与者都可能拥有多重角色。
模型厂商在出售 Token 的同时,也在购买算力。
算力供应商在出售 GPU 容量的同时,也可能采购模型 API 来升级自身服务。
企业可能采购外部模型,也可能将自己的垂直模型或闲置推理能力接入平台。
市场不再是一条单向供应链,而会逐渐成为多边交易网络。
结语:工厂已经出现,下一步是让 Token 真正流动
一周内落地四座词元工厂,并不是终点。
它说明 AI 算力已经进入新的工业化阶段。
产业正在完成一系列转变:
从硬件走向服务;
从卡时走向词元;
从峰值算力走向每瓦词元数;
从碎片化集群走向标准化产出;
从固定 API 走向动态市场。
词元工厂解决生产问题。
下一道挑战是流通。
Token 需要聚合、智能路由、动态定价与实时清结算的市场基础设施。
这正是 Unexhub 希望承担的角色。
通过连接模型与算力供应商,提供毫秒级路由,筹备 API 竞价,并推动实时收益结算,Unexhub 正在尝试建设新一代词元工厂之上的市场层。
对开发者而言,部分线路最低约为市场或官方标准价格三折,可以为 Agent、批量任务和模型测试释放更多预算空间。
对供应商而言,统一接入、实时分成和早期低佣金机制,则能够缩短闲置算力转化为可持续现金流的路径。
电力之所以成为通用商品,是因为“度电”通过统一电网和市场,把发电厂与用户连接了起来。
AI 算力正在接近同一个时刻。
词元工厂负责生产标准化智能,算力交易中心负责让智能流动。
真正的“度电时刻”,不仅是 Token 能够规模化生产的那一天。
而是 Token 能像电力一样,被自然地计量、定价、路由、交易与结算的那一天。