
When 950 Million Users Start Consuming Tokens, Pricing Has to Change
Bilingual Article|English first, followed by the complete Chinese edition.
双语文章|英文全文在前,完整中文版请继续向下阅读。
On July 23, Alphabet released a quarterly report that left the market with two very different impressions.
Total revenue reached $119.796 billion, up 24% year over year, while Google Cloud revenue surged 82% to approximately $24.8 billion.
Those numbers were impressive. Yet for engineers and AI developers, two other figures were even more revealing:
The Gemini app reached 950 million monthly active users, while its model APIs processed approximately 22 billion tokens every minute.
One quarter earlier, API throughput stood at around 16 billion tokens per minute. In just three months, it increased by nearly 40%.
Twenty-two billion tokens per minute means more than 366 million tokens are moving through Google’s infrastructure every second.
Every token represents real activity:
A user asking a question;
A developer calling an API;
An enterprise agent processing a document;
A model reading context;
A system retrieving data;
An AI application generating an answer.
Tokens are no longer merely a technical term used to calculate context windows and API bills.
They are becoming one of the AI economy’s most fundamental units of production, consumption, and settlement.
And when nearly one billion users begin consuming the same resource at enormous scale, its pricing model can no longer remain unchanged.
1. The Utility Moment of the Token Economy
The source material cites data from the China Academy of Information and Communications Technology showing that average daily token usage in China increased from approximately 100 billion at the beginning of 2024 to around 140 trillion by March 2026.
That is growth of more than one thousand times in roughly two years.
An IDC forecast cited in the original text also expects global annual token consumption to rise dramatically by 2030. Exact figures depend on statistical definitions and forecasting methods, but the direction is unmistakable:
Token demand is growing nonlinearly.
The reason is simple.
AI is evolving from a tool that answers one question into an infrastructure layer that continuously performs work.
A single user request may now trigger multiple actions:
Understand the user’s intent;
Retrieve relevant information;
Read documents or databases;
Call external tools;
Generate an answer;
Check the result;
Continue working through follow-up tasks.
One request may consume thousands or tens of thousands of tokens.
An enterprise Agent may perform thousands of tasks every day.
A global application may continuously route traffic across multiple models, regions, and suppliers.
As AI becomes embedded in software development, search, customer service, finance, healthcare, education, and business operations, Token consumption begins to resemble the consumption of a utility.
Water measures physical consumption.
Electricity measures energy consumption.
Bandwidth measures network consumption.
Tokens increasingly measure AI inference consumption.
Nvidia CEO Jensen Huang has repeatedly compared tokens to a new form of commodity output produced by AI factories.
The comparison is useful because commodities need more than production.
They also need:
Standardized measurement;
Transparent pricing;
Efficient distribution;
Stable settlement;
Sufficient market liquidity.
Today, token production is expanding rapidly, but the market infrastructure surrounding it remains fragmented.
2. The Limits of Pure Usage-Based Pricing
The dominant API pricing model is currently simple:
Use a certain number of tokens and pay for that amount.
This approach has clear advantages.
It is flexible, transparent at the unit level, and relatively friendly to developers who are experimenting with small workloads.
A startup does not need to sign a large annual contract before testing an idea. It can begin with a few API calls and increase usage as the product grows.
But usage-based pricing also creates new problems.
For enterprises, costs become difficult to predict
Once AI is deeply integrated into business operations, Token usage no longer grows in a smooth, linear pattern.
Usage can rise suddenly when:
More employees gain access;
A new Agent workflow is launched;
Customer-service traffic increases;
A large document set is processed;
A model enters a retry loop;
A long-context task is repeatedly executed.
A monthly bill may remain stable for several weeks, then rise sharply after one workflow is deployed at scale.
Unit prices may be clear, but the final cost can still be unpredictable.
For finance teams, this creates a budgeting problem.
For engineering teams, it creates pressure to restrict usage just as employees begin discovering productive applications.
For individuals, every call feels like a meter running
Pure usage-based pricing also creates psychological friction.
Users may hesitate before uploading a long document or starting a deep conversation because they do not know how many tokens the task will consume.
That anxiety may be small for one request, but it discourages continuous use.
Consumers generally prefer a predictable monthly payment to an invisible meter that increases with every conversation.
This is why the AI industry is unlikely to rely on only one pricing mechanism forever.
Usage-based billing will remain important, especially for developers and infrastructure buyers.
But as Token consumption expands, subscriptions, reserved capacity, dynamic pricing, and market-based trading will increasingly coexist.
3. Subscriptions: Consuming Tokens Like Mobile Data
A second pricing layer is already emerging: Token subscriptions.
According to the examples cited in the original material, Chinese telecom operators have begun experimenting with personal and household AI-compute packages.
Some plans combine a monthly subscription fee with a defined Token allowance, while others advertise a fixed quantity of tokens for a low entry price.
Mainstream AI applications are also shifting toward a hybrid model:
Subscription access for regular users;
Usage limits or fair-use policies;
Additional Token purchases for heavier workloads;
API-based billing for developers and businesses.
This resembles the evolution of internet access.
In the dial-up era, users paid according to the number of minutes they stayed online.
The visible cost discouraged usage.
Broadband subscriptions changed the behavior completely. Once users paid a predictable monthly fee, they stopped counting every minute and began integrating the internet into everyday life.
Token subscriptions can have a similar effect.
Users no longer ask:
“How much will this conversation cost?”
Instead, they ask:
“How much AI capability is included in my plan?”
For individuals and small teams, this lowers the psychological threshold.
For providers, subscriptions create recurring revenue and more predictable demand.
For enterprises, reserved Token packages can make budgeting easier than purely variable monthly billing.
But subscriptions mainly solve the demand-side problem.
They make Token consumption easier to purchase.
They do not fully address the other half of the market:
How should fragmented Token supply be priced, distributed, and settled?
That requires a different layer of infrastructure.
4. Why Token Pricing Cannot Remain One-Dimensional
A million tokens do not always represent the same economic product.
Their production cost and business value can differ according to:
Model intelligence;
Input or output usage;
Cache hits;
Context length;
Response latency;
Region;
Service stability;
Peak demand;
Domain specialization;
Tool-use capability.
A million tokens used for casual conversation are not equivalent to a million tokens used in pharmaceutical research, legal analysis, industrial design, or a revenue-generating enterprise Agent.
A cached input token costs less to process than an uncached token.
A request that can wait for an hour should not necessarily be priced like a real-time request that must respond in two seconds.
A general model with spare capacity should not always carry the same price as a specialized model with limited availability.
Pure usage-based pricing treats Token quantity as the primary variable.
A more mature market must also consider:
Quality;
Urgency;
Reliability;
Scarcity;
Business requirements;
Supply-and-demand conditions.
This is where Token pricing begins to resemble energy, bandwidth, and cloud-capacity markets.
The market may eventually contain several layers at the same time:
Usage-based pricing for flexible consumption;
Subscriptions and packages for predictable access;
Reserved capacity for enterprise stability;
Dynamic bidding for real-time supply and demand;
Specialized pricing for high-value models and workloads.
These approaches do not replace one another.
They serve different users and different scenarios.
5. Unexhub: Allowing Tokens to Be Traded Like Digital Commodities
Subscriptions reduce the purchasing barrier for consumers and small businesses.
But the supply side of the Token economy still needs liquidity.
This is the market layer Unexhub is attempting to build.
Unexhub positions itself as a Token aggregation and distribution compute marketplace, connecting fragmented AI suppliers with enterprises and developers that require model capacity.
Its role is not simply to place another API gateway between the customer and the model provider.
The larger objective is to help Token resources become easier to:
Discover;
Compare;
Route;
Purchase;
Price;
Settle.
On the supply side, Unexhub brings together providers of different sizes, including services associated with mainstream models such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, as well as resources supplied by smaller AI companies and independent providers.
On the demand side, developers and enterprises can access multiple model resources through a unified interface.
This reduces the need to manage a separate account, balance, authentication system, API format, and bill for every provider.
More importantly, Unexhub introduces several mechanisms intended to improve market liquidity.
6. Real-Time Settlement: Turning Delivered Compute Into Available Cash
One of the biggest challenges facing small and medium-sized AI suppliers is not necessarily demand.
It is cash flow.
A supplier may deliver inference capacity today but wait one or two months before receiving payment.
During that period, it still needs to cover:
Electricity;
Hardware depreciation;
Bandwidth;
Cloud-resource costs;
Operations;
Technical staff.
Usage may be increasing while available cash continues to decline.
Unexhub’s real-time settlement mechanism is designed to shorten the gap between delivery and payment.
Its principle can be summarized as:
Once Token capacity has been consumed, the corresponding supplier share should become available for settlement promptly.
For suppliers, this converts delayed accounts receivable into reusable working capital.
The practical value includes:
Faster recovery of operating funds;
Lower financing pressure;
More predictable cash flow;
Greater ability to expand capacity;
Less risk during periods of rapid growth.
The phrase “supply what you can and settle what has been consumed” is not simply a payment slogan.
For smaller suppliers, it can determine whether spare capacity becomes a sustainable business.
7. Zero Commission for Early Suppliers
During the early stage of marketplace development, qualifying suppliers may receive a zero-commission entry benefit.
This is designed to lower the cost of testing a new distribution channel.
A smaller model provider or independent compute supplier can connect resources, observe actual demand, and begin earning usage revenue without immediately losing part of every transaction to platform fees.
The benefit is more meaningful than a temporary discount.
It provides:
A lower-risk route into the market;
More protection for thin margins;
Greater flexibility when testing prices;
Faster access to developer and enterprise demand;
A shorter path from idle capacity to revenue.
This also supports the marketplace’s network effect.
More suppliers create more choice and more competitive pricing for buyers.
More demand makes the marketplace more valuable to suppliers.
Early zero-commission access helps both sides reach meaningful scale more quickly.
8. API Bidding: Letting Supply and Demand Discover the Price
Most model APIs still use fixed pricing.
The supplier announces a price, and the developer decides whether to accept it.
But AI capacity changes constantly.
During off-peak periods, large amounts of inference infrastructure may remain unused.
During peak periods, the same capacity can become scarce.
A fixed price cannot fully reflect both conditions.
Unexhub’s API bidding model is intended to allow pricing to move dynamically with supply and demand.
When capacity is abundant:
Suppliers can offer lower prices;
Developers can schedule flexible tasks more cheaply;
Idle compute can generate additional revenue;
Batch processing can move into lower-cost periods.
When demand rises:
Higher-priority requests can secure capacity;
Suppliers can earn more from scarce resources;
Infrastructure can be allocated toward higher-value workloads.
This mechanism creates advantages for both sides.
Independent developers can access lower prices when their tasks are not time-sensitive.
Suppliers can improve utilization during quiet periods and increase returns during periods of strong demand.
The market can also distinguish between:
Real-time and delay-tolerant workloads;
Peak and off-peak usage;
General-purpose and specialized models;
Standard and guaranteed service levels.
This is how a Token begins to move from a provider-controlled billing unit toward a market-priced digital commodity.
9. Millisecond-Level Routing and Unified Model Access
A Token marketplace cannot function efficiently if every supplier remains isolated behind a separate API.
Unexhub therefore uses a unified access layer and millisecond-level routing to connect demand with suitable model resources.
A request can be routed according to:
Model capability;
Token price;
Current capacity;
Response latency;
Task type;
Regional availability;
Service stability.
A basic classification request does not need to use the most expensive frontier model.
A complex coding or reasoning workflow may require a more capable endpoint.
A flexible batch task can be sent to lower-cost capacity, while an enterprise request with strict latency requirements may prioritize speed and availability.
Developers do not need to rebuild their applications whenever a model’s price, availability, or performance changes.
The routing layer can make the adjustment behind the same interface.
For selected models and usage plans, users may access resources at prices as low as approximately 30% of the corresponding official or prevailing market rate.
The value becomes particularly visible in high-volume workloads.
A small discount on ten requests has limited impact.
Across millions or billions of tokens, it can materially reduce the cost of:
AI search;
Coding Agents;
Enterprise knowledge bases;
Customer-service automation;
Document processing;
Batch content generation;
Multi-Agent workflows.
The objective is not simply to make one API request cheaper.
It is to make continuous and intensive AI usage financially sustainable.
10. From Pay-as-You-Go to a Layered Token Economy
The Token economy is not moving from one pricing model to another.
It is developing into a layered structure.
The foundation: Usage-based billing
Developers pay according to actual consumption.
This remains flexible and suitable for experimentation, APIs, and variable workloads.
The middle layer: Subscriptions and Token packages
Individuals and businesses purchase predictable monthly access.
This lowers psychological friction and makes budgeting easier.
The infrastructure layer: Market-based trading
Suppliers and buyers meet through platforms such as Unexhub, with prices influenced by demand, capacity, latency, and service requirements.
The enterprise layer: Reserved and customized capacity
Large organizations purchase guaranteed throughput, customized routing, integration support, and more predictable settlement terms.
Each layer solves a different problem.
Usage-based pricing provides flexibility.
Subscriptions provide predictability.
Trading markets provide liquidity.
Reserved capacity provides stability.
The future is unlikely to belong to only one model.
Instead, Token pricing will become increasingly segmented according to customer size, workload, urgency, and business value.
11. The Real Meaning of 950 Million Users
The significance of Gemini’s 950 million monthly active users is not simply that Google has built another large consumer application.
It means Token consumption is entering a mass-market phase.
When hundreds of millions of users begin asking AI to:
Search;
Write;
Translate;
Analyze;
Code;
Plan;
Create;
Operate tools;
Token demand stops being a niche infrastructure metric.
It becomes a recurring economic activity.
At the same time, API throughput of 22 billion tokens per minute shows that developer and enterprise demand is growing alongside consumer adoption.
This combination changes the nature of the market.
Token demand is becoming:
Large;
Continuous;
Global;
Multi-model;
Price-sensitive;
Infrastructure-dependent.
A market of this scale cannot rely indefinitely on fragmented accounts and one-sided provider pricing.
It needs better mechanisms for price discovery, distribution, routing, and settlement.
12. My View: Token Pricing Is Entering Its Next Phase
The 950 million Gemini users and 22 billion tokens processed per minute demonstrate the scale and reality of Token demand.
The rapid growth of daily Token consumption shows that AI is moving from occasional use into continuous production.
Tokens are becoming the water, electricity, and bandwidth of the AI economy.
But a utility-scale resource needs more than a meter.
It needs a complete market system.
Usage-based pricing will continue to serve developers.
Subscriptions will make AI easier for individuals and small teams to consume.
Reserved capacity will help large enterprises control risk.
Dynamic marketplaces will allow supply and demand to discover more efficient prices.
Unexhub is attempting to build this market infrastructure by:
Aggregating domestic and international AI suppliers;
Connecting mainstream resources such as Gemini, ChatGPT, Claude, DeepSeek, and GLM;
Providing unified API access;
Routing requests in milliseconds;
Supporting market-based API bidding;
Offering qualifying resources at prices as low as approximately 30% of prevailing rates;
Allowing suppliers to connect once and receive usage revenue through real-time settlement;
Providing enterprise integration and customized service support.
When Tokens are consumed like utilities, purchased like data packages, and traded like commodities, AI commercialization enters a new stage.
The transition is not from “pay per token” to one universal replacement.
It is from a single pricing model to a complete Token economy.
And when nearly one billion people begin consuming that economy every month, the way Tokens are priced, distributed, and settled must evolve with them.
当 9.5 亿人开始消耗 Token,定价方式必须随之改变
中文完整版|上方为英文全文,以下为完整中文内容。
Chinese Edition|The complete English version appears above.
7 月 23 日,Alphabet 交出了一份让市场出现明显分歧的季度财报。
总营收达到 1197.96 亿美元,同比增长 24%;Google Cloud 营收同比大涨 82%,达到约 248 亿美元。
这些数字已经足够亮眼。
但对于工程师和 AI 开发者而言,真正值得停下滑轮细看的,是另外两个数字:
Gemini 应用月活跃用户达到 9.5 亿,模型 API 每分钟处理约 220 亿 Token。
一个季度前,后一个数字还是每分钟约 160 亿。短短三个月,Token 处理量增长了近 40%。
每分钟 220 亿 Token,意味着每秒有超过 3.66 亿 Token 在 Google 的算力基础设施中流动。
每一个 Token 背后,都对应着一次真实活动:
一名用户提出问题;
一名开发者调用 API;
一个企业 Agent 处理文件;
一个模型读取上下文;
一套系统检索数据;
一个 AI 应用生成答案。
Token 已经不再只是用于计算上下文长度和 API 账单的技术术语。
它正在成为 AI 时代最基础的生产、消费与结算单位之一。
当接近十亿用户开始大规模消耗同一种资源时,它的定价方式也不可能继续维持原状。
一、Token 经济正在迎来“水电气时刻”
原文援引中国信通院的数据称,国内日均 Token 调用量从 2024 年初约 1000 亿,增长至 2026 年 3 月约 140 万亿。
短短两年左右,增幅超过一千倍。
原文引用的 IDC 预测也认为,到 2030 年,全球年度 Token 消耗量将出现极其显著的增长。具体数字会受到统计定义和预测方法影响,但趋势已经非常明确:
Token 需求正在呈现非线性扩张。
原因并不复杂。
AI 正在从回答一个问题的工具,变成可以持续完成工作的基础设施。
一个用户请求,可能触发多个动作:
理解用户意图;
检索相关信息;
读取文件或数据库;
调用外部工具;
生成答案;
检查结果;
继续完成后续任务。
一次请求可能消耗数千甚至数万 Token。
一个企业 Agent 每天可能执行数千项任务。
一个全球化应用,则可能持续在多个模型、多个区域和多个供应商之间分发请求。
当 AI 进入软件开发、搜索、客服、金融、医疗、教育和企业运营,Token 消耗便开始呈现出公共资源式的特征。
水用于衡量生活与生产用水;
电用于衡量能源消耗;
带宽用于衡量网络使用量;
Token 则越来越多地用于衡量模型推理消耗。
英伟达 CEO 黄仁勋曾多次将 Token 比作由 AI 工厂持续生产的新型商品。
这个比喻的关键在于:大宗商品不仅需要生产,还需要完整的市场基础设施,包括:
标准化计量;
透明定价;
高效分发;
稳定结算;
足够的市场流动性。
当前,Token 的生产和消耗都在快速增长,但围绕它的市场体系仍然相对分散。
二、单纯按量计费,已经开始暴露局限
目前主流的 API 定价方式很简单:
使用多少 Token,就支付多少钱。
这种模式具有明显优势。
计量方式清晰,使用灵活,也比较适合正在试验产品的中小开发者。
一家初创公司不必先签署一份庞大的年度合同,只需要调用少量 API,就能测试自己的业务设想,再随着产品增长逐步扩大使用量。
但按量计费也产生了新的问题。
对企业来说,成本很难预测
当 AI 深度进入企业流程后,Token 使用量不会按照平滑的直线增长。
以下任何一个变化,都可能让调用量突然上升:
更多员工获得 AI 使用权限;
新的 Agent 工作流正式上线;
智能客服流量增加;
一批长文档被集中处理;
模型进入异常重试循环;
长上下文任务被反复执行。
月度账单可能连续几周保持稳定,却在某条自动化流程规模化后突然上涨。
单价看起来透明,最终成本却依然可能不可预测。
这给财务部门带来预算压力,也可能迫使工程团队限制调用量——而此时员工才刚刚开始发现 AI 的真实价值。
对个人来说,每次调用都像计价器在跳
纯按量收费还会形成心理门槛。
用户在上传长文档或开启深度对话之前,可能会担心:
“这次到底会消耗多少 Token?”
“这轮对话会不会很贵?”
单次顾虑或许不大,但长期来看,会阻碍用户把 AI 融入日常工作。
消费者通常更喜欢固定、可预测的月费,而不是每次对话都在增长的隐形计价器。
因此,AI 产业未来不太可能永远依赖唯一的定价模式。
按量计费仍会长期存在,特别适合开发者和基础设施采购方。
但随着 Token 消耗持续扩大,订阅套餐、预留容量、动态定价和市场化交易将逐渐共存。
三、订阅制:让 Token 像手机流量一样被消费
第二层定价方式已经出现:Token 订阅。
根据原文列举的案例,国内电信运营商已开始尝试个人版和家庭版 AI 算力套餐。
部分方案采用固定月费搭配明确的 Token 额度,也有产品以较低门槛提供一定数量的 Token。
主流 AI 应用也在逐渐走向混合商业模式:
普通用户按月订阅;
套餐内设置使用额度或公平使用规则;
高强度用户额外购买 Token;
开发者和企业继续按 API 用量付费。
这一变化很像互联网接入方式的演进。
拨号上网时代,用户按照在线分钟数付费。
明显的时间成本,会让人不自觉地压缩使用量。
宽带包月出现后,用户不再计算每一分钟,而是把互联网真正融入生活与工作。
Token 套餐也可能产生类似作用。
用户不再反复考虑:
“这次对话会花多少钱?”
而是关心:
“我的套餐里包含多少 AI 能力?”
对个人和小型团队而言,订阅制可以降低心理门槛。
对服务商而言,订阅能够带来持续收入和更加稳定的需求预期。
对企业而言,预购 Token 套餐也比完全浮动的月度账单更容易制定预算。
但订阅制主要解决的是需求侧问题。
它让 Token 更容易被购买和消费,却没有彻底回答供给侧的另一个问题:
分散的 Token 产能应该怎样定价、分发和结算?
这需要另一层基础设施。
四、Token 的价格不能只由数量决定
一百万 Token 并不总是同一种经济商品。
其生产成本和业务价值,会受到多种因素影响:
模型能力;
输入或输出类型;
是否命中缓存;
上下文长度;
响应延迟;
服务区域;
稳定性保障;
高峰供需;
垂直领域能力;
工具调用能力。
用于普通闲聊的一百万 Token,与用于药物研发、法律分析、工业设计或企业 Agent 的一百万 Token,并不等价。
缓存命中的输入 Token,处理成本低于完全重新计算的输入 Token。
可以等待一小时的任务,也不应该始终按照两秒内必须响应的实时任务定价。
拥有大量空闲容量的通用模型,与容量有限的专业模型,也很难长期维持同样的价格逻辑。
纯按量计费,主要将 Token 数量作为变量。
一个更加成熟的市场,还需要同时考虑:
质量;
时效;
可靠性;
稀缺程度;
业务要求;
实时供需。
这时,Token 定价就开始接近能源、带宽和云容量市场。
未来可能长期共存多层定价方式:
按量计费,服务弹性使用;
订阅与套餐,提供可预测的消费体验;
预留容量,保障企业稳定性;
动态竞价,匹配实时供需;
专业化定价,服务高价值模型和场景。
它们并不是互相替代,而是在服务不同规模、不同任务和不同预算的用户。
五、Unexhub:让 Token 像数字商品一样交易
订阅制可以降低个人和中小企业的消费门槛。
但 Token 经济的供给侧,仍然需要足够的流动性。
这正是 Unexhub 试图建立的市场层。
Unexhub 将自身定位为一个 Token 聚合与分发算力交易中心,连接分散的 AI 供应商,以及需要模型能力的企业和开发者。
它并不只是简单地在客户与模型厂商之间增加一层 API 网关。
更大的目标,是让 Token 资源更容易被:
发现;
比较;
路由;
采购;
定价;
结算。
在供给侧,Unexhub 汇聚不同规模的服务商,包括与 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型相关的资源,也包括中小 AI 企业和独立供应商提供的优质能力。
在需求侧,开发者和企业可以通过统一接口访问不同模型,减少为每家供应商分别维护账户、余额、鉴权、API 规范和账单的重复投入。
更重要的是,Unexhub 通过几项机制改善 Token 市场的流动性。
六、实时结算:让已经交付的算力尽快变成现金
对中小 AI 供应商来说,最大的挑战未必是没有需求,而是现金流。
供应商今天已经交付推理算力,但调用收入可能要等待一两个月才能到账。
在这段时间里,仍然需要承担:
电费;
硬件折旧;
网络带宽;
云资源费用;
运维投入;
技术人员成本。
使用量可能持续增长,可用现金却不断减少。
Unexhub 的实时结算机制,旨在缩短算力交付与收入回收之间的距离。
其逻辑可以概括为:
Token 产能一旦被真实消耗,相应的供应商分成就应尽快进入结算。
对于供应商而言,这能够把长期应收款更快转化为可周转资金。
具体价值包括:
更快回收运营成本;
降低融资压力;
提高现金流可预测性;
获得更强的扩容能力;
降低高速增长阶段的经营风险。
“供多少,结算多少”并不只是支付口号。
对中小供应商而言,它可能决定闲置算力能否成为一项可持续业务。
七、早期供应商零佣金权益
在平台市场启动阶段,符合条件的早期供应商还可享受 零佣金入驻权益。
这能降低测试新分发渠道的成本。
中小模型供应商或独立算力提供者,可以先接入资源、观察实际需求并获得调用收入,而不必从第一笔交易开始就向平台让出一部分收益。
它的价值不只是一项短期折扣,还包括:
以更低风险进入市场;
为较薄的利润空间提供保护;
获得更灵活的价格测试条件;
更快触达企业和开发者需求;
缩短闲置产能转化为收入的路径。
这一机制也有利于平台形成网络效应。
供应商越多,需求方获得的选择和价格竞争越充分。
需求越多,平台对供应商的渠道价值也越高。
早期零佣金能够帮助供需两侧更快达到有效规模。
八、API 动态竞价:让供需共同发现价格
大多数模型 API 目前仍采用固定价格。
供应商公布价目表,开发者决定是否接受。
但算力供给一直在变化。
低谷时段,可能存在大量闲置推理容量。
高峰时段,同样的资源又可能变得紧张。
固定价格很难同时准确反映这两种状态。
Unexhub 推动的 API 竞价模式,可以让价格根据供需动态变化。
当容量充足时:
供应商可以降低价格吸引任务;
开发者可以更低成本运行弹性工作;
闲置算力能够转化为收入;
批处理任务可以转移到低价时段。
当需求上升时:
高优先级任务可以锁定容量;
供应商可以从稀缺资源中获得更高收益;
基础设施能够优先分配给高价值工作负载。
这一机制同时改善供需两端。
独立开发者可以在任务不紧急时获得更低价格。
供应商可以提升低谷期利用率,并在需求旺盛时提高收益。
市场还可以进一步区分:
实时任务与可延迟任务;
高峰调用与低谷调用;
通用模型与专业模型;
普通服务与保障型服务。
Token 由此开始从模型厂商单方面定价的计费单位,转向由市场动态发现价格的数字商品。
九、毫秒级路由与多模型统一接入
如果每个供应商都封闭在独立 API 之后,Token 市场就难以实现高效流动。
因此,Unexhub 通过统一接入层和毫秒级路由,将需求与更加合适的模型资源连接起来。
路由时可以综合考虑:
模型能力;
Token 价格;
当前容量;
响应延迟;
任务类型;
区域可用性;
服务稳定性。
简单分类任务不一定需要调用价格最高的旗舰模型。
复杂编程和推理任务,则可能需要能力更强的端点。
具有时间弹性的批处理任务,可以被分配到成本更低的容量;对响应速度要求较高的企业请求,则可以优先选择低延迟与高可用服务。
当模型的价格、容量或性能发生变化时,开发者不必反复重构应用。
路由层可以在同一接口背后完成资源切换。
在部分模型和调用方案中,用户最低可以获得约为相应官方标准价格或市场价格 30% 的综合调用成本。
这一权益在高调用量业务中尤其明显。
十次请求节省的费用可能有限。
但当业务持续消耗数百万、数亿甚至数十亿 Token 时,更具竞争力的价格可以显著降低以下场景的成本:
AI 搜索;
编程 Agent;
企业知识库;
智能客服;
文档处理;
批量内容生成;
多智能体工作流。
平台追求的并不只是让某一次 API 调用更便宜。
而是让持续、高强度的 AI 使用真正具备经济可持续性。
十、从按量付费,走向分层 Token 经济
Token 经济并不是从一种定价模式完全切换到另一种模式。
它正在形成一套分层结构。
基础层:按量计费
开发者根据真实 Token 消耗付款。
这种模式保持了灵活性,适合试验项目、API 服务和波动较大的业务。
中间层:订阅制与 Token 套餐
个人和企业按月购买可预测的使用额度。
它可以降低心理门槛,也更方便制定预算。
市场层:供需交易
供应商和需求方通过 Unexhub 等平台进行匹配,价格受到容量、需求、延迟和服务质量影响。
企业层:预留与定制容量
大型企业采购吞吐保障、定制路由、系统集成服务与更加稳定的结算条件。
每一层都在解决不同问题。
按量计费提供灵活性;
订阅套餐提供可预测性;
交易市场提供流动性;
预留容量提供稳定性。
未来很难只剩下一种定价方案。
Token 价格会根据客户规模、工作负载、紧迫程度和业务价值,变得越来越精细。
十一、9.5 亿用户真正意味着什么?
Gemini 月活达到 9.5 亿,意义并不只是 Google 又打造了一款大型消费级应用。
它说明 Token 消费正在进入大众市场。
当数亿用户开始让 AI:
搜索;
写作;
翻译;
分析;
编程;
规划;
创作;
操作工具;
Token 需求就不再只是基础设施团队内部关注的技术指标。
它开始成为持续发生的经济活动。
与此同时,每分钟 220 亿 Token 的 API 处理量,说明开发者和企业需求也在与消费端同步增长。
两者叠加后,Token 市场开始具备几个明显特征:
规模巨大;
持续发生;
覆盖全球;
涉及多个模型;
对价格敏感;
高度依赖基础设施。
如此庞大的市场,很难长期依靠分散账户和单一供应商定价维持运转。
它需要更成熟的价格发现、资源分发、模型路由与交易结算机制。
十二、我的判断:Token 定价正在进入下一个阶段
Gemini 的 9.5 亿月活用户和每分钟 220 亿 Token 处理量,证明了 Token 需求的真实性与规模。
日均 Token 调用量的快速增长,则说明 AI 正在从偶尔使用的工具,变成持续运行的生产系统。
Token 正在成为 AI 经济中的水、电和带宽。
但达到公共资源规模的生产要素,不能只有一只计价器。
它需要完整的市场体系。
按量计费将继续服务开发者。
订阅套餐会降低个人和小型团队的使用门槛。
预留容量可以帮助大型企业控制风险。
动态交易市场则能够让供需找到更加高效的价格。
Unexhub 正在尝试建设这一市场基础设施:
聚合海内外不同规模的 AI 供应商;
连接 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型资源;
提供统一 API 接入;
以毫秒级速度完成请求路由;
支持市场化 API 竞价;
让符合条件的资源最低达到约市场价格的 30%;
让供应商一次接入,并通过实时结算获得调用分成;
为企业提供系统集成和定制化服务支持。
当 Token 像水电一样被消耗,像流量套餐一样被购买,又像大宗商品一样被交易时,AI 商业化便进入了新的阶段。
这场变化并不是从“按 Token 计费”,转向另一个唯一答案。
而是从单一计价模式,走向一套完整的 Token 经济体系。
当接近十亿人每个月都开始参与这套经济体系时,Token 的定价、分发和结算方式,也必须随之改变。