每分钟 220 亿 Token 调用,Token 交易市场如何撑起谷歌云 248 亿单季收入

Alphabet’s total revenue reached $119.796 billion, up 24% year over year. Google Cloud revenue surged 82% to $24.768 billion, marking the fastest growth in the segment’s history.


22 Billion Tokens per Minute: The Engine Behind Google Cloud’s $24.8 Billion Quarter

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.

On one side, the growth numbers were exceptional.

Alphabet’s total revenue reached $119.796 billion, up 24% year over year. Google Cloud revenue surged 82% to $24.768 billion, marking the fastest growth in the segment’s history.

Cloud operating profit jumped from $2.826 billion to $8.814 billion, while its operating margin expanded from 20.7% to 35.6%.

For investors, these figures demonstrated that Google Cloud had moved beyond the stage of “growing at any cost.” It was growing rapidly while becoming significantly more profitable.

For engineers, however, the most revealing number was elsewhere:

Gemini APIs are now processing approximately 22 billion tokens every minute.

In the previous quarter, that figure was about 16 billion. In just three months, token throughput increased by nearly 40%.

Twenty-two billion tokens per minute means that more than 366 million tokens are flowing through Google’s infrastructure every second.

Behind each token lies a real chain of costs:

  • GPUs and AI accelerators;

  • Data-center electricity;

  • Networking and storage;

  • Model inference;

  • Context processing;

  • Caching and routing;

  • Security and operational support.

Those costs are not abstract.

They are ultimately converted into cloud contracts, API bills, enterprise subscriptions, and the $24.768 billion in quarterly Google Cloud revenue.

This is the deeper story inside Alphabet’s financial results:

Tokens are no longer merely a technical unit used by model developers. They are becoming one of the AI economy’s most important revenue carriers.


1. From a Technical Unit to a Commercial Asset

In the early days of large language models, tokens were mostly discussed inside engineering teams.

Developers used them to estimate context length, calculate API expenses, and prevent prompts from exceeding model limits.

Today, tokens have moved far beyond that technical role.

They have become the unit through which AI demand is measured, priced, delivered, and monetized.

According to the figures cited in the source material, Gemini’s API token-processing volume increased from approximately 16 billion tokens per minute in the previous quarter to 22 billion per minute in the latest period.

At the same time, the Gemini application reached approximately 950 million monthly active users, while nearly 90% of Fortune 100 companies were reported to be using Gemini Enterprise.

These figures reflect three layers of demand:

  1. Consumer demand from individual users;

  2. Developer demand through APIs;

  3. Enterprise demand through cloud and business systems.

Every layer ultimately consumes tokens.

When a consumer asks Gemini to summarize a document, tokens are processed.

When a developer calls the API to generate code, tokens are processed.

When an enterprise uses an AI agent to inspect contracts, answer customer questions, or analyze internal data, tokens are processed.

Tokens therefore provide a common measurement layer across very different products.

They connect consumer applications, enterprise software, models, and infrastructure within the same economic system.


2. The Global Token Explosion Has Only Just Begun

The scale of future token consumption is difficult to comprehend using conventional software metrics.

The source text cites an IDC projection suggesting that annual global token consumption could rise from an extremely small petatoken base in 2025 to more than 152,000 petatokens by 2030, implying extraordinary compound growth.

The exact forecast depends heavily on definitions and statistical methodology, but the direction is clear:

Token consumption is not growing linearly. It is expanding nonlinearly as AI moves into every workflow.

A traditional software application typically responds only when a user performs an explicit action.

An AI-native system may continue operating after the initial request:

  • Reading files;

  • Searching databases;

  • Calling external tools;

  • Generating plans;

  • Running code;

  • Correcting mistakes;

  • Checking results;

  • Preparing follow-up actions.

One user request can trigger dozens of model calls.

One enterprise agent may run thousands of tasks per day.

A global application may process billions of requests across multiple models.

As AI evolves from single-turn chat into continuous agent execution, token consumption rises much faster than user counts.

The number of people using AI matters.

But the number of tasks each person delegates—and the number of model calls inside each task—may matter even more.


3. The $24.8 Billion Revenue Engine

Google Cloud’s $24.768 billion quarterly revenue cannot be attributed to Gemini tokens alone.

It includes infrastructure, databases, storage, security, enterprise software, networking, AI platforms, and other cloud services.

Yet tokens increasingly sit at the center of how those services interact.

An enterprise AI workflow may require:

  • Cloud storage to hold company documents;

  • Databases to retrieve customer information;

  • Security tools to control access;

  • AI models to understand and generate content;

  • Compute infrastructure to run inference;

  • Monitoring systems to track usage;

  • Billing systems to settle costs.

The token is the measurable unit moving through this entire stack.

A customer may believe it is purchasing an enterprise assistant, a coding platform, or an intelligent customer-service system.

Underneath the product, however, the customer is purchasing a stream of model inference.

That stream is measured in tokens.

This is why cloud demand is increasingly shifting from “buying hardware capacity” toward “securing future inference capacity.”

Google Cloud’s contracted backlog reportedly reached $514 billion, approximately five times its current annualized cloud revenue.

That figure does not mean Google has already earned $514 billion.

It represents future contractual commitments that may be recognized gradually over several years.

Nevertheless, the backlog reveals something important:

Enterprises are committing real budgets to future AI and cloud consumption before all of that capacity has even been delivered.

At the commercial level, they are signing cloud contracts.

At the infrastructure level, they are reserving compute, storage, networking, and AI services.

At the most fundamental level, they are locking in their future ability to consume tokens.


4. Why Token Demand Is Becoming More Predictable

AI demand was once difficult to forecast.

Companies experimented with chatbots, small pilot projects, and limited internal tools. Usage could rise temporarily and then disappear.

That phase is ending.

AI is becoming embedded in recurring workflows:

  • Customer support runs every day;

  • Coding agents work through every development cycle;

  • Knowledge bases answer questions continuously;

  • AI search processes every new query;

  • Document systems review every uploaded file;

  • Sales agents analyze every customer account;

  • Automated operations run around the clock.

Once AI enters these workflows, token demand becomes recurring rather than experimental.

This changes its financial character.

A pilot project may consume tokens for one month.

A production workflow may consume them every hour for several years.

The expansion of Google Cloud’s backlog is therefore not only a signal of current demand. It is a sign that enterprises expect token consumption to become a long-term operating expense.

Tokens are moving into the same category as:

  • Cloud storage;

  • Bandwidth;

  • Database capacity;

  • Electricity;

  • Other recurring production inputs.

They are becoming part of the cost of operating a modern company.


5. A Rapidly Growing Market Without Mature Trading Infrastructure

Despite this growth, the token economy remains fragmented.

The industry still lacks universal answers to several basic questions:

  • Who should determine the price of tokens?

  • How should two model providers be compared?

  • How should unused capacity reach developers?

  • How should suppliers be paid?

  • How should cross-model usage be billed?

  • How should enterprises change providers without rebuilding their systems?

At present, most model providers publish their own fixed prices.

Developers must manage separate accounts, balances, authentication systems, service limits, and billing rules.

The cost of one million tokens can vary dramatically depending on:

  • The model;

  • The supplier;

  • Input or output usage;

  • Cache hits;

  • Response latency;

  • Geographic location;

  • Peak demand;

  • Service guarantees.

Even the economic value of those tokens is different.

Tokens used for casual conversation may have limited commercial value.

Tokens used for pharmaceutical research, financial analysis, legal work, industrial design, or revenue-generating agents may produce far greater economic returns.

A token therefore has two different forms of value:

  1. Production value—the compute cost required to generate it;

  2. Application value—the business result it helps create.

Current API pricing primarily reflects production cost, provider strategy, and model capability.

A mature token market would also need to reflect availability, urgency, quality, demand, and application requirements.


6. Existing Token Service Models—and What Is Still Missing

Several service models have already emerged around token consumption.

Direct Model APIs

Developers purchase tokens directly from the model provider.

This provides clear access to the original service, but it creates fragmented integration and billing when multiple models are needed.

Cloud MaaS Platforms

Cloud providers offer model-as-a-service platforms with unified enterprise billing, security, and compliance.

These platforms are well suited to large organizations, but prices, model choices, and routing flexibility may still be constrained by a specific cloud ecosystem.

Model Aggregation Platforms

Platforms represented by OpenRouter bring multiple models under one interface and generate revenue through service charges, routing, or pricing differences.

They have demonstrated substantial market demand for unified access.

However, aggregation alone does not automatically create a fully functioning token market.

A mature market also requires:

  • Transparent price discovery;

  • Dynamic supply-and-demand matching;

  • Supplier-side liquidity;

  • Real-time settlement;

  • Standardized service levels;

  • Efficient capacity allocation.

This is the infrastructure gap Unexhub seeks to address.


7. Unexhub: Building a Marketplace for Token Supply and Distribution

Unexhub positions itself as a token aggregation and distribution compute marketplace.

It connects AI suppliers of different sizes with enterprises and developers seeking model capacity.

The platform includes access to services associated with mainstream domestic and international models, including Gemini, ChatGPT, Claude, DeepSeek, and GLM, while also creating distribution opportunities for smaller companies, independent model developers, and compute suppliers.

Its objective is not merely to place another API gateway between the model and the customer.

It aims to turn fragmented AI capacity into a resource that can be compared, routed, purchased, and settled through a unified market layer.

The platform’s value can be understood through several mechanisms.


8. Real-Time Settlement: From Accounts Receivable to Usable Cash

For small and medium-sized suppliers, long settlement cycles can become a major obstacle.

Inference capacity may be delivered immediately, while payment arrives several weeks or months later.

During that time, suppliers still need to pay for:

  • Electricity;

  • Bandwidth;

  • Equipment depreciation;

  • Cloud infrastructure;

  • Technical operations;

  • Staffing.

A supplier may have growing usage but still face severe cash-flow pressure.

Unexhub’s real-time settlement mechanism is designed to reduce this delay.

Its principle is straightforward:

When tokens have been delivered and consumed, the corresponding revenue should enter settlement as quickly as possible.

For suppliers, the value is not simply receiving money earlier.

It means:

  • Less working-capital pressure;

  • More predictable cash flow;

  • Faster infrastructure reinvestment;

  • Greater ability to respond to demand;

  • Lower dependence on financing.

The original concept of “supply as much as you can and settle what has been consumed” turns delayed receivables into working liquidity.

For smaller suppliers, this can significantly reduce the financial burden of entering the AI market.


9. Zero Commission for Early Suppliers

During the platform’s early development stage, qualifying suppliers may receive a zero-commission participation benefit.

This lowers the risk of entering a new distribution channel.

A smaller AI provider can connect its model or compute capacity, test actual market demand, and begin receiving usage revenue without immediately surrendering part of every transaction to platform fees.

The practical advantages include:

  • Lower integration and market-entry costs;

  • Greater protection for thin margins;

  • Faster access to enterprise and developer demand;

  • More room to test pricing;

  • A quicker route from idle capacity to revenue.

This is a classic two-sided marketplace strategy.

The more supply the platform attracts, the more options buyers receive.

The more demand enters the platform, the more valuable the channel becomes to suppliers.

Zero commission can help accelerate this initial network effect.


10. Dynamic API Bidding: Allowing the Market to Discover Prices

A fixed API price cannot accurately reflect changing supply conditions.

During off-peak hours, large amounts of inference capacity may remain idle.

During peak periods, the same capacity may become scarce.

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

When supply is abundant:

  • Suppliers can lower prices to attract workloads;

  • Developers can run flexible tasks more cheaply;

  • Idle compute can generate revenue;

  • Batch workloads can move to lower-cost periods.

When demand is strong:

  • High-priority workloads can secure capacity;

  • Suppliers can earn higher returns;

  • Scarce resources can be allocated more efficiently.

This makes it possible to distinguish between:

  • Peak and off-peak calls;

  • Real-time and delay-tolerant tasks;

  • General-purpose and specialized models;

  • Standard and guaranteed service;

  • Low-cost and premium capacity.

For developers, this means more opportunities to obtain model access at competitive prices.

For suppliers, it creates a way to monetize both idle capacity and peak demand.

This is how tokens begin moving from a fixed billing unit toward a market-priced resource.


11. Millisecond-Level Routing and Unified 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 distribute requests across available model resources.

Routing decisions can take account of:

  • Model capability;

  • Token price;

  • Current capacity;

  • Response time;

  • Task requirements;

  • Geographic availability;

  • Service stability.

For example, a simple classification task does not need to use the most expensive frontier model.

A complex coding or reasoning task may need to be sent to a more capable endpoint.

A flexible batch workload can be routed toward lower-cost capacity, while a real-time enterprise request may prioritize latency and reliability.

The application does not need to rebuild its integration every time prices or availability change.

The routing layer can select a more suitable resource behind the same interface.

This reduces:

  • Repeated API development;

  • Account fragmentation;

  • Manual model switching;

  • Operational maintenance;

  • Billing complexity.

For selected models and usage plans, developers may access prices as low as approximately 30% of the corresponding official or prevailing market rate.

This benefit is especially meaningful in high-volume use cases.

A small price reduction has limited impact on a few API calls.

Across millions or billions of tokens, it can materially reduce the cost of:

  • AI search;

  • Coding agents;

  • Enterprise knowledge bases;

  • Customer-service automation;

  • Document analysis;

  • Batch content generation;

  • Multistep agent workflows.

The goal is not simply to offer a cheap call.

It is to make continuous, high-intensity AI usage economically sustainable.


12. From Selling GPUs to Selling Continuous Token Production

The transformation taking place across cloud computing can be summarized in one sentence:

The industry is moving from selling hardware capacity to selling continuous token production.

In the previous phase of AI, companies purchased GPUs primarily for model training.

Training was expensive, but it took place in relatively concentrated cycles.

Today, inference has become a continuous production process.

Whenever a user asks a question, generates code, analyzes a contract, searches with AI, or launches an agent, the infrastructure must produce new tokens.

Compute is no longer merely an upfront cost.

It is a productive asset that continuously generates digital output.

This changes what customers care about.

They increasingly ask:

  • How many tokens can be delivered?

  • At what latency?

  • At what cost?

  • With what level of stability?

  • Can workloads switch between models?

  • Can usage be monitored in real time?

  • Can costs be settled clearly?

The commercial product is no longer just “access to a model.”

It is reliable access to an ongoing flow of inference capacity.

Google Cloud’s $514 billion backlog is one expression of this transition.

Customers are reserving future access to that flow.


13. Unexhub as the Liquidity Layer

Unexhub does not need to train every model or own every GPU to play a role in this market.

Its function is to provide liquidity.

For suppliers, this means turning available capacity into revenue.

For developers and enterprises, it means finding appropriate model resources at more competitive prices.

For the overall market, it means establishing mechanisms for:

  • Supply aggregation;

  • Demand distribution;

  • Price discovery;

  • Intelligent routing;

  • Usage measurement;

  • Transaction settlement.

The platform acts as an intermediary infrastructure layer:

It aggregates supply

Large providers, smaller AI companies, independent developers, and distributed compute resources can enter the same marketplace.

It distributes demand

Enterprise customers and developers can access multiple models through one interface.

It discovers prices

Competition and API bidding allow prices to respond to supply and demand.

It routes workloads

Requests can be matched with models according to cost, capability, latency, and availability.

It settles transactions

Suppliers can receive usage-based revenue more promptly.

When these elements work together, Token supply becomes more liquid and easier to consume.


14. My View: $24.8 Billion Is Only the Beginning

Google Cloud’s $24.768 billion in quarterly revenue and Gemini’s 22 billion tokens processed per minute demonstrate that token demand is already real and operating at enormous scale.

The $514 billion backlog shows that enterprises are locking in future demand.

Alphabet’s capital expenditure guidance approaching $205 billion demonstrates how much infrastructure is still required to meet that demand.

These figures tell one consistent story:

Tokens are not a speculative concept. They are becoming a large-scale industrial input.

Every AI search, coding agent, enterprise assistant, and automated workflow requires a continuous supply of model inference.

But a production input of this scale needs more than model providers.

It needs a market.

That market must provide:

  • Diverse supply;

  • Transparent pricing;

  • Stable routing;

  • Efficient settlement;

  • Flexible purchasing;

  • Enterprise-grade service.

Unexhub is attempting to build this infrastructure layer.

By aggregating model and compute suppliers, connecting mainstream services such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, supporting millisecond-level routing, enabling market-based API bidding, offering selected resources at prices as low as approximately 30% of prevailing rates, and settling supplier revenue in real time, the platform is working to transform tokens from a provider-controlled billing unit into a more liquid digital commodity.

Google Cloud’s $24.768 billion quarter is not the conclusion of the AI commercialization story.

It is evidence that the market has begun.

When tokens can move more transparently, efficiently, and freely between suppliers and users, the next stage of AI commercialization will truly begin.


每分钟 220 亿 Token,撑起 248 亿美元云业务

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。

一边,是极其亮眼的增长数据。

Alphabet 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 247.68 亿美元,创下历史最快增速。

云业务运营利润则从 28.26 亿美元跃升至 88.14 亿美元,运营利润率由 20.7% 提升至 35.6%。

对于投资者来说,这组数据证明 Google Cloud 已经不再只是依靠投入换取规模,而是在高速增长的同时,显著提升盈利能力。

但真正让工程师停下滑轮的,是另一个数字:

Gemini 模型 API 每分钟处理约 220 亿 Token。

上一季度,这个数字还是约 160 亿。仅仅一个季度,Token 处理量便增长了近 40%。

每分钟 220 亿 Token,意味着每秒有超过 3.66 亿 Token 在 Google 的算力集群中流动。

每一个 Token 背后,都对应一条真实的成本链:

  • GPU 与 AI 加速器;

  • 数据中心电力;

  • 网络与存储;

  • 模型推理;

  • 上下文处理;

  • 缓存与路由;

  • 安全和运维服务。

这些成本并非抽象数字。

它们最终会转化为云服务合同、API 账单、企业订阅,以及 Google Cloud 单季度 247.68 亿美元的收入。

这才是 Alphabet 财报中更深层的故事:

Token 已经不再只是模型开发者使用的技术单位,而是在成为 AI 经济最重要的收入载体之一。


一、从技术概念,走向商业资产

在大模型发展的早期阶段,Token 主要出现在工程团队的讨论中。

开发者用它计算上下文长度、估算 API 成本,或者判断一段提示词是否会超过模型限制。

今天,Token 已经远远超出了这一技术角色。

它开始成为衡量、交付、计价和变现 AI 需求的统一单位。

根据原文引用的数据,Gemini API 的 Token 处理量从上一季度每分钟约 160 亿,上升至本季度的 220 亿。

与此同时,Gemini 应用的月活跃用户达到约 9.5 亿;接近 90% 的《财富》100 强企业,也已开始使用 Gemini Enterprise。

这些数据代表了三层不同需求:

  1. 来自个人用户的消费级需求;

  2. 来自开发者的 API 调用需求;

  3. 来自企业业务系统的生产需求。

而每一层需求,最终都会消耗 Token。

个人让 Gemini 总结一份文件,需要处理 Token。

开发者调用 API 生成代码,需要处理 Token。

企业用 AI Agent 审查合同、回答客户问题、分析内部数据,同样需要处理 Token。

因此,Token 为不同类型的产品建立了一套共同计量方式。

它将消费级应用、企业软件、模型和算力基础设施,连接进了同一个经济系统。


二、全球 Token 爆发才刚刚开始

未来 Token 消耗量的规模,已经很难再用传统软件指标描述。

原文援引 IDC 的预测称,全球年度 Token 消耗量可能从 2025 年极低的 Peta Token 基数,增长到 2030 年的 15 万 Peta Token 以上。

具体预测结果会受到统计定义与测算方式影响,但大方向已经非常明确:

当 AI 进入所有工作流程后,Token 消耗不会线性增长,而会出现非线性扩张。

传统软件通常只有在用户主动操作时,才会执行一次任务。

AI 原生系统则可能在用户提出需求后继续运行:

  • 阅读文件;

  • 检索数据库;

  • 调用外部工具;

  • 生成执行计划;

  • 运行代码;

  • 修正错误;

  • 检查结果;

  • 准备后续动作。

一次用户请求,可能触发数十次模型调用。

一个企业 Agent,每天可能运行数千项任务。

一个全球化 AI 应用,则可能同时跨多个模型处理数十亿次请求。

当 AI 从单轮问答走向持续运行的智能体,Token 消耗增速会远远超过用户增长速度。

使用 AI 的人数很重要。

但每个人交给 AI 多少任务,以及每项任务内部需要多少次模型调用,可能更加重要。


三、247.68 亿美元收入背后的引擎

Google Cloud 单季度 247.68 亿美元收入,当然不能全部归因于 Gemini Token。

其中还包括基础设施、数据库、存储、安全、企业软件、网络和其他云服务。

但 Token 正在逐渐成为这些服务相互连接的核心。

一条企业 AI 工作流,可能同时需要:

  • 使用云存储保存企业资料;

  • 通过数据库查询客户信息;

  • 借助安全工具管理访问权限;

  • 调用大模型理解和生成内容;

  • 使用算力基础设施完成推理;

  • 依靠监控系统记录调用情况;

  • 通过计费系统结算使用成本。

贯穿这套系统、可以被统一计量的单位,正是 Token。

客户表面上购买的,可能是企业助手、编程平台或智能客服系统。

但在产品底层,客户真正购买的是一股持续输出的模型推理能力。

而这股推理能力,通过 Token 进行计量。

因此,云服务需求正在从“购买硬件容量”,逐渐转向“锁定未来推理产能”。

Google Cloud 已签约但尚未确认收入的积压订单,据称已经达到 5140 亿美元,约为当前云业务年化收入的五倍。

这并不代表 Google 已经赚到 5140 亿美元。

它代表未来数年等待逐步交付和确认的合同需求。

但这个数字仍然传递出一个重要信号:

企业正在用真实预算提前锁定未来的 AI 与云计算消耗能力。

从商业合同看,企业购买的是云服务。

从基础设施看,企业锁定的是算力、存储、网络和模型服务。

从最底层看,它们是在确保自己未来拥有持续消耗 Token 的能力。


四、为什么 Token 需求正在变得更确定?

早期的企业 AI 需求很难预测。

公司可能尝试一个聊天机器人、一个小型试点项目,或者少量内部工具,使用量短暂上升后又迅速回落。

这种阶段正在结束。

AI 开始进入持续运行的业务流程:

  • 智能客服每天都要工作;

  • 编程 Agent 贯穿整个开发周期;

  • 企业知识库持续回答员工问题;

  • AI 搜索处理每一个新查询;

  • 文档系统审查每一份上传文件;

  • 销售智能体分析每一个客户账户;

  • 自动化运营流程全天候运行。

一旦 AI 进入这些业务,Token 需求就不再是实验性的,而会变成周期性、持续性的需求。

这也改变了 Token 的财务属性。

一个试点项目可能只消耗一个月的 Token。

一条生产工作流,却可能连续数年、每小时都在消耗 Token。

Google Cloud 积压订单持续增长,不只是当前需求旺盛的信号。

它还说明,企业已经开始将 Token 消耗视为一项长期经营成本。

Token 正在与以下资源进入同一类目:

  • 云存储;

  • 网络带宽;

  • 数据库容量;

  • 电力;

  • 其他持续消耗的生产资料。

它正在成为现代企业运营成本的一部分。


五、快速增长的 Token 市场,还缺少成熟交易机制

尽管需求快速增长,Token 经济仍然高度分散。

行业尚未对一些基础问题形成统一答案:

  • Token 价格应该由谁决定?

  • 不同模型供应商应该如何比较?

  • 闲置算力如何触达开发者?

  • 中小供应商怎样获得及时结算?

  • 不同模型的调用如何统一计费?

  • 企业如何在不重构系统的情况下切换供应商?

目前,大多数模型厂商仍然独立公布固定价格。

开发者需要分别管理不同的账户、余额、鉴权系统、调用限制和计费规则。

同样是一百万 Token,实际价格却可能受到多种因素影响:

  • 使用的模型;

  • 接入供应商;

  • 输入还是输出;

  • 是否命中缓存;

  • 响应速度;

  • 服务区域;

  • 高峰需求;

  • 稳定性保障。

甚至 Token 本身所能创造的经济价值,也存在巨大差异。

用于普通闲聊的 Token,商业价值可能很低。

用于药物研发、金融分析、法律服务、工业设计或收入型 Agent 的 Token,却可能创造远高于调用成本的收益。

因此,Token 同时具有两种价值:

  1. 生产价值:生成 Token 所消耗的算力成本;

  2. 应用价值:Token 最终帮助用户创造的业务成果。

目前的 API 价格主要反映生产成本、供应商策略和模型能力。

一个成熟的 Token 市场,还需要考虑资源可用性、紧迫程度、服务质量、市场需求和业务要求。


六、现有 Token 服务模式,还缺少什么?

围绕 Token 消耗,目前已经形成几类主要服务模式。

模型官方 API

开发者直接向模型厂商购买 Token。

这种方式可以获得原生模型服务,但一旦需要多个模型,就会产生接口、账户与账单分散的问题。

云厂商 MaaS

云服务商通过模型即服务平台,为企业提供统一计费、安全与合规能力。

这类平台更适合大型组织,但模型选择、价格与路由灵活性,可能仍受特定云生态限制。

模型聚合平台

以 OpenRouter 为代表的平台,通过统一接口整合多个模型,并通过手续费、路由或价格差获得收入。

这类平台已经证明,市场确实存在强烈的多模型统一接入需求。

但仅仅完成模型聚合,还不足以形成真正成熟的 Token 市场。

市场还需要:

  • 透明的价格发现;

  • 动态供需匹配;

  • 供应商资金流动性;

  • 实时交易结算;

  • 标准化服务等级;

  • 高效率容量分配。

这正是 Unexhub 希望补上的基础设施缺口。


七、Unexhub:建设 Token 聚合与分发市场

Unexhub 将自身定位为一个 Token 聚合与分发算力交易中心。

平台连接不同规模的 AI 供应商,以及需要模型能力的企业与开发者。

平台汇聚了与 Gemini、ChatGPT、Claude、DeepSeek、GLM 等海内外主流模型相关的服务,也为中小企业、独立模型开发者和算力供应商提供分发机会。

它的目标并不是简单地在模型与客户之间增加一层 API 网关。

而是将分散的 AI 产能,转化为可以通过统一市场进行比较、路由、采购和结算的资源。

其价值主要体现在几项机制上。


八、实时结算:把应收账款变成可用现金

对于中小型供应商而言,漫长的结算周期可能成为业务扩张的重要阻碍。

算力已经即时交付,但收入可能需要等待数周甚至数月才会到账。

在此期间,供应商仍然需要支付:

  • 电费;

  • 网络带宽;

  • 设备折旧;

  • 云基础设施;

  • 技术运维;

  • 人员成本。

一家供应商可能拥有不断增长的调用量,却依然承受严重的现金流压力。

Unexhub 的实时结算机制,希望缩短算力交付与资金到账之间的时间。

其逻辑非常直接:

Token 已经被交付和使用,相应的收入就应尽快进入结算。

对供应商而言,其价值不只是“更早收到钱”,还包括:

  • 降低营运资金压力;

  • 提高现金流可预测性;

  • 更快投入基础设施扩容;

  • 增强响应市场需求的能力;

  • 减少对外部融资的依赖。

“供多少,结算多少”,能够让漫长账期中的应收款,更快转化为可以周转的资金。

对于中小供应商来说,这有助于显著减轻进入 AI 市场时的资金压力。


九、早期供应商零佣金权益

在平台启动阶段,符合条件的早期供应商可以享受 零佣金入驻权益。

这降低了尝试新分发渠道的风险。

中小模型供应商可以先接入自己的模型或算力,测试市场真实需求,并获得调用收入,而不必从第一笔交易开始就向平台让出收益。

这项权益带来的实际价值包括:

  • 降低接入与市场开拓成本;

  • 为利润较低的供应商保留更多收益;

  • 更快触达企业和开发者需求;

  • 获得更大的价格测试空间;

  • 缩短闲置算力转化为收入的路径。

这是一种典型的双边市场启动策略。

平台吸引的供给越多,开发者获得的模型选择就越丰富。

需求越多,平台对供应商的渠道价值也就越高。

早期零佣金,可以加速这一网络效应的形成。


十、API 动态竞价:让市场发现真实价格

固定 API 价格,很难准确反映算力供给的实时变化。

在低谷时段,大量推理容量可能处于闲置状态。

到了高峰期,同样的容量却可能变得紧张。

Unexhub 推动的 API 动态竞价模式,可以让价格随供需变化。

当供给充足时:

  • 供应商可以降低价格吸引任务;

  • 开发者可以低成本运行弹性工作负载;

  • 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。

当需求旺盛时:

  • 高优先级任务能够锁定容量;

  • 供应商可以获得更高收益;

  • 稀缺资源得到更加高效的配置。

这种模式可以进一步区分:

  • 高峰与低谷调用;

  • 实时与可延迟任务;

  • 通用与专业模型;

  • 普通与保障型服务;

  • 低价与高性能容量。

对开发者而言,这意味着有更多机会获得具有竞争力的调用价格。

对供应商而言,则可以同时提高闲置期与高峰期的资源收益。

Token 也由此开始从固定的模型计费单位,走向由市场动态定价的数字资源。


十一、毫秒级路由与统一接入

如果每个供应商都被封闭在独立 API 后面,Token 市场就无法高效流动。

因此,Unexhub 通过统一接入层与毫秒级路由,将请求分配给不同模型资源。

路由过程可以综合考虑:

  • 模型能力;

  • Token 价格;

  • 当前容量;

  • 响应速度;

  • 任务要求;

  • 区域可用性;

  • 服务稳定性。

例如,一个简单的分类任务,并不需要调用价格最高的旗舰模型。

复杂编程或推理任务,则可能需要更强的模型端点。

具有时间弹性的批处理任务,可以被路由到价格更低的资源;实时企业请求,则可以优先考虑低延迟与稳定性。

当供应商价格或可用性变化时,应用不需要重新搭建接口。

路由层可以在统一接口背后选择更加合适的资源。

这能减少:

  • 重复 API 开发;

  • 账户与余额分散;

  • 人工切换模型;

  • 日常运维压力;

  • 账单管理复杂度。

在部分模型和调用方案中,开发者最低可以获得约为相应官方标准价格或市场价格 30% 的综合成本。

这一权益在高调用量场景中尤其明显。

少量 API 请求节省的费用可能有限。

但当业务需要消耗数百万、数亿甚至数十亿 Token 时,更低的单价可以显著降低以下应用的运营成本:

  • AI 搜索;

  • 编程智能体;

  • 企业知识库;

  • 智能客服;

  • 文档分析;

  • 批量内容生成;

  • 多步骤 Agent 工作流。

平台追求的并不只是提供一次便宜的调用。

而是让长期、高强度的 AI 使用,在成本上真正具备可持续性。


十二、从“卖 GPU”走向“持续生产 Token”

云计算行业正在发生的变化,可以用一句话概括:

行业正在从销售硬件容量,转向销售持续的 Token 生产能力。

在上一阶段,企业购买 GPU,主要是为了进行模型训练。

训练成本很高,但通常集中发生在某些周期。

今天,模型推理已经变成持续生产过程。

每当用户提出问题、生成代码、审查合同、使用 AI 搜索或启动 Agent,底层基础设施都需要生产新的 Token。

因此,算力不再只是模型训练前期的一项成本。

它变成了能够持续生产数字输出的生产性资产。

客户关心的问题也随之改变:

  • 可以交付多少 Token?

  • 响应延迟是多少?

  • 每百万 Token 成本是多少?

  • 服务能否长期稳定?

  • 能否在多个模型之间切换?

  • 使用量能否被实时监控?

  • 费用是否能够清晰结算?

商业产品已经不再只是“使用一个模型”。

而是获得一股稳定、持续的模型推理产能。

Google Cloud 的 5140 亿美元积压订单,就是这种变化的一种表现。

客户正在提前锁定未来使用这股推理产能的权利。


十三、Unexhub:成为 Token 市场的流动性层

Unexhub 不需要训练所有模型,也不必拥有每一张 GPU,仍然可以在 Token 市场中承担重要角色。

它提供的是流动性。

对供应商而言,是把可用的算力转化为收入。

对企业和开发者而言,是以更具竞争力的价格找到适合任务的模型资源。

对整体市场而言,则是建立以下机制:

  • 聚合供给;

  • 分发需求;

  • 发现价格;

  • 智能路由;

  • 计量使用量;

  • 完成交易结算。

平台作为中间基础设施层运行:

聚合供应

让大型供应商、中小 AI 企业、独立开发者与分散算力资源进入同一个市场。

分发需求

让企业客户与开发者通过统一接口调用多个模型。

发现价格

通过供应商竞争和 API 动态竞价,让价格响应市场供需。

路由任务

根据成本、能力、延迟和可用性,为请求匹配合适的资源。

结算交易

按照真实 Token 消耗,更及时地向供应商结算调用收入。

当这些环节协同运行时,Token 供给会获得更强流动性,也会变得更容易被企业与开发者使用。


十四、我的判断:247.68 亿美元只是开始

Google Cloud 单季度 247.68 亿美元的收入,以及 Gemini API 每分钟处理 220 亿 Token 的规模,证明 Token 需求已经真实存在,并进入了极高的流量水平。

5140 亿美元的积压订单说明,企业已经开始提前锁定未来需求。

接近 2050 亿美元的全年资本开支指引则说明,为了满足这些需求,市场仍然需要建设大量基础设施。

所有数字共同指向一个结论:

Token 不是一个用于资本炒作的抽象概念,而是在成为被大规模消耗的产业生产资料。

每一次 AI 搜索、每一个编程 Agent、每一套企业助手和每一条自动化流程,都需要持续获得模型推理能力。

但如此大规模的生产资料,不能只依赖分散的模型供应商。

它需要一个市场。

这个市场需要提供:

  • 多元供给;

  • 透明价格;

  • 稳定路由;

  • 高效结算;

  • 灵活采购;

  • 企业级服务能力。

Unexhub 正在尝试建设这层基础设施。

通过聚合模型与算力供应商,连接 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型资源,支持毫秒级请求路由、市场化 API 竞价,为部分资源提供最低约市场价格 30% 的调用权益,并实时结算供应商调用分成,平台正在推动 Token 从模型厂商控制的计费单位,转化为流动性更强的数字商品。

Google Cloud 单季度 247.68 亿美元的收入,并不是 AI 商业化故事的终点。

它只是证明,这个市场已经真正启动。

当 Token 能够在供应商和用户之间更加透明、高效、自由地流动时,AI 商业化的下一个阶段,才会真正开始。