
Behind Google’s $514 Billion Backlog: Tokens Are Repricing the Entire AI Economy
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 delivered a quarterly report that divided the market.
Total revenue reached $119.796 billion, up 24% year over year. Google Cloud revenue surged 82% to $24.768 billion, marking its fastest growth on record.
Cloud operating profit climbed from approximately $2.8 billion to $8.814 billion, while the segment’s operating margin expanded from 20.7% to 35.6%.
By almost every conventional measure, the cloud business was performing exceptionally well.
Yet Alphabet’s shares fell in after-hours trading.
The reason was equally striking: free cash flow turned negative for the first time in decades, falling to negative $5.855 billion. At the same time, Alphabet raised its full-year capital expenditure guidance from $180–190 billion to approximately $195–205 billion.
One side of the financial statement showed rapidly expanding AI and cloud revenue.
The other showed enormous pressure from data centers, chips, energy, networking, and inference infrastructure.
The market was asking one question:
Where is all this money going, and when will it generate an acceptable return?
Part of the answer lies in another number:
$514 billion.
This figure does not represent revenue that Google has already earned. It represents contracted demand waiting to be delivered.
And behind much of that demand is something more fundamental than servers or cloud accounts:
Future token consumption.
1. A $514 Billion Backlog Is the Monetary Expression of Token Demand
How large is $514 billion?
It is roughly five times the current annualized revenue of Google Cloud.
It is also larger than the projected annual economic output of many countries.
But the more important story is not the size alone. It is the speed at which the backlog has expanded.
According to the figures cited in the original material:
The backlog stood at approximately $106 billion in the second quarter of the previous year;
It reached around $460 billion in the first quarter;
It then increased to $514 billion in the latest quarter.
That means contracted demand rose by another $54 billion in just three months.
This cannot be explained solely by better sales execution.
Enterprise customers are signing long-term agreements because they are competing to secure future AI capacity.
Google management reportedly summarized the situation clearly:
Market demand continues to exceed our investment capacity.
In other words, the main problem is no longer whether customers want AI infrastructure.
The problem is whether enough infrastructure can be built quickly enough to serve them.
At the contract level, customers may be buying cloud services, reserved capacity, model access, databases, storage, networking, and enterprise AI products.
At the economic level, however, much of this demand eventually becomes token consumption.
Customers are not merely asking:
“How many GPUs can we rent?”
They are effectively asking:
“How much AI inference can we reliably consume over the next several years?”
Every enterprise assistant, coding agent, document-analysis workflow, customer-service system, and AI search request must eventually pass through model inference.
Inference is measured in tokens.
The $514 billion backlog can therefore be understood as a monetary expression of future token demand.
2. Tokens Are Becoming a New Unit of AI Valuation
The scale of token usage is increasing at a rate that traditional infrastructure metrics struggle to capture.
According to the figures cited in the article, average daily token consumption in China rose from approximately 100 billion tokens at the beginning of 2024 to around 140 trillion tokens per day by March 2026.
In only two years, usage expanded by more than one thousand times.
Gemini’s API token-processing volume also reportedly increased from approximately 16 billion tokens per minute in the first quarter to 22 billion tokens per minute in the second quarter.
These numbers point to a broader structural change.
AI adoption is no longer measured only by:
The number of companies purchasing GPUs;
The number of models released;
The number of API accounts opened;
The number of people using an AI application.
It is increasingly measured by how much inference is actually consumed.
A company can announce a powerful model, but if few users call it, the economic value remains limited.
A smaller model may look less impressive on a benchmark, but if it processes trillions of tokens across production workflows, it may generate significantly more commercial value.
Tokens connect several parts of the AI economy:
Compute infrastructure produces inference capacity;
Models convert that capacity into usable intelligence;
Applications consume the resulting tokens;
Businesses pay according to usage and outcomes.
That makes token volume a bridge between technical capacity and commercial demand.
In this sense, tokens are becoming a new valuation unit for the AI industry.
They help answer questions such as:
How much productive demand does a model actually have?
How efficiently does infrastructure convert compute into usable output?
How sticky is an AI application?
How much future capacity have customers already committed to buying?
How much revenue can be generated from each unit of inference?
The rapid expansion of token consumption is driving a repricing of the entire AI value chain—from chips and data centers to models, applications, routing platforms, and settlement systems.
3. The Missing Infrastructure of the Token Economy
Token demand is growing quickly, but the underlying market infrastructure remains fragmented.
Today, several basic questions still lack standardized answers:
Who determines the price of a token?
How should different models and service levels be compared?
How should idle compute capacity reach buyers?
How should small suppliers receive payment?
How should cross-platform usage be settled?
How can enterprises switch providers without rebuilding their systems?
Most model vendors still set prices independently.
Developers must accept different billing standards, authentication methods, account balances, rate limits, and settlement cycles.
The same one million tokens may carry dramatically different prices depending on:
The model;
Whether the input is cached;
The region;
The provider;
The response speed;
The time of day;
The required availability level;
The type of workload.
This creates a market with growing demand but limited liquidity.
Supply exists, but it is scattered.
Demand exists, but it cannot always find the most suitable capacity.
Prices exist, but they are difficult to compare.
Usage is growing, but supplier settlement is often slow.
This is the market gap Unexhub aims to address.
4. Unexhub: Building an Exchange Layer for Tokens
Unexhub is positioned as a token aggregation and distribution compute marketplace.
Its goal is not simply to add another API gateway between developers and model providers.
Its broader ambition is to create a market layer where fragmented AI capacity can be aggregated, priced, routed, consumed, and settled more efficiently.
On the supply side, the platform connects:
Large model providers;
Regional AI companies;
Independent model developers;
Small and medium-sized compute suppliers;
Organizations with underused inference capacity.
On the demand side, it serves:
Enterprise customers;
Software developers;
AI application teams;
Agent builders;
High-volume model users.
Unexhub’s model focuses on three central mechanisms: real-time settlement, early-stage zero commission, and market-based API bidding.
Together, these mechanisms attempt to move tokens from a fixed billing unit toward a more liquid, tradable digital resource.
5. Real-Time Settlement: Turning Locked Revenue Into Working Capital
Small and medium-sized AI suppliers often face a serious cash-flow problem.
They provide compute capacity immediately, but payment may not arrive for weeks or even months.
During that period, operating expenses continue:
Electricity must be paid;
GPUs and servers continue depreciating;
Bandwidth bills arrive;
Operations teams must be funded;
Cloud-resource fees remain due.
A supplier may appear profitable on paper while still lacking the cash needed to expand capacity.
Unexhub’s real-time settlement mechanism is designed to shorten the distance between token delivery and revenue collection.
The basic logic is:
When capacity is consumed, the corresponding supplier revenue should become available for settlement as quickly as possible.
This turns accounts receivable into working capital.
For smaller suppliers, the value is direct:
Faster cash recovery;
Lower financing pressure;
More predictable revenue;
Greater ability to expand infrastructure;
Less dependence on long billing cycles.
“Supply what you can, settle what has been consumed” is more than a payment feature.
It can determine whether a smaller supplier is able to survive periods of rapid growth.
6. Early-Stage Zero Commission: Lowering the Cost of Entering the Market
During its market-building stage, Unexhub offers qualifying early suppliers a zero-commission participation benefit.
This is a typical two-sided marketplace strategy.
The platform must first attract sufficient supply. Only then can it create enough choice, competition, and availability to draw sustained demand.
For suppliers, zero commission reduces the cost of testing the channel.
A smaller model company can connect its service, observe actual developer demand, and begin earning usage revenue without immediately giving up part of every transaction.
The benefit is not merely “paying fewer fees.”
It provides suppliers with:
A lower-risk entry point;
More room to test pricing;
Better protection for thin margins;
Faster exposure to new customers;
A chance to convert idle capacity into revenue.
As the supply side becomes more diverse, buyers gain more model options and more competitive prices.
The incentive therefore supports both sides of the market.
7. API Bidding: Allowing Supply and Demand to Discover the Price
The most market-oriented part of Unexhub’s design is its API bidding mechanism.
In the traditional model, a provider publishes one price and users either accept it or leave.
That approach does not fully reflect changing capacity conditions.
At off-peak times, large amounts of inference capacity may sit idle.
During peak periods, the same provider may face heavy congestion and limited availability.
A dynamic bidding system allows prices to respond to these conditions.
When supply exceeds demand:
Suppliers can lower prices;
Developers can run flexible workloads more cheaply;
Idle servers can produce revenue;
Batch and non-urgent tasks can move into lower-cost periods.
When demand exceeds supply:
Higher-value requests can secure priority capacity;
Suppliers can receive better returns;
Scarce infrastructure can be allocated more efficiently.
This structure can distinguish among:
Real-time and delay-tolerant tasks;
Peak and off-peak usage;
General-purpose and specialized models;
Standard and guaranteed service levels;
High-cost and low-cost inference regions.
For independent developers and smaller businesses, dynamic pricing creates an opportunity to access model capacity at significantly lower prices when timing is flexible.
For suppliers, it creates a mechanism for earning more when their capacity becomes scarce.
The result is a more efficient token market.
8. Unified Access and Millisecond-Level Routing
The token market cannot become liquid if every model remains isolated behind a separate interface.
Unexhub therefore aggregates services associated with major domestic and international models, including Gemini, ChatGPT, Claude, DeepSeek, and GLM, while also providing distribution opportunities for models offered by smaller companies and independent suppliers.
Developers can access these resources through a unified interface.
The platform can then route requests according to:
Model capability;
Price;
Availability;
Latency;
Task type;
Regional requirements;
Service stability.
Millisecond-level routing is intended to make this selection process largely invisible to the application.
When one provider becomes expensive, congested, or unavailable, the routing layer can direct suitable workloads toward another resource without requiring developers to rebuild the entire integration.
For selected models and usage plans, developers may receive prices as low as approximately 30% of the corresponding official or prevailing market rate.
This pricing advantage becomes particularly meaningful at scale.
A discount on a few requests is limited.
A lower unit price across millions or billions of tokens can materially change the economics of:
AI search;
Coding agents;
Enterprise knowledge bases;
Customer-service systems;
Document processing;
Batch content generation;
Multistep agent workflows.
The objective is not merely to provide a cheap API call.
It is to make sustained, high-volume AI usage more financially predictable.
9. From Selling GPUs to Selling Tokens
Google’s $514 billion backlog signals a deeper transformation in the AI business model.
The industry is moving from selling infrastructure as a static product toward selling inference as a continuous service.
In the earlier stage of AI development, GPUs were purchased primarily to train models.
Training was a large but relatively discrete event.
Once a model was trained, the industry moved into an era of continuous inference.
Every time a user:
Asks a question;
Generates code;
Processes a document;
Runs an agent;
Requests an image;
Uses AI search;
the infrastructure must produce new output.
Compute is therefore no longer only a cost required to create a model.
It is a productive asset that continuously generates tokens.
This changes the commercial unit.
Customers increasingly care less about the physical number of chips and more about:
How much inference they can consume;
How quickly it can be delivered;
How reliably it remains available;
How much each million tokens costs;
Whether workloads can be routed dynamically;
Whether expenses can be monitored and settled clearly.
The product is gradually shifting from “access to a model” toward “access to continuous token production.”
Some cloud platforms have already reflected this change by separating pricing for:
Input tokens;
Output tokens;
Cached input;
Priority service;
Batch processing;
Reserved capacity.
That degree of pricing precision is a prerequisite for token commercialization.
Once inference can be measured precisely, it can also be compared, routed, traded, and financed.
10. Unexhub as the Liquidity Layer of Continuous Token Production
Unexhub does not need to train every model or own every GPU to participate in this transformation.
Its role is different.
It aims to provide liquidity for continuous token production.
For suppliers, that means turning unused or underdistributed capacity into revenue.
For buyers, it means accessing suitable model resources at a more competitive price.
For the broader market, it means creating mechanisms for price discovery, routing, and settlement.
The platform’s infrastructure can be understood through several layers:
Supply Aggregation
Bringing together large providers, smaller AI companies, independent models, and distributed compute resources.
Demand Distribution
Connecting that supply with enterprises, developers, and AI applications that need inference capacity.
Price Discovery
Allowing competition and dynamic bidding to influence API pricing.
Intelligent Routing
Matching workloads with appropriate models according to cost, speed, capability, and availability.
Transaction Settlement
Calculating actual token consumption and settling supplier revenue promptly.
Enterprise Services
Supporting model selection, API adaptation, concurrency, billing, workflow integration, and ongoing technical operations.
When these layers work together, token supply becomes easier to discover and consume.
That is what transforms scattered compute capacity into a functioning market.
11. Why the $514 Billion Backlog Matters More Than Current Revenue
It is important to understand what the $514 billion figure does—and does not—mean.
It does not mean Google has already received and recognized $514 billion in revenue.
Backlog represents contracted future obligations that may be delivered over several years.
Some agreements may change, expand, or be recognized gradually.
But the number remains highly significant because customers have committed real budgets to future cloud and AI usage.
They are not merely expressing interest.
They are signing contracts.
That makes the backlog a stronger demand signal than:
Survey responses;
AI adoption forecasts;
Model benchmark excitement;
Product waitlists;
Conference announcements.
It shows that major enterprises expect AI consumption to continue growing and are willing to reserve capacity in advance.
The combination of several financial indicators tells a consistent story:
Google Cloud revenue is accelerating;
Operating margins are expanding;
Contracted demand is rising;
Capital expenditure is increasing sharply;
Free cash flow is under pressure;
Management says demand exceeds investment.
Together, these signals suggest that token demand is real, but supply remains capital-intensive.
The market is not debating whether AI will consume more inference.
It is debating who will finance, produce, distribute, and monetize that inference most efficiently.
12. My View: Tokens Are Repricing Everything
Google Cloud’s 82% revenue growth and its $514 billion backlog demonstrate that enterprise AI demand is both substantial and durable.
Negative free cash flow and capital expenditure approaching $205 billion demonstrate how difficult and expensive it is to build enough supply.
This is the defining tension of the AI economy:
Token demand is expanding faster than the infrastructure needed to produce it.
Tokens are not a speculative abstraction.
They are a production input being consumed at enormous scale.
Every coding agent, enterprise assistant, AI search engine, and automated workflow requires a continuous supply of inference.
But a productive resource needs a functioning market.
It needs:
Transparent pricing;
Diverse suppliers;
Reliable routing;
Efficient settlement;
Flexible purchasing;
Scalable distribution.
This is the infrastructure layer Unexhub is attempting to build.
By aggregating supply, routing requests in milliseconds, supporting market-based API pricing, offering qualifying resources at prices as low as approximately 30% of prevailing rates, and settling supplier revenue in real time, the platform aims to turn tokens from a provider-controlled billing unit into a more liquid digital commodity.
The real meaning of Google’s $514 billion backlog is not that Google has already earned that amount.
It is that the market has committed real money to future AI consumption.
Future token demand is already being locked in.
What remains missing is a more transparent and efficient market through which those tokens can flow.
Unexhub is building toward that market.
5140 亿美元积压订单背后:Token 正在重估一切
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 营收同比增长 82%,达到 247.68 亿美元,创下历史最快增速。
云业务运营利润从约 28 亿美元大幅增长至 88.14 亿美元,运营利润率也从 20.7% 跃升至 35.6%。
按照传统财务指标衡量,这几乎是一份无可挑剔的云业务成绩单。
但 Alphabet 股价却在盘后下跌。
原因同样醒目:自由现金流数十年来首次转负,降至 负 58.55 亿美元;与此同时,公司将全年资本开支指引从 1800 亿至 1900 亿美元,上调至约 1950 亿至 2050 亿美元。
财报的一边,是快速增长的 AI 与云业务收入。
另一边,则是数据中心、芯片、电力、网络和推理基础设施带来的巨大资金压力。
市场都在追问同一个问题:
这些钱究竟花到哪里去了,又要到什么时候才能形成足够的回报?
部分答案,藏在另一个数字里:
5140 亿美元。
这并不是 Google 已经确认的收入,而是等待未来交付的已签约需求。
在这笔巨额需求背后,比服务器和云账户更基础的东西,是未来将被持续消耗的 Token。
一、5140 亿美元积压订单,是 Token 需求的货币化表达
5140 亿美元是什么概念?
它约等于 Google Cloud 当前年化收入的五倍,也超过了许多国家一整年的经济总量。
但比绝对规模更加重要的,是这个数字的增长速度。
根据原文引用的数据:
去年第二季度,积压订单约为 1060 亿美元;
今年第一季度上升至约 4600 亿美元;
本季度进一步增长至 5140 亿美元。
短短一个季度,新增签约需求便达到 540 亿美元。
这很难单纯用销售能力增强来解释。
企业客户正在争相签订长期合同,提前锁定未来的 AI 与云计算容量。
Google 管理层在财报电话会上给出的判断非常直接:
市场需求仍然超过公司的投资能力。
这意味着,当前最核心的问题已经不是客户是否愿意购买 AI 基础设施。
而是供应商能否足够快地建设基础设施,满足已经出现的需求。
从合同形式看,客户购买的可能是云服务、预留容量、模型调用、数据库、存储、网络和企业 AI 产品。
但从经济本质看,这些需求中的很大一部分,最终都会转化为 Token 消耗。
客户真正锁定的,不只是“未来能租到多少张 GPU”,而是:
未来几年,我们能够稳定获得多少模型推理能力?
每一个企业助手、编程智能体、文档分析流程、智能客服系统和 AI 搜索请求,最终都要经过模型推理。
而推理能力,正是通过 Token 被计量的。
因此,5140 亿美元的积压订单,可以被理解为未来 Token 需求的一种货币化表达。
二、Token 正在成为 AI 产业的新估值单位
Token 使用量正在以传统基础设施指标难以描述的速度增长。
根据文章援引的数据,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。
两年时间,使用规模扩大超过一千倍。
Gemini API 每分钟处理的 Token 数量,也据称从第一季度约 160 亿增长至第二季度的 220 亿。
这组数据意味着,AI 产业的增长已经不能只用以下指标衡量:
企业购买了多少张 GPU;
市场发布了多少个模型;
有多少开发者开通了 API;
有多少用户打开了 AI 应用。
更加关键的问题是:
到底有多少模型推理正在被真实消耗?
一家企业可以发布参数规模庞大、基准测试优秀的模型,但如果没有足够调用量,其经济价值仍然有限。
另一个规模较小的模型,可能在排行榜上并不耀眼,却因为进入了大量生产工作流、持续处理数万亿 Token,而创造出更高的商业价值。
Token 将 AI 产业链的多个环节连接起来:
算力基础设施提供推理产能;
模型将算力转化为可使用的智能;
应用通过 Token 消耗这些能力;
企业再根据使用量和业务结果支付费用。
Token 因而成为连接技术能力与商业需求的桥梁。
从这个角度看,Token 正在成为 AI 行业新的估值单位,用于衡量:
一个模型究竟拥有多少真实需求;
基础设施将算力转化为输出的效率;
一个 AI 应用的使用黏性;
客户已经锁定了多少未来容量;
每一单位推理可以产生多少收入。
Token 消耗量的高速增长,正在推动芯片、数据中心、模型、应用、路由平台和结算系统的全面重估。
三、Token 经济还缺少一套成熟的市场基础设施
Token 需求正在迅速增长,但底层市场仍然高度分散。
目前,一系列基础问题尚未形成统一答案:
Token 的价格由谁决定?
不同模型与服务等级如何比较?
闲置算力怎样找到买方?
中小供应商怎样及时收到款项?
跨平台调用如何统一结算?
企业怎样在不重构系统的情况下切换供应商?
现阶段,大部分模型厂商仍然独立制定价格。
开发者不得不适应不同的计费标准、鉴权方式、账户余额、速率限制与结算周期。
同样是一百万 Token,价格却可能因以下因素出现显著差异:
模型类型;
是否命中缓存;
服务区域;
接入渠道;
响应速度;
使用时段;
可用性要求;
业务场景。
这形成了一个需求快速增长、流动性却不足的市场。
算力供给已经存在,却分散在不同供应商手中;
Token 需求已经存在,却未必能找到最合适的资源;
市场价格已经存在,却难以透明比较;
调用量持续增长,供应商结算却可能依然缓慢。
Unexhub 希望填补的,正是这层市场缺口。
四、Unexhub:为 Token 建设一层“交易所”基础设施
Unexhub 的定位,是建设一个 Token 聚合与分发算力交易中心。
它的目标并不是简单地在开发者与模型厂商之间增加一道 API 网关。
更大的愿景,是建立一个能够聚合、定价、路由、消费并结算分散 AI 产能的市场层。
在供给侧,平台连接:
大型模型供应商;
地区性 AI 企业;
独立模型开发者;
中小算力供应商;
拥有闲置推理容量的机构。
在需求侧,平台服务:
企业客户;
软件开发者;
AI 应用团队;
Agent 开发者;
高调用量模型用户。
Unexhub 重点围绕三个机制展开:实时结算、早期零佣金和 API 市场化竞价。
这些机制共同指向一个目标:
推动 Token 从固定的计费单位,逐渐变成流动性更强、可以交易的数字资源。
五、实时结算:把被账期锁住的收入变成周转资金
中小 AI 供应商经常面对一个现实问题:现金流。
算力已经即时交付,但收入可能需要等待数周甚至数月才能到账。
在此期间,运营成本仍在持续产生:
电费需要支付;
GPU 与服务器持续折旧;
网络带宽需要续费;
运维团队需要投入;
云资源账单不会暂停。
一家供应商可能在账面上拥有可观收入,却没有足够现金扩大容量。
Unexhub 的实时结算机制,正是为了缩短 Token 交付与资金到账之间的距离。
其核心逻辑可以概括为:
算力一旦被真实消耗,相应的供应商收入就应尽快进入结算流程。
这能把应收账款更快转化为营运资金。
对于中小供应商而言,具体价值十分直接:
更快回收资金;
降低融资压力;
提高收入可预测性;
获得更强的扩容能力;
减少对漫长账期的依赖。
“供多少、结算多少”不只是一项支付功能。
在快速增长阶段,它甚至可能决定一家中小供应商能否持续经营。
六、早期零佣金:降低进入 Token 市场的门槛
在平台市场启动阶段,符合条件的早期供应商可享受 零佣金入驻权益。
这是一种典型的双边市场启动策略。
平台首先需要汇聚足够丰富的供给,才能为需求方提供更多选择、更稳定的容量和更具竞争力的价格。
对于供应商来说,零佣金降低了测试新渠道的成本。
中小模型企业可以先接入服务,观察实际开发者需求,并开始获得调用收入,不必在市场效果尚未得到验证前,就让出一部分交易收益。
这项权益并不只是“少交一点费用”,还意味着:
以更低风险进入新市场;
获得更灵活的价格测试空间;
为利润较薄的服务保留更多收益;
更快触达企业和开发者客户;
让闲置算力更快转化为收入。
当平台拥有更加多元的供应商后,需求方也会得到更多模型选择和更具竞争力的使用价格。
因此,早期激励最终能够同时改善供需两端。
七、API 竞价:让供需共同发现价格
Unexhub 更具市场化特点的设计,是 API 竞价机制。
在传统模式中,供应商公布一个固定价格,用户只能选择接受或离开。
这种方式无法充分反映算力容量的实时变化。
在低谷时段,大量推理资源可能处于闲置状态;到了高峰时段,同一个供应商又可能面临拥堵和容量不足。
动态竞价机制允许价格随供需发生变化。
当供给大于需求时:
供应商可以主动降低价格;
开发者可以更低成本运行弹性任务;
闲置服务器能够产生收入;
批处理与非紧急工作可以转移到低价时段。
当需求大于供给时:
高价值请求可以获得优先容量;
供应商能够获得更高收益;
稀缺基础设施可以被更高效地分配。
这种模式可以进一步区分:
实时任务与可延迟任务;
高峰调用与低谷调用;
通用模型与垂直模型;
普通服务与保障型服务;
高成本与低成本节点。
对于独立开发者和中小企业而言,动态定价意味着,只要任务在时间上具有一定弹性,就可能以更低价格获得优质模型能力。
对于供应商来说,则可以在容量紧张时获得更合理的回报。
最终形成更加高效的 Token 市场。
八、统一接入与毫秒级路由
如果每一种模型都被封闭在独立接口之后,Token 市场就很难形成真正的流动性。
因此,Unexhub 聚合了不同规模的海内外 AI 服务,包括与 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型相关的资源,也为中小企业和独立供应商提供模型分发渠道。
开发者可以通过统一接口调用不同资源。
平台则可根据以下条件对请求进行分配:
模型能力;
调用价格;
当前可用性;
响应延迟;
任务类型;
区域要求;
服务稳定性。
毫秒级路由的目标,是让模型选择过程尽可能对应用无感。
当一家供应商价格上涨、容量紧张或暂时不可用时,路由层可以将合适的任务分配给其他资源,而不需要开发者重新改造整套系统。
在部分模型和调用方案中,开发者最低可享受约为相应官方标准价格或市场价格 30% 的综合使用成本。
这项价格权益在规模化场景中尤其明显。
少量请求的折扣价值有限,但当业务需要消耗数百万甚至数十亿 Token 时,较低的单位价格可以显著改变以下业务的成本结构:
AI 搜索;
编程智能体;
企业知识库;
智能客服;
文档处理;
批量内容生成;
多步骤 Agent 工作流。
平台的目标并不只是提供一次便宜的 API 调用。
而是让长期、高强度的 AI 使用拥有更加可预测的成本。
九、从“卖 GPU”转向“卖 Token”
Google 5140 亿美元的积压订单,释放出一个更深层的商业模式信号。
AI 行业正在从销售静态基础设施,转向销售持续推理服务。
在 AI 发展的早期阶段,GPU 主要用于模型训练。
训练成本巨大,但相对集中。
模型完成训练后,整个产业进入了持续推理时代。
每当用户:
提出一个问题;
生成一段代码;
分析一份文件;
运行一个 Agent;
生成一张图片;
使用一次 AI 搜索;
底层基础设施都需要生产新的模型输出。
因此,算力已经不再只是训练模型的一次性成本。
它成为可以持续生产 Token 的生产性资产。
商业计量单位也随之改变。
客户越来越少只关心供应商拥有多少张芯片,而更加关心:
可以稳定消耗多少推理资源;
Token 能以多快速度交付;
服务能否长期保持可用;
每百万 Token 成本是多少;
工作负载能否被动态路由;
费用能否清晰监控与结算。
AI 产品正在从“使用一个模型”,转向“获得持续的 Token 生产能力”。
部分云平台已经开始分别计价:
输入 Token;
输出 Token;
缓存输入;
优先服务;
批量处理;
预留容量。
这种精细计价,是 Token 商品化的必要前提。
推理资源只有被准确计量,才能进一步被比较、路由、交易和金融化。
十、Unexhub:成为持续 Token 生产的流动性层
Unexhub 不需要训练所有模型,也不需要拥有全球全部 GPU。
它所承担的是另一种角色:
为持续生产的 Token 提供流动性。
对供应商来说,是将闲置或缺少渠道的算力转化为收入。
对需求方来说,是以更加合理的价格获得适合业务的模型资源。
对整个市场来说,则是建立价格发现、智能路由与交易结算机制。
平台的基础设施可以被拆分为几个层次:
供给聚合
汇聚大型模型厂商、中小 AI 企业、独立模型与分散的算力资源。
需求分发
将这些资源连接到需要推理能力的企业、开发者与 AI 应用。
价格发现
通过供应商竞争和动态竞价影响 API 价格。
智能路由
根据成本、速度、能力与可用性,为任务匹配更合适的模型。
交易结算
按照真实 Token 消耗计算费用,并及时结算供应商收入。
企业服务
为企业提供模型选择、接口适配、并发支持、账单管理、工作流接入和持续技术服务。
当这些环节协同运行时,Token 供给将变得更容易被发现和使用。
分散的算力资源,也能逐步形成真正的市场。
十一、为什么 5140 亿美元积压订单比当前收入更重要?
需要明确的是,5140 亿美元代表什么,又不代表什么。
它并不代表 Google 已经收到并确认了 5140 亿美元收入。
积压订单是未来数年内等待履行的合同义务,收入可能按照项目进度逐步确认,部分合同也可能发生调整。
但这个数字仍然非常重要。
因为客户已经把真实预算投入未来的云与 AI 使用。
他们不是在口头表达兴趣,而是在签订合同。
因此,积压订单比以下信号更能证明需求:
调查问卷;
AI 应用预测;
模型基准热度;
产品预约名单;
行业会议声明。
它说明大型企业普遍预期 AI 消耗将继续增长,并愿意提前锁定容量。
将几项财务数据放在一起,会看到一条一致的逻辑:
Google Cloud 收入快速增长;
运营利润率持续提升;
已签约需求不断增加;
资本开支大幅上升;
自由现金流承压;
管理层承认需求超过投资能力。
这些信号共同说明:
Token 需求是真实的,但生产 Token 的供给仍然需要投入巨额资本。
市场不再争论 AI 是否会消耗更多推理资源。
真正的竞争是:谁来融资、生产、分发并高效兑现这些推理需求?
十二、我的判断:Token 正在重估一切
Google Cloud 82% 的收入增速和 5140 亿美元的积压订单,证明了企业 AI 需求的规模与持续性。
负自由现金流,以及接近 2050 亿美元的年度资本开支,则证明了供给建设的难度和成本。
这正是 AI 经济最核心的矛盾:
Token 需求的增长速度,正在超过生产 Token 的基础设施建设速度。
Token 并不是一个虚构概念,也不是资本市场包装出来的泡沫。
它是正在被大规模消耗的数字生产资料。
每一个编程 Agent、企业助手、AI 搜索引擎和自动化工作流,都需要持续获得模型推理能力。
但任何生产资料要实现高效配置,都需要一套成熟市场。
它需要:
透明的价格;
多元的供应商;
稳定的模型路由;
高效的收入结算;
灵活的采购方式;
可规模化的分发渠道。
这正是 Unexhub 希望建设的基础设施层。
通过聚合海内外模型与算力供给、以毫秒级速度完成路由、支持供应商参与市场化竞价、让部分符合条件的资源最低达到约市场价格的 30%,并实时结算供应商调用分成,平台正在尝试把 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。
5140 亿美元积压订单真正表达的,不是 Google 已经赚到了这么多钱。
而是企业已经用真金白银,提前锁定了未来的 AI 消耗。
未来 Token 需求已经被写入合同。
现在市场缺少的,是一套更加透明、高效的 Token 流通机制。
而 Unexhub,正在朝着这个市场走去。