
9 Million Developers, 22 Billion Tokens per Minute: Monetizing the Token Marketplace at Scale in the Gemini Ecosystem
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 gave the market two very different signals.
Total revenue reached $119.796 billion, up 24% year over year. Google Cloud revenue surged 82% to $24.768 billion, while its contracted backlog expanded to $514 billion.
Those figures were impressive. But for developers and AI infrastructure teams, two other numbers were even more revealing:
More than 9 million developers use Gemini models through APIs every month, while those APIs process approximately 22 billion tokens per minute.
One quarter earlier, the figure was around 16 billion tokens per minute. In only three months, throughput increased by nearly 40%.
That means more than 366 million tokens are flowing through Google’s AI infrastructure every second.
Behind every token is a real action:
A developer calls a model;
An Agent reads a document;
An enterprise application searches a knowledge base;
A coding assistant analyzes a repository;
A customer-service system generates a response;
A workflow invokes tools and verifies the result.
Nine million developers represent the demand side of the ecosystem.
Twenty-two billion tokens per minute represent its flow.
Together, they reveal a new commercial reality:
Tokens are no longer merely a technical unit. They are becoming a continuously produced, consumed, priced, and settled digital resource.
The question is no longer whether token demand exists.
The real question is how this demand can be monetized at scale—and how token supply can be priced, routed, and settled efficiently.
1. From API Volume to a Scalable Monetization Model
In the early stage of the large-model market, API usage was mainly viewed as a technical metric.
Developers counted tokens to estimate context length, control model output, and calculate the cost of an individual request.
Today, token throughput has become a commercial indicator.
It reveals:
How often a model is used;
How deeply it is embedded in workflows;
How much inference infrastructure is required;
How much recurring demand exists;
How much revenue may be generated from that demand.
The original material cites an IDC projection suggesting that annual global token consumption could expand dramatically between 2025 and 2030.
The exact forecast depends on how tokens and Peta Tokens are defined, but the broader direction is clear:
Global token consumption is growing nonlinearly.
This is because AI is moving from single-turn conversations into continuous production workflows.
A single request can now trigger multiple stages:
Understand the user’s intent;
Retrieve information;
Read documents or databases;
Call external tools;
Generate an answer;
Validate the output;
Retry failed steps;
Continue with subsequent actions.
One user request may produce dozens of model calls.
One enterprise Agent may execute thousands of tasks every day.
One global application may distribute requests across multiple models, providers, and geographic regions.
As AI enters coding, search, finance, customer service, education, healthcare, and enterprise operations, token consumption grows much faster than the number of users alone would suggest.
Nine million developers are not simply opening a chat window.
They are building products that continuously generate future token demand.
2. Why Token Value Cannot Be Defined by Production Cost Alone
The token-pricing market remains highly fragmented.
A million tokens may have very different prices depending on:
Model intelligence;
Input or output usage;
Cache hits and misses;
Context length;
Response latency;
Geographic region;
Service stability;
Peak demand;
Tool-use capability;
Domain specialization.
More importantly, the business value of a token also differs by use case.
A token used for casual conversation may create little direct revenue.
A token used in pharmaceutical research, financial analysis, contract review, industrial design, or a revenue-generating enterprise Agent may help produce substantially more economic value.
The original article uses an extreme comparison between high-value drug-research tokens and low-cost conversational tokens. Exact prices vary by provider and workload, but the underlying logic is valid:
The value of a token depends not only on how much compute was required to produce it, but also on what the token helps the user accomplish.
This creates two distinct forms of value:
Production value
The cost of GPUs, electricity, networking, storage, inference, and model operations required to generate the token.
Application value
The commercial result generated by that token—such as reducing labor costs, improving customer conversion, accelerating research, or automating a critical workflow.
Current API prices mainly reflect model capability, provider strategy, and infrastructure cost.
A mature token marketplace must also account for:
Supply and demand;
Urgency;
Reliability;
Capacity scarcity;
Service guarantees;
Business requirements.
The industry has already developed enormous token demand, but it still lacks a sufficiently efficient market for real-time price discovery.
3. Three Questions the Token Economy Must Answer
As token volume expands, the market must solve three basic problems:
Who sets the price?
Today, most model providers publish prices independently.
Developers can compare official rate cards, but those prices do not always reflect real-time capacity, regional availability, latency, or idle supply.
Who matches buyers and sellers?
Large providers can build their own distribution networks.
Smaller model companies, independent developers, and regional compute suppliers may have useful capacity but lack direct access to enterprise demand.
Who settles the transaction?
A developer may consume tokens immediately, while the supplier waits weeks or months to receive payment.
That delay creates significant cash-flow pressure, especially for smaller suppliers that must continue paying for electricity, hardware, bandwidth, and operations.
Existing service models solve only part of the problem.
Official model APIs provide direct access but create fragmented accounts and integrations.
Cloud MaaS platforms offer unified enterprise billing and compliance but often operate inside a specific cloud ecosystem.
Aggregation platforms simplify multi-model access, but aggregation alone does not automatically create a complete trading market.
A scalable token economy also requires:
Transparent price discovery;
Dynamic supply-and-demand matching;
Intelligent routing;
Supplier-side liquidity;
Standardized measurement;
Efficient settlement.
This is the infrastructure layer Unexhub is attempting to build.
4. Unexhub: Building an Exchange for Token Supply and Distribution
Unexhub positions itself as a Token aggregation and distribution compute marketplace.
Its purpose is not merely to insert another API gateway between developers and model providers.
Its broader ambition is to create a market layer where fragmented AI capacity can be:
Aggregated;
Compared;
Priced;
Routed;
Purchased;
Settled.
On the supply side, Unexhub brings together providers of different sizes, including access to services associated with mainstream domestic and international models such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, as well as model and compute resources supplied by smaller companies and independent providers.
On the demand side, enterprises and developers can use a unified interface to access multiple model resources without separately maintaining:
Multiple API formats;
Separate authentication systems;
Fragmented balances;
Different billing rules;
Multiple operational processes.
The platform is designed to answer the three core questions of the token economy:
Who discovers the price?
Who connects supply with demand?
Who settles the transaction?
Its solution is built around real-time settlement, early-stage zero commission, dynamic API bidding, and millisecond-level routing.
5. Real-Time Settlement: Giving Suppliers Access to Their Revenue Faster
For small and medium-sized AI suppliers, one of the largest obstacles is not necessarily a lack of demand.
It is the delay between delivering compute and receiving payment.
A supplier may provide inference capacity today but wait one or two months before the revenue is settled.
During that period, it must still pay for:
Electricity;
GPU and server depreciation;
Network bandwidth;
Cloud infrastructure;
Operations;
Technical staff.
Usage may be increasing while available cash continues to shrink.
Unexhub’s real-time settlement model is designed to shorten this cycle.
Its core principle is simple:
Once token capacity has been delivered and consumed, the corresponding supplier share should enter settlement promptly.
This turns delayed accounts receivable into usable working capital.
For suppliers, the practical benefits include:
Faster recovery of operating costs;
Lower cash-flow pressure;
More predictable revenue;
Greater capacity for infrastructure expansion;
Less dependence on long settlement cycles.
The value is not merely “getting paid earlier.”
Faster settlement allows suppliers to reinvest revenue, maintain capacity, and respond more quickly to market demand.
It transforms capital locked in long billing periods into money that can circulate again.
6. Early-Supplier Zero Commission: Lowering the Cost of Entering the Market
During the early development of the marketplace, qualifying suppliers may receive a zero-commission participation benefit.
This is a common but important strategy for building a two-sided market.
The platform first needs sufficient supply to offer buyers real choice, stable capacity, and competitive prices.
For a smaller model company or independent compute provider, zero commission reduces the risk of testing a new distribution channel.
Suppliers can connect resources, observe real demand, and begin earning usage revenue without immediately giving up part of every transaction.
The benefit includes:
Lower market-entry costs;
Better protection for thin margins;
More freedom to test pricing strategies;
Faster access to developer and enterprise demand;
A shorter path from idle capacity to revenue.
As more suppliers join, buyers gain more model options.
As more developers and enterprises use the platform, the channel becomes more valuable to suppliers.
The zero-commission benefit helps the marketplace reach this network effect more quickly.
7. API Bidding: Letting the Market Discover Token Prices
Most model APIs still use fixed pricing.
A provider publishes a price, and the developer decides whether to accept it.
But compute capacity changes constantly.
During off-peak hours, large amounts of inference infrastructure may remain underused.
During periods of heavy demand, the same capacity can become scarce.
A fixed price cannot accurately represent both conditions.
Unexhub’s API bidding model allows prices to respond dynamically to supply and demand.
When capacity is abundant
Suppliers can reduce prices to attract workloads;
Developers can run flexible tasks more cheaply;
Idle compute can be converted into revenue;
Batch tasks can move into lower-cost periods.
When demand rises
High-priority workloads can secure capacity;
Suppliers can earn higher returns;
Scarce resources can be allocated more efficiently.
The market can further distinguish between:
Real-time and delay-tolerant workloads;
Peak and off-peak usage;
General-purpose and specialized models;
Standard and guaranteed service levels;
Low-cost and premium-capacity endpoints.
For independent developers, this creates opportunities to access lower-cost models when tasks are not time-sensitive.
For suppliers, it improves utilization during quiet periods while increasing potential returns during demand peaks.
Tokens begin to move from provider-controlled billing units toward market-priced digital resources.
8. Millisecond-Level Routing: Turning Multiple Models Into One Usable Market
A token marketplace cannot operate efficiently if every model remains isolated behind a separate API.
Unexhub therefore uses a unified access layer and millisecond-level routing to match requests with suitable model capacity.
Routing decisions can take into account:
Model capability;
Current token price;
Available capacity;
Response latency;
Task type;
Regional availability;
Service stability.
A basic text-classification task may not need the most expensive frontier model.
A complex coding or reasoning workflow may require a more capable endpoint.
A flexible batch job may be routed toward lower-cost capacity, while a real-time enterprise request may prioritize speed and service guarantees.
When a provider’s price, capacity, or availability changes, developers do not have to rebuild their applications.
The routing layer can redirect appropriate workloads behind the same interface.
This reduces:
Repeated API development;
Manual model switching;
Account fragmentation;
Operational maintenance;
Billing complexity.
For eligible models and usage plans, developers may access prices as low as approximately 30% of the corresponding official or prevailing market rate.
That benefit is especially significant in high-volume applications.
A discount on a handful of requests has limited value.
Across millions or billions of tokens, lower unit prices can materially reduce the cost of:
AI search;
Coding Agents;
Enterprise knowledge bases;
Customer-service automation;
Document analysis;
Batch content generation;
Multi-Agent workflows.
The goal is not simply to make one API request cheaper.
It is to make sustained and intensive AI usage financially manageable.
9. Scale Is the Foundation; Liquidity Is the Answer
Nine million developers and 22 billion tokens per minute matter because they show that token demand has reached a scale requiring professional market infrastructure.
In the previous phase of AI, GPUs were mainly purchased for training.
Training was expensive but occurred in relatively concentrated cycles.
Today, inference has become continuous production.
Whenever a user:
Asks a question;
Generates code;
Analyzes a document;
Runs an Agent;
Searches with AI;
Calls an enterprise assistant;
the infrastructure must produce new tokens.
Compute is no longer only an upfront cost used to create a model.
It has become a productive asset that continuously generates digital output.
Tokens measure that output.
A market gives it liquidity.
This transformation is changing the commercial model from:
Selling hardware capacity
to:
Selling a continuous stream of model inference.
Customers increasingly care less about the physical number of GPUs and more about:
How many tokens can be delivered;
At what speed;
At what price;
With what level of stability;
Through which model;
Under what settlement terms.
The token is becoming the commercial bridge between infrastructure and application demand.
10. Why Google Cloud’s $514 Billion Backlog Matters
Google Cloud’s reported backlog of $514 billion does not mean that Google has already recognized that amount as revenue.
It represents contracted future commitments that may be delivered and recognized over several years.
Even so, it provides a powerful demand signal.
Customers are not merely expressing interest in AI.
They are committing real budgets to future cloud and inference consumption.
At the contract level, enterprises may be purchasing:
Cloud infrastructure;
Reserved capacity;
Storage;
Databases;
AI model access;
Enterprise software.
At the economic level, much of that capacity will ultimately support future token consumption.
Nearly 90% of Fortune 100 companies are reportedly using Gemini Enterprise.
These organizations are not simply asking how many GPUs they can acquire.
They are effectively asking:
How much reliable AI inference can we consume over the next several years?
As enterprise assistants, coding Agents, customer-service systems, and automated workflows expand, token demand becomes recurring rather than experimental.
The backlog represents an advance commitment to that future demand.
11. From Subscription Revenue to Market-Based Token Trading
The Token economy is developing through several commercial layers.
Subscription
Consumers and teams pay a predictable monthly fee for access to AI services.
This reduces purchasing friction and creates recurring revenue.
Usage-based billing
Developers and enterprises pay according to the number of tokens consumed.
This provides flexibility and clearer unit economics.
Reserved capacity
Large organizations purchase guaranteed throughput, stability, and long-term access.
This improves predictability but may require substantial commitments.
Market-based trading
Suppliers and buyers meet through platforms such as Unexhub, with prices influenced by capacity, demand, latency, model quality, and service requirements.
These models are not mutually exclusive.
They address different needs.
Subscriptions make AI easier to consume.
Usage-based billing provides flexibility.
Reserved capacity provides stability.
Trading markets provide liquidity and price discovery.
As the market matures, Token pricing will become increasingly layered rather than dependent on one universal rate card.
12. What Unexhub Is Really Monetizing
Unexhub is not simply monetizing API access.
It is monetizing several forms of market efficiency.
Supply aggregation
Fragmented model and compute resources become easier for buyers to discover.
Demand distribution
Suppliers gain access to enterprise and developer usage without independently building global sales channels.
Price discovery
API bidding allows the market to reflect current capacity and demand.
Intelligent routing
Requests can be assigned according to price, capability, speed, and availability.
Transaction settlement
Token usage can be measured and supplier revenue settled more promptly.
Enterprise services
Organizations can receive support for model selection, API integration, concurrency, billing, Agent workflows, and ongoing operations.
The deeper product is therefore not a single model.
It is the infrastructure that allows model capacity to flow.
13. My View: Nine Million Developers Are Only the Beginning
Google Cloud’s quarterly revenue of $24.768 billion, its $514 billion backlog, more than 9 million monthly API developers, and Gemini’s 22 billion tokens processed per minute all point in the same direction:
Token demand is real, recurring, and already operating at industrial scale.
Tokens are not a speculative concept.
They are becoming a new production input for the AI era.
Every AI search, coding Agent, enterprise assistant, and automated workflow depends on a continuous stream of model inference.
But a production resource of this scale needs more than model providers and fixed price lists.
It needs a functioning market with:
Diverse supply;
Transparent pricing;
Efficient routing;
Flexible purchasing;
Fast settlement;
Enterprise-grade reliability.
Unexhub is attempting to build that infrastructure by:
Aggregating AI suppliers of different sizes;
Connecting resources associated with Gemini, ChatGPT, Claude, DeepSeek, GLM, and other models;
Providing unified API access;
Routing requests in milliseconds;
Supporting market-based API bidding;
Offering eligible model resources at prices as low as approximately 30% of prevailing rates;
Allowing suppliers to connect once and receive usage-based revenue through real-time settlement;
Providing customized enterprise integration services.
Nine million developers are the sparks on the demand side.
Twenty-two billion tokens per minute are the bloodstream of the ecosystem.
A professional token marketplace is the infrastructure that keeps the entire system circulating.
The real commercialization opportunity does not end when a model processes more tokens.
It begins when those tokens can be priced, routed, purchased, and settled more efficiently.
900 万开发者、每分钟 220 亿 Token:Gemini 生态中 Token 交易市场的规模化变现
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。
公司总营收达到 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;已签约但尚待未来履行的积压订单,则进一步增长至 5140 亿美元。
这组数据已经足够亮眼。
但真正让工程师和 AI 基础设施从业者停下滑轮的,是另外两个数字:
每月有超过 900 万名开发者通过 API 使用 Gemini 模型,Gemini API 每分钟处理约 220 亿 Token。
一个季度前,这个数字还是每分钟约 160 亿。短短三个月,Token 吞吐量增长了近 40%。
每分钟 220 亿 Token,意味着每秒有超过 3.66 亿 Token 在 Google 的 AI 基础设施中流动。
每一个 Token 背后,都是一次真实活动:
开发者调用模型;
Agent 阅读企业文档;
应用检索知识库;
编程助手分析代码仓库;
智能客服生成回答;
自动化流程调用工具并验证结果。
900 万开发者,代表 Gemini 生态的需求侧规模。
每分钟 220 亿 Token,代表这套生态的流动速度。
两组数字共同揭示了一个新的商业现实:
Token 已经不再只是技术计量单位,而是在成为一种被持续生产、消耗、定价与结算的数字资源。
市场现在面对的问题,已经不是 Token 需求是否存在。
真正需要回答的是:如何将如此庞大的需求规模化变现,又如何让 Token 供给获得更加高效的定价、路由与结算机制?
一、从“API 调用量”走向“规模化变现逻辑”
在大模型市场的早期阶段,API 调用量主要是一项技术指标。
开发者通过 Token 数量估算上下文长度、控制输出规模,并计算单次请求的调用费用。
今天,Token 吞吐量已经开始成为商业指标。
它可以反映:
一个模型被使用的频率;
模型进入业务流程的深度;
底层需要建设多少推理基础设施;
市场存在多少持续性需求;
这些需求可能产生多少收入。
原文援引 IDC 的预测认为,2025 年至 2030 年,全球年度 Token 消耗量可能出现极其显著的增长。
具体结果会受到 Token 与 Peta Token 统计定义的影响,但大趋势已经很清晰:
全球 Token 消耗量正在非线性扩张。
原因在于,AI 正从单轮问答走向持续运行的生产流程。
一次用户请求,现在可能触发多个步骤:
理解意图;
检索信息;
阅读文件或数据库;
调用外部工具;
生成答案;
检查输出;
重试失败环节;
继续完成后续动作。
一次请求可能带来数十次模型调用。
一个企业 Agent 每天可能执行数千项任务。
一个全球化应用可能同时在多个模型、供应商和地区之间分发请求。
当 AI 进入编程、搜索、金融、客服、教育、医疗和企业运营后,Token 消耗的增长速度会远远超过用户数量的增长速度。
900 万开发者并不只是打开 Gemini 聊天窗口。
他们正在构建未来会持续消耗 Token 的产品与工作流。
二、Token 的价值不能只由生产成本决定
当前 Token 的价格体系仍然高度分散。
一百万 Token 的价格可能受到以下因素影响:
模型能力;
输入或输出类型;
是否命中缓存;
上下文长度;
响应速度;
服务区域;
稳定性保障;
高峰期需求;
工具调用能力;
垂直领域能力。
更重要的是,Token 在不同业务中的应用价值也完全不同。
用于普通闲聊的 Token,可能不会直接带来多少商业收入。
用于药物研发、金融分析、合同审查、工业设计或收入型企业 Agent 的 Token,却可能帮助企业创造远高于调用成本的价值。
原文使用了药物研发 Token 与低价闲聊 Token 之间的极端比较。具体数字会随模型、服务商和任务发生变化,但其核心逻辑成立:
Token 的价值不仅取决于生产它消耗了多少算力,还取决于它最终帮助用户完成了什么。
Token 因而同时具有两种价值。
生产价值
生成 Token 所需要的 GPU、电力、网络、存储、模型推理和运维成本。
应用价值
Token 最终创造的商业结果,例如减少人工成本、提高客户转化、加快研发速度或自动化关键流程。
目前的 API 价格,主要反映模型能力、供应商策略和基础设施成本。
一个成熟的 Token 市场,还需要考虑:
供需变化;
任务紧急程度;
服务可靠性;
容量稀缺程度;
吞吐保障;
真实业务要求。
Token 需求已经形成规模,但市场仍然缺少足够高效的实时价格发现机制。
三、Token 经济必须回答三个问题
随着 Token 流量不断扩大,市场需要解决三个基础问题。
谁来定价?
目前,大多数模型供应商仍然独立公布价目表。
开发者可以比较不同厂商的官方价格,但这些固定价格未必能够及时反映实际容量、区域供给、响应延迟和闲置算力。
谁来撮合?
大型模型企业可以自己建立销售和分发网络。
中小模型公司、独立开发者和地区性算力供应商,可能拥有优质资源,却缺少触达企业客户的渠道。
谁来结算?
开发者可以即时消耗 Token,但算力供应商可能要等待数周甚至数月才能收到收入。
对于需要持续支付电费、硬件、网络和运维成本的中小供应商而言,漫长账期会造成明显的现金流压力。
现有服务模式只解决了其中一部分问题。
模型官方 API 提供原生访问,但会形成账户和接口分散。
云厂商 MaaS 平台提供统一计费与合规支持,但通常受限于特定云生态。
多模型聚合平台简化了统一接入,但仅仅完成聚合,并不等于建立完整的交易市场。
真正可规模化的 Token 经济,还需要:
透明的价格发现;
动态供需匹配;
智能请求路由;
供应商资金流动性;
标准化 Token 计量;
高效率交易结算。
这正是 Unexhub 希望建设的基础设施层。
四、Unexhub:为 Token 建立聚合与分发交易市场
Unexhub 将自身定位为一个 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和模型供应商之间增加一道 API 网关。
更大的愿景,是建立一套市场基础设施,让分散的 AI 产能能够被:
聚合;
比较;
定价;
路由;
采购;
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 等海内外主流模型相关的资源,也包括中小企业和独立供应商提供的模型与算力能力。
在需求侧,企业与开发者可以通过统一接口访问多个模型,不必再为每家供应商分别维护:
不同的 API 格式;
多套鉴权系统;
分散的账户余额;
不同的计费规则;
重复的运维流程。
Unexhub 希望回答 Token 经济的三个核心问题:
谁来发现价格?
谁来连接供需?
谁来完成结算?
其解决方案主要建立在实时结算、早期零佣金、API 动态竞价和毫秒级路由之上。
五、实时结算:让供应商更快拿回已经产生的收入
对中小 AI 供应商来说,最大的难题未必是没有需求,而是算力交付与资金回收之间存在较长时间差。
供应商今天已经提供推理能力,收入却可能需要等待一两个月才能到账。
在此期间,仍然需要持续支付:
电费;
GPU 与服务器折旧;
网络带宽;
云基础设施;
日常运维;
技术人员成本。
调用量可能持续上升,可周转资金却越来越紧张。
Unexhub 的实时结算机制,旨在缩短这一周期。
其核心逻辑是:
Token 产能已经交付并被真实消耗,相应的供应商分成就应尽快进入结算流程。
这能帮助供应商将长期应收款,更快转化为可使用的营运资金。
具体价值包括:
更快回收运营成本;
显著缓解现金流压力;
提高收入的可预测性;
获得更强的基础设施扩容能力;
降低对漫长结算周期的依赖。
实时结算的意义,不只是“更早收到钱”。
更短的资金周转周期,可以让供应商将收入继续投入容量建设与服务优化,使原本被账期锁住的资金重新流动起来。
六、早期供应商零佣金:降低进入市场的试错成本
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。
这是搭建双边市场的重要策略。
平台需要先聚合足够丰富的供给,才能让需求方获得更多选择、更稳定的容量和更具竞争力的价格。
对于中小模型企业或独立算力供应商而言,零佣金可以降低测试新分发渠道的风险。
供应商可以先接入资源、观察真实需求并获得调用收入,不必从第一笔交易开始就向平台让出部分收益。
其具体价值包括:
降低进入市场的成本;
为较薄的利润空间提供保护;
获得更灵活的定价测试机会;
更快触达开发者和企业需求;
缩短闲置资源转化为收入的路径。
加入平台的供应商越多,需求方获得的模型选择就越丰富。
平台拥有的需求越多,渠道对供应商的价值也越高。
早期零佣金权益,有助于让这一网络效应更快形成。
七、API 动态竞价:让市场发现 Token 的真实价格
大多数模型 API 仍然使用固定价格。
供应商公布价格,开发者决定是否接受。
但推理算力的供给始终在发生变化。
低谷时段,大量算力可能处于闲置状态。
高峰时段,同一批资源又可能迅速变得紧张。
单一固定价格,很难准确反映两种完全不同的供需状态。
Unexhub 的 API 竞价模式,让价格可以根据市场供需动态变化。
当算力充足时
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负载;
闲置算力能够转化为实际收入;
批处理任务可以转移到价格更低的时段。
当需求上升时
高优先级工作负载可以锁定容量;
供应商可以获得更高收益;
稀缺算力能够得到更合理的分配。
市场还可以进一步区分:
实时任务与可延迟任务;
高峰与低谷调用;
通用模型与专业模型;
普通服务与保障型服务;
低价节点与高性能节点。
对独立开发者而言,只要任务具有时间弹性,就可能以更低价格获得优质模型服务。
对供应商而言,则可以提高低谷时段的利用率,并在需求旺盛时获得更合理的回报。
Token 也由此开始从模型厂商单方面定价的计费单位,转向由市场供需发现价格的数字资源。
八、毫秒级路由:把多个模型变成一个可用市场
如果每个模型都被封闭在独立 API 后面,Token 市场就很难高效运转。
Unexhub 因此通过统一接入层和毫秒级路由,为请求匹配更加合适的模型资源。
路由决策可以综合考虑:
模型能力;
当前 Token 价格;
剩余算力容量;
响应延迟;
任务类型;
区域可用性;
服务稳定性。
简单的分类任务,没有必要始终调用价格最高的旗舰模型。
复杂的编程与推理工作流,则可能需要能力更强的端点。
时间要求不高的批处理任务,可以被分配到成本更低的算力;实时企业请求,则可以优先选择低延迟和高可用资源。
当某个供应商的价格、容量或可用性发生变化时,开发者不必重构整个应用。
路由层可以在相同接口背后,将合适的任务切换到其他资源。
这能够减少:
重复的 API 开发;
手动切换模型;
账户与余额分散;
日常运维投入;
账单管理复杂度。
在部分符合条件的模型与调用方案中,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% 的综合调用成本。
这一权益在高调用量场景中更为明显。
少量请求的价格差异可能有限。
但当业务持续消耗数百万、数亿甚至数十亿 Token 时,更低的单位成本可以显著优化以下应用的预算:
AI 搜索;
编程 Agent;
企业知识库;
智能客服;
文档分析;
批量内容生成;
多智能体工作流。
平台的目标并不只是让某一次 API 请求更便宜。
而是让长期、高频、规模化的 AI 使用,真正进入企业和开发者能够承受的成本区间。
九、规模是前提,流动性才是答案
900 万开发者与每分钟 220 亿 Token 的意义在于:
Token 需求已经达到需要专业市场基础设施的规模。
在上一阶段,企业购买 GPU,主要用于模型训练。
训练成本虽然巨大,但通常集中在某些周期。
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;
分析文件;
启动 Agent;
使用 AI 搜索;
调用企业助手;
底层基础设施都需要生产新的 Token。
算力不再只是用于创造模型的一次性投入。
它正在成为持续生产数字输出的生产性资产。
Token 衡量这种产出。
交易市场则为这种产出提供流动性。
整个商业模式正在从:
销售硬件容量
转向:
销售持续的模型推理能力。
客户越来越少只关心供应商有多少张 GPU,而开始更加关心:
能稳定交付多少 Token;
交付速度有多快;
单位价格是多少;
服务能否保持稳定;
应该选择哪个模型;
采用什么结算条件。
Token 正在成为连接算力基础设施与应用需求的商业桥梁。
十、为什么 5140 亿美元积压订单如此重要?
Google Cloud 的 5140 亿美元积压订单,并不代表 Google 已经确认了同等规模的收入。
这部分订单是将在未来数年内逐步交付和确认的合同承诺。
即便如此,它仍然是一个非常强烈的需求信号。
企业客户不再只是对 AI 表达兴趣,而是已经用真实预算锁定未来的云服务与推理能力。
从合同形式看,企业采购的可能是:
云基础设施;
预留容量;
数据库存储;
网络服务;
模型调用;
企业 AI 软件。
但在经济底层,其中很大一部分资源最终都会支撑未来的 Token 消耗。
接近 90% 的《财富》100 强企业,据称已经开始使用 Gemini Enterprise。
这些企业真正关心的,不只是可以买到多少 GPU,而是:
未来几年,我们能够稳定消耗多少 AI 推理能力?
当企业助手、编程 Agent、智能客服和自动化流程不断扩张时,Token 需求会从实验性支出转变为周期性生产成本。
5140 亿美元的积压订单,正是对这部分未来需求的提前锁定。
十一、从订阅收入走向市场化 Token 交易
Token 经济正在形成多层商业模式。
订阅制
个人和团队支付可预测的月费,获得 AI 服务使用权。
这可以降低使用门槛,并为平台带来持续收入。
按量计费
开发者和企业按照真实 Token 消耗付款。
这种模式更加灵活,单位成本也更清晰。
预留容量
大型企业提前购买稳定吞吐、服务保障和长期容量。
这有助于提高可预测性,但通常需要更高承诺。
市场化交易
供给方与需求方通过 Unexhub 等平台进行撮合,价格受到算力容量、市场需求、响应延迟、模型质量和服务要求影响。
这些模式不是互相替代,而是在解决不同问题。
订阅制让 AI 更容易被消费。
按量计费提供灵活性。
预留容量提供稳定性。
交易市场则提供流动性和价格发现。
随着市场成熟,Token 定价不会只依赖一张固定价目表,而会逐渐形成多层、分场景的价格体系。
十二、Unexhub 真正变现的是什么?
Unexhub 变现的并不只是 API 接入能力。
它真正提供的是多种市场效率。
供给聚合
让分散的模型与算力资源更容易被需求方发现。
需求分发
让供应商无需独立搭建全球销售体系,也能获得企业和开发者调用。
价格发现
通过 API 竞价,让价格能够反映实时容量与需求。
智能路由
根据价格、能力、速度和可用性,为任务匹配模型。
交易结算
准确计量 Token 用量,并更及时地向供应商结算收入。
企业服务
围绕模型选择、API 对接、并发支持、统一账单、Agent 工作流和持续运维,提供定制化服务。
因此,Unexhub 更深层的产品并不是某一个模型。
而是让模型能力得以持续流通的基础设施。
十三、我的判断:900 万开发者只是开始
Google Cloud 单季度 247.68 亿美元收入、5140 亿美元积压订单、每月超过 900 万名 API 开发者,以及 Gemini 每分钟处理 220 亿 Token,共同指向了一个结论:
Token 需求是真实的、持续的,并且已经进入工业化规模。
Token 不是一个资本市场包装出来的抽象概念。
它正在成为 AI 时代的新型生产资料。
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,不能只依靠分散的模型供应商和固定价目表完成配置。
它需要一个成熟市场,提供:
多元供给;
透明价格;
高效路由;
灵活采购;
快速结算;
企业级稳定性。
Unexhub 正在尝试建设这一市场基础设施:
聚合不同规模的 AI 供应商;
连接 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型资源;
提供多模型统一 API 接入;
以毫秒级速度完成请求路由;
支持市场化 API 动态竞价;
让部分符合条件的模型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与定制服务。
900 万开发者,是需求侧的星火。
每分钟 220 亿 Token,是 Gemini 生态的流动命脉。
而专业的 Token 交易市场,则是让整个系统持续循环的基础设施。
AI 商业化的机会,并不会在模型处理更多 Token 时结束。
恰恰相反,当这些 Token 能够以更加透明、高效的方式被定价、路由、采购和结算时,真正的规模化变现才刚刚开始。