AI Mode 单次应答成本降至最低,Token 交易市场迎来搜索广告变现的黄金拐点

The headline numbers looked impressive: revenue reached $119.8 billion, up 24% year over year, while Google Search and other related businesses generated approximately $63.3 billion, an increase of 17%.

When the Cost of a Single AI Response Approaches Zero, Search Becomes Harder to Monetize

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 its fiscal Q2 2026 results.

The headline numbers looked impressive: revenue reached $119.8 billion, up 24% year over year, while Google Search and other related businesses generated approximately $63.3 billion, an increase of 17%.

AI Overviews and AI Mode continued to expand search usage, bringing more users and more complex questions into Google’s ecosystem.

Yet the more important story was not simply that people were searching more.

It was that the economics underneath every search were being rewritten.

According to some interpretations of the earnings data, AI Mode had surpassed one billion monthly active users. Alphabet’s official disclosures, however, referred to approximately 950 million monthly active users for the Gemini app and 22 billion tokens processed through its APIs every minute.

The reporting scopes may differ, but the direction is clear:

AI search has entered the billion-user era, and every generated answer now carries a measurable token cost.

Behind each query, tokens are being processed, routed, cached, generated, priced, and settled.

The continued decline in the cost of a single AI response is not simplifying the search business.

It is pushing the entire token economy toward a new—and more complicated—commercial threshold.


1. How Did the Cost of AI Responses Fall So Quickly?

One of the most important changes was the shift in default model strategy.

Rather than routing every query through the most capable and expensive model, AI search systems increasingly use lighter, faster, and more cost-efficient models for everyday requests.

In Google’s case, the default model behind certain AI Mode experiences reportedly moved toward Gemini 3.5 Flash-Lite, positioned as one of the fastest and most economical models in the 3.5 family.

This was not a routine product replacement.

It was an exercise in compute economics.

Most search queries do not require the maximum reasoning capacity of a frontier model. A request for a restaurant comparison, a product summary, or a short explanation can often be handled by a smaller model at a fraction of the cost.

Only the most complex queries need to be escalated to a stronger system.

This layered approach allows the platform to:

  • Route simple questions to lower-cost models;

  • Reserve frontier models for difficult requests;

  • Cache repeated context and common information;

  • Increase batch processing where real-time output is unnecessary;

  • Improve hardware utilization across different demand periods.

As models, chips, inference frameworks, and caching systems improve together, the cost of generating a token continues to decline.

The inference cost of earlier GPT-generation models has reportedly fallen hundreds of times from its initial level. Across the broader market, model API prices between 2022 and 2026 declined by as much as 99.9% in some categories.

In China, cost-focused models such as DeepSeek V4 Flash have pushed prices into the range of approximately RMB 1 per million input tokens and RMB 2 per million output tokens, according to the figures cited in the original material.

Exact prices continue to change, but the economic trend is more important than any single price point:

AI inference is becoming dramatically cheaper.

And when something becomes cheaper, people tend to use more of it.


2. Lower Prices Do Not Reduce Spending—They Unlock More Demand

China Academy of Information and Communications Technology data cited in the source material shows that average daily token usage in China rose from approximately 100 billion tokens at the beginning of 2024 to around 140 trillion by the end of March 2026.

That represents growth of more than one thousand times.

This is a classic example of an efficiency rebound.

When the cost of each response falls, companies do not simply maintain the same usage and save money.

They add new use cases:

  • More employees receive access;

  • More documents are processed;

  • More customer conversations are automated;

  • More code is generated and tested;

  • More agent workflows run in the background;

  • More searches become conversational and multi-step.

The result is a self-reinforcing flywheel:

  1. Inference becomes cheaper;

  2. More applications become financially viable;

  3. Usage rises rapidly;

  4. Larger volumes support further infrastructure optimization;

  5. Lower costs create even more demand.

AI Mode strengthens this cycle because every search interaction may involve multiple stages of computation:

  • Understanding the user’s intent;

  • Retrieving relevant information;

  • Reading source content;

  • Ranking and comparing evidence;

  • Generating a structured answer;

  • Responding to follow-up questions.

One search may consume thousands or tens of thousands of tokens.

The cheaper each response becomes, the more freely users ask complex questions. The more questions they ask, the faster overall token consumption grows.

The cost per response may approach zero.

The total cost of operating the system does not.


3. Search Volume Is Rising, but Advertising Is Becoming Harder to Sell

This is where the economics become uncomfortable.

Traditional search was built around traffic distribution.

A user entered a keyword, Google displayed links and advertisements, the user clicked a result, and advertisers paid for access to that traffic.

AI search changes user behavior.

Instead of showing ten blue links and asking users to assemble the answer themselves, AI Mode can generate a complete response inside the search page.

The user may no longer need to visit an external website.

Traditional search already contains a large share of zero-click queries. According to the figures cited in the original article, the zero-click rate within AI Mode may reach as high as 93%.

That would mean fewer than one in ten searches sends a user to an external destination.

This presents a structural challenge.

The AI answer may improve the user experience, but it also weakens the click-based chain on which search advertising and publisher traffic have historically depended.

The system pays for inference before it knows whether the user will click anything.

Every AI-generated response creates a cost.

Not every response creates an advertising event.

AI Overviews has expanded across more than 200 countries and regions and reportedly increased usage for certain types of displayed queries by more than 10%.

At the same time, Alphabet’s infrastructure spending has risen sharply.

According to the figures cited in the source material, second-quarter capital expenditure reached approximately $44.9 billion, nearly doubling year over year, while full-year guidance increased to between $195 billion and $205 billion.

That spending supports:

  • Data centers;

  • AI accelerators;

  • Networking infrastructure;

  • Power capacity;

  • Cooling systems;

  • Model training;

  • Real-time inference.

The numbers reveal a difficult equation:

Search revenue may be growing by 17%, while the infrastructure required to protect that revenue is growing much faster.

AI search can create more engagement and more queries, but it also changes the margin structure of every response.

The search business is no longer only competing for attention.

It is competing to produce each unit of intelligence at the lowest sustainable cost.


4. Why Near-Zero Response Costs Make the Business More Difficult

At first glance, a lower response cost should make AI search easier to monetize.

In reality, it can make competition more intense.

When the cost of producing an AI answer falls:

  • More companies can launch AI search products;

  • Free usage becomes easier to offer;

  • Users expect richer answers without paying more;

  • Competitors can copy basic features faster;

  • The market price of a standard answer falls;

  • Differentiation moves toward data, distribution, speed, and workflow integration.

In other words, lower production costs can reduce the scarcity of the product.

A basic AI answer becomes abundant.

What remains scarce is:

  • Reliable access to high-quality models;

  • Low-latency infrastructure;

  • Accurate and current data;

  • Efficient routing;

  • Stable capacity during peak demand;

  • Trusted enterprise integration;

  • Transparent billing and settlement.

The commercial value therefore shifts away from the response itself and toward the infrastructure that produces and distributes it.

This is why the token economy matters.

When a single response becomes almost free, the decisive question becomes:

Who can process, route, and settle trillions of tokens more efficiently than everyone else?


5. A Token Is Not Worth the Same in Every Scenario

The economic value of tokens varies enormously by use case.

A token used in casual conversation may have very little commercial value.

A token used to support drug discovery, financial decisions, legal analysis, industrial design, or a revenue-generating enterprise agent may be worth far more.

The original text uses an extreme comparison: tokens used in certain high-value drug-research scenarios may be priced at up to $1,000 per million, while low-value conversational tokens may cost only a fraction of a dollar.

The exact numbers vary by model, supplier, service level, and application, but the underlying point is valid:

The value of a token depends not only on what it costs to produce, but on what it helps the user accomplish.

This is similar to electricity.

One kilowatt-hour has a production cost, but its economic value differs depending on whether it powers a household light, a hospital operating room, or a high-value industrial process.

Tokens behave in a similar way.

Their market value depends on:

  • Model quality;

  • Response latency;

  • Output reliability;

  • Domain specialization;

  • Context length;

  • Tool access;

  • Availability guarantees;

  • Business impact.

This creates the foundation for a more sophisticated token market.

Instead of treating every token as an identical technical unit, the market can price different forms of inference according to capability, urgency, quality, and demand.


6. Unexhub: Turning AI Costs Into Tradable Capacity

This is the core problem Unexhub is attempting to address.

Unexhub is positioned as a token aggregation and distribution compute marketplace, connecting model and compute suppliers with enterprises and developers that need AI capacity.

Rather than relying only on fixed prices determined by individual model providers, the platform aims to aggregate fragmented supply and allow pricing to emerge through market demand.

Its infrastructure focuses on several mechanisms.

Real-Time Supplier Settlement

For small and medium-sized AI suppliers, a long settlement cycle can create severe cash-flow pressure.

Compute may be delivered today, while payment arrives one or two months later.

During that period, the supplier still needs to cover:

  • Electricity;

  • Hardware depreciation;

  • Bandwidth;

  • Operations;

  • Staffing;

  • Cloud-resource costs.

Unexhub’s real-time settlement model is intended to shorten the distance between usage and payment.

When capacity is consumed, the corresponding supplier share can be calculated and settled promptly.

For smaller suppliers, this means that revenue does not remain trapped in a long accounts-receivable cycle.

The practical benefit is not merely faster payment.

It is stronger liquidity.

Funds can be reused to expand capacity, maintain infrastructure, and support new demand.

Early-Supplier Zero Commission

During the platform’s early market-building stage, qualifying suppliers may receive a zero-commission entry benefit.

This lowers the cost of opening a new distribution channel.

A smaller model company or independent compute provider can connect its capacity, test market demand, and begin generating usage revenue without immediately losing a portion of each transaction to platform fees.

The value is especially meaningful for suppliers operating on thin margins.

They receive:

  • Lower entry costs;

  • Faster exposure to developer demand;

  • More room to experiment with pricing;

  • A shorter path from capacity to revenue.

Market-Based API Bidding

Unexhub also introduces an API bidding mechanism intended to allow prices to change with supply and demand.

When compute capacity is underused, suppliers can offer lower prices to attract flexible workloads.

Developers can move batch tasks, testing, data processing, or content generation into lower-cost periods.

When demand rises, suppliers may receive higher returns for scarce capacity.

This creates a more efficient allocation system.

Idle infrastructure can become revenue.

Flexible demand can receive better prices.

High-priority demand can obtain capacity by paying a premium.

Rather than forcing every request into one fixed price, the market can distinguish between:

  • Peak and off-peak demand;

  • Real-time and delay-tolerant tasks;

  • General and specialized models;

  • Standard and guaranteed service levels.


7. From Token Aggregation to a Token Trading Infrastructure

Token aggregation platforms have already demonstrated that developers want access to multiple models through one interface.

Platforms represented by OpenRouter have grown rapidly by consolidating model access, billing, and routing.

According to the figures cited in the article, weekly token traffic on large aggregation platforms has reached tens of trillions of tokens. Different public reports use different measurement periods and model categories, so exact totals should be checked against the latest official data.

The important point is not one specific weekly figure.

It is the scale.

AI applications are no longer consuming millions of tokens.

They are consuming trillions.

At that level, the market needs more than a directory of APIs.

It needs:

  • Transparent price discovery;

  • Capacity allocation;

  • Standardized quality levels;

  • Real-time routing;

  • Unified billing;

  • Supplier settlement;

  • Risk controls;

  • Enterprise service guarantees.

Unexhub’s larger ambition is to move from basic model aggregation toward a more market-oriented token infrastructure.

The platform brings together AI suppliers of different sizes, including services associated with Gemini, ChatGPT, Claude, DeepSeek, GLM, as well as models and capacity provided by smaller companies and independent suppliers.

Developers can access these resources through a unified interface and pay according to actual token consumption.

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

This is particularly valuable in high-volume scenarios.

When an application consumes millions or billions of tokens, a small unit-price difference becomes a substantial operating-cost difference.

Lower access prices can support:

  • Larger development and testing cycles;

  • Longer-context workloads;

  • Higher-volume customer-service agents;

  • Batch processing;

  • AI search;

  • Code generation;

  • Enterprise knowledge bases;

  • Multi-agent automation.

The platform also uses millisecond-level routing to match requests with appropriate model resources according to price, capability, availability, and latency.

The user does not need to rebuild the application every time a provider’s price or capacity changes.

The routing layer can make the adjustment behind the interface.


8. The Golden Turning Point Is Not “Lowest Cost”—It Is Repricing

When the cost of an AI response falls far enough, the token market reaches a critical point.

The key word is not simply “cheap.”

It is repricing.

When inference was expensive, companies focused on reducing usage.

When inference becomes extremely cheap, they begin to reconsider which activities can be rebuilt around AI.

Search is only one example.

The same transition is happening in:

  • Software development;

  • Customer service;

  • Advertising;

  • Education;

  • Healthcare;

  • Financial analysis;

  • Legal services;

  • Enterprise operations.

A task that was uneconomical at $10 per execution may become attractive at $0.10.

Once millions of companies make that calculation, demand does not grow linearly.

It accelerates.

This creates what could be described as token inflation—not inflation in price, but inflation in total token volume.

More applications consume more tokens.
More users trigger more workflows.
More agents operate for longer periods.
More business processes become continuously AI-assisted.

The supply chain must then be repriced around:

  • Compute capacity;

  • Model efficiency;

  • Caching;

  • Routing;

  • Latency;

  • Availability;

  • Settlement speed.

The search business is therefore shifting from monetizing clicks toward managing the economic value of token flows.

That does not mean advertising disappears.

It means the advertising model must now absorb an additional layer of economics.

A company must earn enough from the user’s attention, intent, transaction, or subscription to cover the ongoing cost of generating answers.


9. Unexhub’s Role in the Repricing of AI Infrastructure

Unexhub is attempting to build the infrastructure required for this new stage.

For developers, the platform provides:

  • Unified access to multiple mainstream models;

  • Usage-based token billing;

  • Competitive prices for eligible workloads;

  • Millisecond-level routing;

  • Fewer separate accounts and integrations;

  • Greater flexibility when models or prices change.

For suppliers, it provides:

  • A simplified connection to market demand;

  • Real-time usage settlement;

  • Early-stage zero-commission incentives;

  • Dynamic pricing opportunities;

  • A broader distribution channel.

For enterprise customers, it can provide customized services involving:

  • Model selection and routing;

  • API adaptation;

  • Agent-workflow integration;

  • Concurrency management;

  • Stability support;

  • Unified billing;

  • Cost visibility;

  • Ongoing technical assistance.

The platform has developed service processes for customized AI integration involving different types of organizations, including large domestic and international companies, publicly listed businesses, and state-owned enterprises.

The value proposition is therefore broader than “cheaper tokens.”

A low price may attract a user once.

A reliable infrastructure layer determines whether the user can continue operating at scale.

The larger goal is to help turn fragmented global model capacity into a resource that can be more easily discovered, compared, routed, consumed, and settled.


10. My View: When Responses Become Almost Free, the Real Game Begins

Alphabet’s results show that AI can drive more search activity and support continued revenue growth.

They also show that the infrastructure required to deliver that experience is becoming enormously expensive.

This is the central contradiction of AI search:

The cost of each individual response is falling, while the total cost of serving the market is rising.

The unit becomes cheaper.

The volume becomes larger.

The product becomes more capable.

The advertising chain becomes less direct.

The capital requirements become heavier.

That is why near-zero response cost does not make the search business easy.

It forces the industry to rebuild its economics around scale, routing, pricing, and token liquidity.

One billion users are not merely searching.

They are continuously generating demand for model inference.

Every query consumes tokens.
Every token consumes compute.
Every unit of compute must be financed.
Every answer must eventually be monetized.

The next decade of AI search may not be controlled only by the company with the best model or the largest advertising network.

It may also be shaped by the platforms that make token supply more transparent, liquid, and efficient.

That is the opportunity Unexhub is pursuing:

  • Aggregating global AI supply;

  • Connecting enterprises and developers through unified APIs;

  • Routing requests in milliseconds;

  • Allowing suppliers to compete on price and service;

  • Charging users according to actual token consumption;

  • Offering selected resources at prices as low as roughly 30% of prevailing rates;

  • Settling supplier revenue in real time.

When the cost of one AI response approaches zero, the economics of search do not disappear.

They become more complex.

And that is when the real game begins.


当单次响应成本趋近于零,搜索生意反而更难做了

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 发布了 2026 财年第二季度财报。

表面数字依然亮眼:总营收达到 1198 亿美元,同比增长 24%;Google 搜索及其他相关业务收入约为 633 亿美元,同比增长 17%。

AI Overviews 与 AI Mode 继续带动搜索使用量增长,让更多用户开始用自然语言提出更复杂的问题。

但真正值得关注的,并不只是搜索次数增加了多少。

而是每一次搜索背后的经济逻辑,正在被重新书写。

部分财报解读称,AI Mode 的月活跃用户已经突破 10 亿;Alphabet 官方公开材料则提到,Gemini 应用月活约为 9.5 亿,API 每分钟处理约 220 亿 Token。

不同数字可能对应不同产品与统计口径,但趋势已经非常明确:

AI 搜索正在进入十亿级用户时代,每一次生成答案,都开始对应一笔可以被计算的 Token 成本。

每次查询背后,Token 都在被处理、路由、缓存、生成、计价和结算。

单次 AI 响应成本不断下降,并没有让搜索生意变得更简单。

相反,它正在把整个 Token 经济推向一个更加复杂的商业临界点。


一、AI 响应成本是怎样降下来的?

其中一个关键变化,是搜索系统开始采用分层模型策略。

不再把所有问题都交给能力最强、价格最高的旗舰模型,而是将日常查询分配给更轻、更快、成本更低的模型。

在部分 AI Mode 场景中,默认模型逐渐转向 Gemini 3.5 Flash-Lite。Google 将其定位为 3.5 系列中速度更快、性价比更高的模型之一。

这不是一次普通的产品替换。

而是一场精密的算力经济学计算。

大多数搜索问题并不需要旗舰模型的全部推理能力。

餐厅对比、产品总结、概念解释等常规问题,可以由更小的模型以更低成本完成。

只有真正复杂的任务,才需要被升级到能力更强的模型。

这种分层策略可以帮助平台:

  • 把简单问题交给低成本模型;

  • 把旗舰算力留给复杂任务;

  • 对重复上下文和常见信息进行缓存;

  • 将非实时任务放入批处理队列;

  • 在不同需求时段提高硬件利用率。

当模型架构、芯片、推理框架和缓存系统同时进步时,生成一个 Token 的成本便会持续下降。

早期 GPT 系列模型的推理成本,相比最初阶段据称已经降低数百倍。放眼整个行业,2022 年至 2026 年之间,部分模型 API 单价的下降幅度甚至达到 99.9%。

在国内市场,DeepSeek V4 Flash 等成本型模型,已经将价格压低至约每百万输入 Token 1 元、每百万输出 Token 2 元的区间。

具体价格会持续变化,但比某一个价格数字更重要的是背后的趋势:

AI 推理正在快速变便宜。

而任何东西一旦变便宜,人们通常不会少用,而是会用得更多。


二、价格下降并不会减少支出,而会释放更多需求

根据原文援引的中国信通院数据,国内日均 Token 调用量从 2024 年初约 1000 亿,增长到 2026 年 3 月底约 140 万亿,增幅超过一千倍。

这是典型的效率反弹效应。

当每次调用的成本下降时,企业并不会只维持原来的使用量,然后节省预算。

它们会增加更多应用:

  • 让更多员工获得 AI 权限;

  • 处理更多文件;

  • 自动化更多客户对话;

  • 生成和测试更多代码;

  • 在后台运行更多 Agent;

  • 把更多传统搜索升级成对话式搜索。

最终形成一个自我强化的飞轮:

  1. 推理成本降低;

  2. 更多应用变得经济可行;

  3. Token 使用量快速增长;

  4. 更大规模进一步推动基础设施优化;

  5. 更低成本继续创造新需求。

AI Mode 会强化这个循环,因为每一次搜索可能都包含多个计算阶段:

  • 理解用户意图;

  • 检索相关信息;

  • 阅读来源内容;

  • 对证据进行比较与排序;

  • 生成结构化答案;

  • 继续处理用户追问。

一次搜索可能消耗数千甚至数万 Token。

单次响应越便宜,用户越愿意提出复杂问题;问题越多,整体 Token 消耗增长得越快。

单次响应成本可以趋近于零。

但整个系统的运营成本不会。


三、搜索量增加了,广告却变得更难卖了

真正棘手的问题由此出现。

传统搜索建立在流量分发之上。

用户输入关键词,Google 展示链接和广告;用户点击结果,广告主为这次点击付费。

AI 搜索正在改变这种行为。

它不再只提供十条蓝色链接,让用户自己寻找答案,而是直接在搜索结果页中生成完整内容。

用户可能不再需要访问外部网站。

传统搜索本来就存在大量零点击查询。按照原文引用的数据,AI Mode 内部的零点击率甚至可能达到 93%。

这意味着每十次搜索中,可能不到一次会将用户送往外部网站。

这构成了一个结构性难题。

AI 回答改善了用户体验,却削弱了搜索广告和内容网站过去赖以生存的点击链路。

系统在用户看到答案之前,就已经支付了推理成本。

每一条 AI 回答都会产生费用。

但并不是每一条回答都能产生广告点击。

AI Overviews 已经覆盖全球 200 多个国家和地区,并使部分展示 AI 摘要的查询量增长超过 10%。

与此同时,Alphabet 的基础设施投入也在快速上升。

按照原文引用的数据,其第二季度资本开支达到约 449 亿美元,同比接近翻倍,全年资本开支指引上调至 1950 亿至 2050 亿美元。

这些资金被用于:

  • 数据中心;

  • AI 加速芯片;

  • 网络基础设施;

  • 电力容量;

  • 冷却系统;

  • 模型训练;

  • 实时推理。

这组数字揭示了一个难以回避的算式:

搜索业务收入可能增长了 17%,但保护这部分收入所需的基础设施成本,增长得更快。

AI 搜索可以带来更多查询和更高参与度,却也改变了每一次响应的利润结构。

搜索业务不再只是争夺用户注意力。

它还要竞争:谁能以最低的可持续成本,生产每一单位智能。


四、为什么单次响应越便宜,生意反而越难做?

从直觉上看,响应成本下降,AI 搜索应该更容易赚钱。

现实却可能恰恰相反。

当生成 AI 答案的成本下降后:

  • 更多公司能够推出 AI 搜索产品;

  • 免费服务更容易长期提供;

  • 用户期待更丰富的答案,却不愿支付更多费用;

  • 基础功能更容易被竞争对手复制;

  • 普通 AI 答案的市场价格不断下降;

  • 竞争焦点转向数据、渠道、速度与业务整合。

换句话说,生产成本下降,也会降低产品的稀缺性。

一个普通的 AI 答案正在变得充足。

真正稀缺的资源开始变成:

  • 稳定的高质量模型供给;

  • 低延迟基础设施;

  • 准确且及时的数据;

  • 高效率路由;

  • 高峰期的容量保障;

  • 可信的企业集成;

  • 透明的账单与结算。

商业价值因此从“答案本身”,转移到“生产和分发答案的基础设施”。

这也是 Token 经济越来越重要的原因。

当单次回答几乎免费时,决定胜负的问题变成了:

谁能够比其他人更高效地处理、路由和结算数万亿 Token?


五、不同场景中的 Token,并不具有相同价值

Token 的经济价值,会随着应用场景发生巨大变化。

用于普通聊天的 Token,商业价值可能非常低。

用于药物研发、金融决策、法律分析、工业设计或企业智能体的 Token,则可能创造远高于调用成本的收益。

原文使用了一个极端对比:部分药物研发场景中的 Token,价值可能达到每百万 1000 美元;普通闲聊 Token 的价格,则可能只有几美分甚至更低。

具体价格会因模型、供应商、服务等级和使用场景不同而变化,但其背后的逻辑成立:

Token 的价值不只取决于生产成本,还取决于它最终帮助用户完成了什么。

这与电力市场十分相似。

一度电具有生产成本,但它用于家庭照明、医院手术室或高价值工业生产时,产生的经济价值并不相同。

Token 也是如此。

它的市场价值取决于:

  • 模型质量;

  • 响应延迟;

  • 结果可靠性;

  • 垂直领域能力;

  • 上下文长度;

  • 工具调用权限;

  • 可用性保障;

  • 最终业务价值。

这为更复杂的 Token 交易市场奠定了基础。

市场不必再将所有 Token 视为完全相同的技术单位,而可以根据能力、时效、质量与需求,为不同推理资源定价。


六、Unexhub:把 AI 成本转化为可交易的算力资源

这正是 Unexhub 试图解决的核心问题。

Unexhub 的定位,是建设一个 Token 聚合与分发算力交易中心,连接提供模型和算力的供应方,以及需要 AI 能力的企业与开发者。

不同于完全由单一模型厂商决定固定价格的模式,Unexhub 希望聚合分散的算力供给,让 Token 价格在真实市场需求中形成。

平台重点构建了几项核心机制。

供应商实时结算

对于中小型 AI 供应商而言,过长的结算周期会带来明显的现金流压力。

算力今天已经交付,收入却可能要一两个月以后才能到账。

在此期间,供应商仍然需要支付:

  • 电费;

  • 硬件折旧;

  • 网络带宽;

  • 运维成本;

  • 人员费用;

  • 云资源成本。

Unexhub 的实时结算模式,希望缩短调用发生与收入到账之间的距离。

算力被使用后,系统便可以及时计算供应商应得的调用分成并进行结算。

对于中小供应商而言,这不只是“回款更快”。

更重要的是提高资金流动性。

回收的资金可以继续投入扩容、设备维护和新业务供给,让原本被账期锁住的“死钱”,重新变成可以周转的“活钱”。

早期供应商零佣金

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。

这降低了测试新分发渠道的成本。

中小模型企业或独立算力供应商可以先接入资源、验证需求并获得调用收入,不必在市场效果尚未确定之前,就先让出一部分交易收益。

其具体价值包括:

  • 降低进入市场的成本;

  • 更快触达开发者需求;

  • 为低利润供应商保留更多收益空间;

  • 提供更灵活的定价测试条件;

  • 缩短算力资源变现路径。

市场化 API 竞价

Unexhub 还希望通过 API 竞价机制,让价格根据供需动态变化。

当算力处于空闲状态时,供应商可以通过更低价格吸引弹性任务。

开发者则可以把批处理、测试、数据加工或内容生成等非紧急工作,安排在成本更低的时段。

当需求上升时,供应商可以根据稀缺容量获得更高收益。

这会形成更有效率的资源配置方式。

闲置基础设施可以转化为收入;
弹性需求可以获得更低价格;
高优先级需求可以通过溢价获得稳定容量。

平台不必让所有请求都接受同一种固定价格,而可以区分:

  • 高峰与低谷需求;

  • 实时与可延迟任务;

  • 通用与专业模型;

  • 普通与保障型服务。


七、从 Token 聚合,走向 Token 交易基础设施

Token 聚合平台已经证明,开发者确实需要通过一套接口调用多个模型。

以 OpenRouter 为代表的平台,通过统一接入、路由和账单管理快速增长。

根据文章引用的行业数据,大型聚合平台的每周 Token 流量已经达到数十万亿级别。不同公开报告采用的统计周期和模型范围可能不同,具体数字仍应以平台最新披露为准。

真正重要的,并不是某一周究竟是 26.9 万亿还是 54.6 万亿。

而是这个市场已经进入万亿级 Token 流通时代。

当 AI 应用的使用量不再以百万 Token 计算,而是以万亿 Token 计算时,市场需要的不只是一个 API 目录。

它还需要:

  • 透明的价格发现;

  • 算力容量分配;

  • 标准化服务等级;

  • 实时模型路由;

  • 统一账单管理;

  • 供应商收入结算;

  • 风险控制;

  • 企业级稳定性保障。

Unexhub 更大的目标,是从基础模型聚合,进一步走向市场化的 Token 基础设施。

平台聚合不同规模的 AI 供应商,包括与 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型相关的服务,也包括中小企业与独立供应商提供的优质模型和算力资源。

开发者可以通过统一接口使用这些资源,并根据实际消耗的 Token 付费。

在部分模型和调用方案中,价格最低可以达到相应官方标准价或市场价格的约 30%。

这一权益在高调用量业务中更为明显。

当应用消耗数百万甚至数十亿 Token 时,即使单价只有小幅差异,最终也会形成显著的运营成本差距。

更低的综合价格,可以支持:

  • 更多轮开发与测试;

  • 更长的上下文处理;

  • 更大规模的智能客服;

  • 批量任务处理;

  • AI 搜索;

  • 代码生成;

  • 企业知识库;

  • 多智能体自动化流程。

平台还可以通过毫秒级路由,根据价格、能力、可用性和响应速度,为请求匹配更合适的模型资源。

当供应商价格或容量变化时,开发者不必反复改造应用。

路由层可以在统一接口之后完成调整。


八、黄金转折点不是“成本最低”,而是“重新定价”

当 AI 单次响应成本下降到足够低时,Token 市场便进入一个关键转折点。

关键词不是单纯的“便宜”。

而是 重估。

当推理成本较高时,企业的重点是减少调用。

当推理成本足够低时,企业开始重新评估:哪些业务可以完全围绕 AI 重新设计?

搜索只是其中一个例子。

同样的变化还在发生于:

  • 软件开发;

  • 客户服务;

  • 广告营销;

  • 教育;

  • 医疗;

  • 金融分析;

  • 法律服务;

  • 企业运营。

一个执行成本为 10 美元的任务,可能不具备商业价值。

当成本降低到 0.1 美元时,它可能立刻成为值得规模化部署的应用。

当数百万家公司同时完成这笔计算时,需求不会线性增长,而会快速加速。

这会形成一种“Token 通胀”。

这里的通胀并不是 Token 单价上涨,而是 Token 总量持续膨胀:

更多应用消耗更多 Token;
更多用户启动更多工作流;
更多 Agent 运行更长时间;
更多业务流程开始持续接受 AI 辅助。

整个供应链需要围绕以下因素重新定价:

  • 算力容量;

  • 模型效率;

  • 缓存命中;

  • 请求路由;

  • 响应延迟;

  • 可用性保障;

  • 结算速度。

搜索商业模式也因此从单纯售卖点击,转向管理 Token 流动所产生的经济价值。

这并不意味着广告会消失。

而是广告模式必须承担额外的成本结构。

平台从用户的注意力、意图、交易或订阅中获得的收入,必须足以覆盖持续生成答案的推理成本。


九、Unexhub 在 AI 基础设施重估中的位置

Unexhub 正在尝试建设这一阶段所需要的交易基础设施。

对开发者而言,平台可以提供:

  • 多个主流模型的统一接入;

  • 按实际 Token 使用量付费;

  • 符合条件场景下更具竞争力的价格;

  • 毫秒级模型路由;

  • 更少的账户与接口维护;

  • 模型和价格变化时更灵活的切换能力。

对供应商而言,平台能够提供:

  • 更简单的市场需求接入渠道;

  • 实时调用分成结算;

  • 早期零佣金激励;

  • 动态定价机会;

  • 更广泛的模型分发渠道。

对企业客户而言,平台还可以围绕以下需求提供定制服务:

  • 模型选择与智能路由;

  • API 接口适配;

  • Agent 工作流集成;

  • 并发调用管理;

  • 稳定性保障;

  • 统一账单;

  • 成本可视化;

  • 持续技术支持。

目前,相关服务流程已覆盖国内外大型公司、上市公司和国有企业等不同类型组织的大模型接入与定制场景。

因此,平台所提供的价值并不只是“Token 更便宜”。

低价或许可以吸引用户完成第一次调用。

但只有稳定的基础设施,才能支持用户长期扩大业务规模。

更大的目标,是将分散在全球各处的模型和算力资源,变成更容易发现、比较、路由、消费和结算的数字资源。


十、我的判断:当响应接近免费,真正的游戏才刚刚开始

Alphabet 的财报证明,AI 可以提升搜索使用量,并继续推动收入增长。

同时,它也证明了,为这种体验提供支撑的基础设施,正在变得极其昂贵。

这正是 AI 搜索最核心的矛盾:

单次响应成本不断下降,服务整个市场的总成本却不断上升。

单位成本更低;
总使用量更大;
产品能力更强;
广告点击链路更短;
资本投入更重。

因此,单次响应成本趋近于零,并不会让搜索生意变得轻松。

它迫使整个行业围绕规模、路由、定价与 Token 流动性,重新设计商业模式。

十亿用户不只是在搜索信息。

他们正在持续创造模型推理需求。

每一次查询都在消耗 Token;
每一个 Token 都在消耗算力;
每一份算力都需要被支付;
每一个答案最终都要找到变现方式。

未来十年的 AI 搜索市场,未必只由拥有最强模型或最大广告网络的公司控制。

它也可能由那些能够让 Token 供给更加透明、流动和高效的平台共同塑造。

这正是 Unexhub 希望抓住的机会:

  • 聚合全球 AI 模型与算力供给;

  • 通过统一 API 连接企业和开发者;

  • 以毫秒级路由分配调用请求;

  • 让供应商通过竞价提供价格与服务;

  • 让用户按照实际 Token 消耗付费;

  • 让符合条件的模型资源最低达到约市场价格的 30%;

  • 让供应商的调用分成能够实时结算。

当一次 AI 响应的成本真正接近于零时,搜索的经济问题不会消失。

它只会变得更加复杂。

而真正的游戏,也正是从那一刻才刚刚开始。