算力交易中心上线 AI Agent App Store 企业入驻通道:字节豆包 1200 个 Agent 打样,99 元/月能买到的不只是工具

For years, the industry repeated the same slogan: “Compute is the new oil.” GPU clusters, massive-scale interconnects, and compute rental services dominated the infrastructure narrative.


When “Selling Compute” Becomes “Selling Capability”: The Paradigm Shift Behind AI Agent App Stores

The underlying business logic of the AI industry is undergoing a quiet but fundamental rewrite.

For years, the industry repeated the same slogan: “Compute is the new oil.” GPU clusters, massive-scale interconnects, and compute rental services dominated the infrastructure narrative.

But by the summer of 2026, the direction is changing.

From the “Bay Area Smart City” initiatives in Guangzhou to Token aggregation marketplaces in Nanjing, from Token marketplaces in Wuxi to Token trading centers in Guangdong, a new consensus is emerging:

Compute should not simply be sold as infrastructure. It should be transformed into usable AI capability.

The shift sounds subtle, but it could fundamentally change how AI services are bought, sold, and monetized.


From “Selling Hardware” to “Selling AI Services”

What does traditional compute trading look like?

You purchase GPUs, rent servers, configure an environment, deploy a model—and then figure out what to do with it.

It is not fundamentally different from buying raw materials.

But most businesses do not actually want GPUs.

They want the results those GPUs can produce.

The “Bay Area Smart City” approach is straightforward: combine trusted compute services, trusted AI-agent services, and trusted data spaces to provide companies with customized AI solutions through a unified platform.

More importantly, companies can list industry models and AI agents developed on the platform for other businesses to use and purchase.

Nanjing’s Token marketplace pushes the concept further by using Tokens as a standardized measurement layer connecting compute supply, model development, industrial applications, and supporting financial services.

Wuxi’s Token marketplace similarly attempts to abstract away the underlying hardware. Companies do not need to understand which GPU is running behind the scenes. They simply pay according to the AI capability they actually consume—much like paying for bandwidth or mobile data.

That is why the transition from “selling compute” to “selling capability” is more than a change in terminology.

It represents a restructuring of the business model itself.


AI Agent App Stores: Turning Agents Into Tradable Products

Once compute becomes measurable through Tokens, AI agents can become tradable digital products.

On platforms such as the “Bay Area Smart City,” enterprises can list AI agents they have developed and monetize them while other organizations consume their capabilities.

Nanjing’s Token marketplace provides standardized Token-based compute trading, scheduling, and application services for enterprises, universities, and institutions.

China Unicom Guangdong’s “Yuanjing Tianrong Agent Supermarket” goes one step further by adopting an “Agent + Token” operating model, using Tokens as a common value carrier connecting AI-agent supply and demand.

The commercial data from AI·Modfang’s Agent distribution center is even more revealing.

The platform has reportedly connected 135 companies and 287 AI-agent products, facilitating 302 transactions with cumulative transaction value approaching RMB 30 million.

That is no longer merely a conceptual experiment.

It is evidence that the commercial loop is beginning to work:

Develop → List → Match → Use → Pay → Monetize.


ByteDance’s 1,200 Agents: A Prototype for Scale

While much of the industry is still debating how AI agents should be commercialized, ByteDance has already put a large-scale prototype on the table.

At the Volcano Engine FORCE conference in March 2026, ByteDance presented a three-layer Agent ecosystem consisting of:

  • Consumer-facing Doubao

  • Enterprise-oriented Coze

  • Vertical industry Agents

HiAgent was introduced as an important component of the enterprise Agent platform ecosystem.

Meanwhile, Doubao Enterprise began customer testing with native integration into Feishu documents, spreadsheets, meetings, group chats, and other workplace scenarios.

The organizational restructuring that followed was even more revealing.

On July 30, 2026, ByteDance merged the Feishu and Doubao product teams, while also combining the Feishu GTM team with the Volcano Engine team.

More than 90% of Feishu’s new customers reportedly also purchased Feishu AI products.

The message is clear:

AI agents cannot rely solely on consumer chatbots. Their real commercial value comes from becoming part of enterprise workflows.

An AI agent becomes far more valuable when it can access documents, knowledge bases, meetings, communication channels, and business systems—and actually complete tasks.


RMB 99 a Month: A Price Point or a Signal of the New Era?

The emergence of a RMB 99/month price point is particularly interesting.

In June 2026, Doubao Professional launched with standard recurring subscriptions starting from RMB 68 per month. Doubao Coding plans were also promoted at prices as low as RMB 9.9 for the first month.

China Unicom’s Nanjing OPC service platform introduced Token packages starting from around RMB 15 per month.

RMB 99 sits in an interesting middle ground.

It is slightly more expensive than a typical personal AI subscription, but dramatically cheaper than traditional enterprise software procurement.

More importantly, the product being sold is changing.

The customer is no longer simply purchasing access to a piece of software.

They are effectively “hiring a digital employee.”

That digital employee can work with documents, knowledge bases, meetings, chats, and other enterprise context to complete tasks with increasing levels of autonomy.

This points toward a broader shift:

AI-agent pricing may gradually move from seat-based subscriptions toward usage-based economics.

Instead of paying for the right to use software, companies increasingly pay for the amount of work an AI agent actually performs.


Compute Marketplaces + AI Agent Stores: The Paradigm Shift Is Already Underway

Put all these developments together, and the bigger picture becomes much clearer.

Infrastructure Layer: Compute Trading Centers

Compute is standardized into Tokens.

Tokens become measurable, transferable, and tradable units of AI consumption.

Application Layer: AI Agent Marketplaces

Agents become products that can be developed, listed, purchased, deployed, and reused—much like apps.

Scale Layer: Thousands of Agents

ByteDance’s 1,200-agent ecosystem demonstrates what happens when the number of AI agents moves from dozens to hundreds and eventually thousands.

When these three layers converge, the rules of the AI economy begin to change:

From selling models → to selling capabilities

From buying compute → to buying outcomes

From software subscriptions → to hiring AI agents

Two years ago, building an AI application typically meant purchasing compute, selecting models, configuring infrastructure, and building a toolchain before the actual product development could even begin.

The barrier to entry was enormous.

Two years from now, the workflow could look very different:

Choose a model → allocate Tokens → deploy an Agent → connect enterprise context → publish the service.

The infrastructure becomes invisible.

The capability becomes the product.

And the outcome becomes the unit of value.


Unexhub: Turning Token Consumption Into a More Flexible AI Marketplace

This is also where platforms such as Unexhub fit into the emerging ecosystem.

Unexhub is positioning itself as a global Token aggregation and distribution platform, connecting AI model providers with developers and enterprise users through a compute trading marketplace.

The platform brings together AI providers across different scales, including major model ecosystems such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, alongside high-quality models developed by individual developers and smaller AI companies.

Instead of locking users into a single model provider, the aggregation approach gives developers and businesses more flexibility to compare capabilities and costs across multiple models.

Its value proposition is particularly relevant in a usage-based AI economy:

  • Multi-model access: connect to multiple mainstream domestic and international AI models through one platform.

  • Competitive pricing: providers compete for enterprise and developer demand, creating room for lower-cost AI consumption.

  • Millisecond-level routing: intelligently route requests to appropriate model resources based on requirements.

  • Pay by actual Token usage: developers can pay according to real consumption rather than committing to unnecessary infrastructure.

  • Potential savings: eligible users can access Token prices as low as 30% of prevailing market prices, depending on model and provider conditions.

  • Provider monetization: AI suppliers can connect to distribution channels and receive usage-based revenue sharing with real-time settlement.

For high-volume AI applications, the economic value is not simply about finding a cheaper API.

It is about gaining more choices, more flexible routing, and better control over the total cost of AI consumption.


The AI economy is gradually moving from “How much compute do you own?” toward a more practical question:

“How much useful work can your AI capability deliver—and what does each unit of that work cost?”

That may be the real meaning behind the transition from selling compute to selling capability.

And RMB 99/month may only be the beginning.


中文版|中文读者请继续向下阅读

当“卖算力”变成“卖能力”:AI Agent 应用商店的范式迁移

AI 产业的底层商业逻辑,正在经历一场安静却可能十分彻底的重构。

过去几年,行业反复强调一句话:

“算力是新时代的石油。”

GPU 集群、万卡互联、算力租赁,一轮又一轮地成为 AI 基础设施叙事的核心。

但进入 2026 年夏天之后,风向开始发生变化。

从广州的“湾区智城”,到南京的 Token 集合店;从无锡的 Token 超市,到广东探索中的词元交易中心,一个新的共识正在逐渐形成:

算力不应该只是作为基础设施被“卖掉”,而应该被转化成可以直接使用的 AI 能力。

这看似只是概念上的变化,实际上可能正在重新定义 AI 服务未来如何被购买、交易和变现。


从“卖硬件”到“卖 AI 服务”

传统的算力交易是什么?

购买 GPU、租赁服务器、搭建运行环境、部署模型,然后再想办法利用这些资源。

它本质上和购买原材料非常接近。

但对于绝大多数企业来说,真正需要的其实不是 GPU。

企业真正需要的是:

GPU 最终能够产生什么结果。

“湾区智城”的思路非常直接:将可信算力服务、可信智能体服务与可信数据空间结合起来,为企业提供一站式 AI 定制方案。

更重要的是,企业还可以将基于平台开发的行业模型和 AI Agent 上架,让其他企业使用甚至购买。

南京的 Token 集合店则进一步尝试以 Token 作为统一计量载体,把算力供给、模型开发、产业应用以及配套服务连接起来。

无锡的 Token 超市也在尝试将底层硬件进一步“隐藏”。

企业不需要关心背后到底使用了什么型号的 GPU,只需要根据实际消耗的 AI 能力付费。

它更像购买流量:

用了多少,就支付多少。

所以,从“卖算力”走向“卖能力”,绝不仅仅是换了一个说法。

它实际上意味着商业模式本身正在发生变化。


AI Agent 应用商店:让智能体成为可交易商品

当算力可以通过 Token 被标准化计量之后,AI Agent 就开始具备成为数字商品的条件。

在“湾区智城”等平台上,企业可以将自己开发的 AI Agent 上架,让其他企业使用,同时获得相应收益。

南京的 Token 集合店则为企业、高校和机构提供标准化的 Token 算力交易、调度与综合应用服务。

中国联通(广东)推出的“元景·天融智能体超市”进一步采用 “Agent + Token” 的新模式,让 Token 成为连接智能体供需双方的统一价值载体。

AI·Modfang 智能体分发中心的数据更加直观:

目前已经连接 135 家企业、287 个智能体产品,完成 302 笔撮合交易,累计交易金额接近 3000 万元。

这已经不再只是概念验证。

一个完整的商业闭环正在形成:

开发 → 上架 → 撮合 → 使用 → 付费 → 变现。


字节跳动的 1200 个 Agent:规模化的原型

当整个行业还在讨论“AI Agent 应该如何商业化”时,字节跳动已经把一个大规模原型摆到了桌面上。

在 2026 年 3 月的火山引擎 FORCE 原动力大会上,字节跳动构建了由:

  • 消费级豆包

  • 企业级扣子

  • 垂直行业 Agent

组成的三层 Agent 生态矩阵。

HiAgent 也作为企业级 Agent 平台的重要组成部分亮相。

与此同时,豆包企业版开始进行客户内测,并原生接入飞书文档、表格、会议、群聊等企业办公场景。

随后发生的组织调整更值得关注。

2026 年 7 月 30 日,字节跳动将飞书产品团队与豆包产品团队合并,同时将飞书 GTM 团队与火山引擎团队合并。

据相关信息,超过 90% 的飞书新客户同时购买了飞书 AI 产品。

这一系列动作释放出的信号非常清晰:

AI Agent 的商业化不能只依靠 C 端聊天,它必须进入企业真实工作流。

当一个 AI Agent 能够连接企业文档、知识库、会议、群聊以及业务系统,并且真正替企业完成工作时,它的商业价值才会真正释放出来。


99 元/月:价格锚点,还是时代信号?

“99 元/月”这个数字出现得很有意思。

2026 年 6 月,豆包专业版正式上线,标准连续包月套餐从 68 元起步;豆包编程套餐首月价格甚至低至 9.9 元。

南京中国联通 OPC 服务平台也推出了 15 元/月起的 Token 套餐。

而 99 元/月恰好处于一个非常微妙的位置。

它比普通个人 AI 订阅稍贵,却远低于传统企业软件采购的成本。

更重要的是,企业购买的东西正在发生变化。

企业购买的不再只是一个软件账号。

它实际上是在:

“雇佣一个数字员工。”

这个数字员工能够读取企业文档、知识库、会议和沟通记录,并根据上下文自主完成越来越复杂的工作。

因此,AI Agent 的商业模式可能正在从传统的“席位订阅”,逐渐转向:

按实际使用量付费。

企业购买的不再只是软件的使用权,而是 AI Agent 实际完成了多少工作。


算力交易中心 + AI Agent 应用商店:范式迁移正在发生

把这些变化放在一起,整个产业结构就变得清晰起来。

基础设施层:算力交易中心

算力被标准化为 Token。

Token 进一步成为可计量、可交易、可流通的 AI 消耗单位。

应用层:AI Agent 应用商店

智能体变成类似 App 的数字产品,可以被开发、上架、购买、部署和重复使用。

规模层:数千个 Agent

字节跳动的 1200 个 Agent 说明,当智能体数量从几十个增长到几百个、几千个之后,一个新的应用生态就可能开始自发形成。

当这三层真正连接起来,AI 产业的游戏规则也会发生变化:

从卖模型 → 卖能力

从买算力 → 买结果

从软件订阅 → 雇佣 AI Agent

两年前,如果想做一个 AI 应用,通常需要先完成算力采购、模型选择、环境搭建和工具链配置。

还没真正开始创业,基础设施门槛可能已经让很多团队望而却步。

而未来的工作方式可能完全不同:

选择模型 → 分配 Token → 部署 Agent → 接入企业数据 → 发布服务。

基础设施逐渐变得透明。

真正被交易的,是 AI 能力。

最终被计价的,则是 AI 能够产生多少有价值的结果。


Unexhub:让 Token 消耗进入更灵活的 AI 交易市场

这也是 Unexhub 所试图切入的方向。

Unexhub 正在构建全球化的 Token 聚合与分发平台,通过算力交易中心的模式,把 AI 模型供应商、开发者和企业用户连接起来。

平台汇聚不同规模的 AI 供应商,包括 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型生态,同时也吸纳个人开发者和中小型 AI 企业提供的优质模型。

相比绑定单一模型厂商,聚合模式可以让企业和开发者在不同模型之间进行能力与成本比较。

对于正在进入“按使用量计费”的 AI 时代来说,这种模式的价值尤其明显:

  • 多模型接入: 一个平台即可连接国内外多个主流 AI 模型。

  • 价格竞争: 供应商围绕开发者和企业用户进行竞争,为更低成本的 AI 调用创造空间。

  • 毫秒级路由: 根据任务需求和模型资源进行智能路由。

  • Token 实付: 按真实 Token 使用量结算,减少不必要的资源投入。

  • 成本优势: 符合条件的用户,根据具体模型和供应商情况,最低可享受约 市场价格 3 折 的 Token 价格。

  • 供应商变现: AI 模型供应商可以一键接入分发渠道,并按照实际调用获得收益,支持实时结算。

对于高频调用 AI 的企业和开发者来说,真正有价值的并不只是“API 更便宜”。

更重要的是:

模型选择更多、路由更灵活、Token 成本更透明,整体 AI 使用成本也更容易控制。


AI 经济正在逐渐从:

“你拥有多少算力?”

转向一个更加现实的问题:

“你的 AI 能力能够完成多少有价值的工作,而每一单位工作的成本是多少?”

这或许才是“从卖算力到卖能力”真正的意义。

而 99 元/月,可能仅仅是这个新时代的开始。