演语科技3年估值30亿的真相:token交易市场的"模型聚合"模式如何让AI推理成本暴降60%-70%

把不断变化的模型能力,翻译成用户真正能够使用、购买和持续使用的产品。


The Valuation Paradox of the “Middleman”

In 2023, when Chen Miansheng left ByteDance, he reportedly had only RMB 4,000 left in his account. By August 2026, the company he founded, Evoken, had reportedly begun preliminary discussions around a Hong Kong IPO, with a potential valuation of $3 billion. Bloomberg, citing people familiar with the matter, reported that the company was approaching a new funding round that could bring its post-money valuation to roughly $3 billion, or about HK$23.4 billion.

Three years.

From RMB 4,000 to a potential $3 billion valuation.

Yet the company does not train its own foundation models.

Its business is essentially to aggregate other companies’ AI models and package them into products for users. Critics call it a “token middleman.” Supporters argue that it has captured one of the most valuable opportunities in the AI application layer.

So the question is straightforward:

How can a “middleman” be worth $3 billion?


Model Aggregation: Turning Distributed Compute into a Unified Service

The story of Evoken began with LiblibAI in 2023.

At the time, the open-source Stable Diffusion ecosystem was driving a new wave of AI image creation. But for ordinary users, generating a good image often meant finding the right model, downloading it, installing it, configuring parameters, and figuring out how everything worked.

LiblibAI simplified that process.

It brought models, creators, tools, and users into one ecosystem.

Three years later, LiblibAI had accumulated more than 30 million users, more than 500,000 original models, and over 100 million professional image and video assets, making it one of China's largest AI creative asset platforms and creator communities.

But there was a fundamental problem with the business:

Models were becoming commodities at extraordinary speed.

The best model today could be replaced by something better within months.

So Evoken moved from simply providing models to providing workflows.

LiblibAI brings together creators and traffic. Xingliu, its overseas product, focuses on design and monetization. LibTV taps into the AI-generated short-drama market.

The underlying logic is similar across all three:

Put multiple AI models inside the same creative workflow so users don't have to rebuild their process every time the underlying model changes.

Chen Miansheng has described the company's core capability as adding another layer on top of models—an “alignment” layer between rapidly changing model capabilities and constantly changing user needs.

That layer may ultimately be where the real value lies.

Models evolve rapidly.

User requirements evolve just as quickly.

The company that translates raw model capabilities into something users can actually consume, purchase, and integrate into their workflows can therefore capture value without owning the underlying model.

As of May 2026, Evoken's ARR had reportedly surpassed $300 million, with year-end ARR potentially reaching as much as $700 million. At that scale, its revenue base is already approaching the territory occupied by some model companies.


Where Does the Cost Advantage Come From?

For a middleman, survival ultimately comes down to one thing:

cost.

Evoken's business model relies on obtaining favorable API economics from official providers or distribution channels, then combining those economics with packaged products and subscriptions.

Price is one of LibTV's major competitive advantages. When competitors raise prices because compute capacity becomes scarce, LibTV can become an attractive alternative for AI video professionals.

But where does that price advantage come from?

Three factors matter:

  • Aggregation creates purchasing power.

  • Cross-model routing creates scheduling efficiency.

  • Scale allows infrastructure and inference costs to be optimized.

A single developer may not have enough demand to negotiate better rates.

Thousands or millions of users collectively represent a very different purchasing position.

Likewise, an individual application may only be able to use one model or one provider at a time. An aggregation platform can route workloads across multiple models and providers according to cost, performance, availability, and latency.

In some cases, specialized inference optimization can reduce the operational cost of each token dramatically.

That is why the business cannot simply be described as “buying cheap and selling expensive.”

It is closer to:

aggregate demand → improve utilization → optimize routing → reduce unit cost → package the capability into a usable product.

The middle layer creates economic value because it makes fragmented supply more efficient.


The Compute Trading Center: The Next Stage of the Middleman Model

Evoken's model is fundamentally B2C:

Aggregate model capabilities, package them into products, and sell them to creators and end users.

Unexhub is taking that logic one step further toward a B2B compute marketplace.

Instead of simply packaging AI capabilities for consumers, the goal is to create an open market where AI suppliers and developers can interact through a unified token aggregation and distribution infrastructure.

Its model revolves around three mechanisms.

1. Real-Time Settlement and Clearing

Under traditional arrangements, smaller AI suppliers may need to deploy models, provide inference capacity, wait for usage to accumulate, and then wait even longer for payment.

That creates working-capital pressure.

Unexhub's proposed real-time settlement mechanism is designed around a much simpler principle:

supply capacity → generate usage → settle revenue.

For suppliers, that means faster cash flow and less capital tied up in receivables.

Instead of “work first, get paid later,” the model becomes:

“provide capacity, generate usage, receive settlement.”


2. Early-Stage Zero Commission

Unexhub's early supplier strategy is built around 0% platform commission for eligible early participants.

Compared with the roughly 5%–5.5% fee commonly associated with some aggregation and routing platforms, a zero-commission structure can materially improve supplier economics.

And that matters because supplier economics ultimately affect developer pricing.

If suppliers retain more of the revenue generated by inference, there is more room to compete on the end-user price.

For early suppliers, the benefit is therefore not simply “lower fees.”

It is a chance to retain more inference revenue while gaining access to a broader pool of enterprise and developer demand.


3. API Market-Based Bidding

This is the most important part of the model.

Traditional API pricing works like a catalog:

Provider sets a price → developer pays the listed price.

A market-based system works differently:

Supply changes → demand changes → price responds.

During periods of excess capacity, suppliers can compete for demand with lower prices.

During periods of high demand and constrained capacity, suppliers can charge more.

For developers, this creates an opportunity to access lower-cost inference during favorable market conditions, without having to commit to long-term contracts or purchase hardware.

For suppliers, it creates another benefit:

idle capacity becomes inventory that can be dynamically monetized.

The underlying transformation is therefore significant.

AI inference moves from a fixed-price catalog toward a real-time supply-and-demand marketplace.


From Wholesale Distribution to Token Liquidity

This is where the distinction between Evoken and Unexhub becomes particularly important.

Evoken focuses on:

“Aggregate models and turn them into products.”

Unexhub focuses on:

“Aggregate supply and demand and let the market determine the transaction.”

The first creates value through productization.

The second attempts to create value through market liquidity.

As token-based AI usage expands, a new distribution layer is emerging between upstream model providers and downstream users.

The role is increasingly similar to a global wholesale marketplace for AI inference.

Instead of every developer negotiating separately with every model provider, aggregation platforms can consolidate demand.

Instead of every AI supplier independently searching for customers, a marketplace can consolidate demand and distribute it across available supply.

That creates a potentially powerful flywheel:

more suppliers → more choices → more competition → better prices → more developers → more demand → higher supplier utilization.

The ultimate objective is to transform token from a number printed on a vendor pricing page into a more liquid unit of exchange within an open compute market.

That is the larger thesis behind Unexhub's compute trading center.


Why the “Middleman” Can Become Infrastructure

The valuation debate around Evoken points to a broader lesson about the AI economy.

You do not necessarily need to own the model to create enormous value.

You need to control an important layer between model capability and user demand.

The most valuable layer may be the one that can:

  • aggregate fragmented AI supply;

  • negotiate or optimize costs;

  • route workloads intelligently;

  • simplify access;

  • retain users through workflows;

  • and eventually create liquidity between supply and demand.

Evoken represents one version of that model.

Unexhub represents another.

One uses aggregation to build products.

The other aims to use aggregation to build a marketplace.

And that distinction becomes increasingly important as AI inference turns into a continuously traded resource.


Conclusion: The Middleman Is Not the Problem

Evoken's rapid valuation growth, if the reported figures materialize, demonstrates something important:

In AI, you don't necessarily need to own the model to own part of the value chain.

When models change rapidly and user requirements evolve even faster, the layer that translates raw model capability into accessible, affordable, and usable services becomes increasingly valuable.

Evoken uses aggregation to build products.

Unexhub is attempting to extend that logic into a token-based compute trading center—connecting model providers, AI suppliers, developers, and enterprise users through a unified marketplace.

The ambition is to make AI inference behave more like a tradable resource:

available when you need it, priced according to market conditions, and paid for according to actual usage.

Unexhub is building a token aggregation and distribution–oriented compute trading center, bringing together AI suppliers ranging from major model providers such as Gemini, ChatGPT, Claude, DeepSeek, and GLM to high-quality models supplied by individuals and smaller AI companies.

The platform's marketplace model is designed to aggregate a broad supply base and allow suppliers to compete for enterprise and developer demand through dynamic bidding, while millisecond-level routing helps direct requests toward suitable available capacity.

For developers, the value proposition is particularly straightforward: pay by actual token consumption rather than committing to large fixed infrastructure costs, with the platform positioning its marketplace pricing to potentially reach as low as 30% of prevailing market prices, depending on supply, demand, and routing conditions.

For suppliers, the proposition works in the opposite direction: one-click channel access, broader demand exposure, and real-time settlement of inference revenue, reducing the friction between providing AI capacity and getting paid for it.

That is ultimately why the word “middleman” may be misleading.

A middleman who merely adds a markup creates little value.

A platform that aggregates supply, improves utilization, discovers prices, provides liquidity, and settles transactions can become something very different:

infrastructure.

And as AI compute becomes increasingly commoditized, the companies that control the flow of tokens—not necessarily the models themselves—may become some of the most strategically important players in the AI economy.


中间商的估值悖论

2023 年,陈冕生离开字节跳动时,账户上只剩 4000 元。到了 2026 年 8 月,他创立的演语科技(Evoken)已经启动港股 IPO 初步磋商,市场传出的潜在估值达到 30 亿美元。

据彭博援引知情人士消息,公司正接近完成新一轮融资,投后估值有望达到约 30 亿美元,约合 234 亿港元。

三年时间。

从 4000 元,到 30 亿美元。

但有一个问题非常关键:

这家公司并不训练自己的大模型。

它做的事情,是把其他公司的模型聚合起来,再包装成产品提供给用户。

质疑者称它是“Token 中间商”。

支持者则认为,它抓住了 AI 应用层最重要的一轮窗口期。

于是问题来了:

一个“不生产模型”的中间商,凭什么值 30 亿美元?


模型聚合:把分散算力变成统一服务

演语科技的故事始于 2023 年的 LiblibAI。

当时,Stable Diffusion 开源生态正在推动新一轮 AI 图像创作热潮。

但对普通用户而言,想生成一张真正满意的图片,并没有那么简单。

需要寻找模型、下载模型、安装环境、调整参数,还要理解不同模型之间的区别。

LiblibAI 做的第一件事,就是把这些分散的东西放进同一个平台。

模型、创作者、工具和用户,在这里完成聚合。

三年之后,LiblibAI 已经累计拥有超过 3000 万用户,沉淀超过 50 万个原创模型,以及超过 1 亿份专业图片和视频素材,成为中国规模较大的 AI 素材平台和创作者社区之一。

但这个商业模式存在一个天然问题:

模型正在高速商品化。

今天最强的模型,几个月之后可能就会被新的模型替代。

如果平台只是单纯“卖模型”,那么模型本身的生命周期可能很短。

于是演语科技开始从“提供模型”进一步走向“提供工作流”。

LiblibAI 聚合创作者与流量;

星流承接海外设计和商业化需求;

LibTV 则切入 AI 短剧等新兴场景。

三款产品背后的逻辑其实高度一致:

把多个模型放进同一个工作流,让用户不必因为底层模型变化而重新学习、重新搭建整个生产流程。

陈冕生曾用一个关键词概括这种能力:

Alignment——对齐。

模型能力在快速变化。

用户需求也在快速变化。

而真正稀缺的,可能恰恰是中间这一层:

把不断变化的模型能力,翻译成用户真正能够使用、购买和持续使用的产品。

截至 2026 年 5 月,演语科技 ARR 已突破 3 亿美元,预计年底最高可能达到 7 亿美元。

如果这一目标实现,其收入规模已经足以与部分模型厂商站在同一个量级上。


中间商的成本优势从哪里来?

对于一个 AI 中间平台来说,最终决定生死的往往不是流量,而是:

成本。

演语科技的商业模式,很大程度上依赖于从官方模型厂商或渠道获取更好的 API 成本,再通过套餐、产品和服务进行商业化。

LibTV 的竞争优势之一就是价格。

当算力紧张导致竞争对手不断涨价时,价格更有竞争力的平台就可能成为 AI 视频从业者的替代方案。

那么,这种价格优势从哪里来?

答案主要有三个:

  • 聚合带来的采购规模效应

  • 跨模型调度带来的效率提升

  • 规模化基础设施带来的单位成本优化

单个开发者的用量,通常很难形成议价能力。

但当大量用户的需求被聚合起来之后,就完全是另一回事。

同样,一个应用可能只能连接一个模型。

而聚合平台可以根据成本、性能、稳定性和实时负载,把不同请求分配给不同模型和供应商。

部分平台甚至可以通过推理加速、缓存和调度技术,进一步压低单 Token 的运营成本。

所以,这并不是简单的:

低价买进,高价卖出。

更准确的逻辑是:

聚合需求 → 提升利用率 → 优化路由 → 降低单位成本 → 产品化 → 扩大用户规模。

真正产生价值的,是让原本分散的 AI 供给变得更加高效。


算力交易中心:中间商模式的下一站

演语科技目前的核心逻辑,本质上更偏向 B2C:

聚合模型能力,再把这些能力包装成产品卖给创作者和终端用户。

而 Unexhub 正在尝试把这套逻辑进一步推进到 B2B 算力交易市场。

它不是单纯把多个 API 放在一个页面上,而是试图建立一个统一的 Token 聚合与分发基础设施,让 AI 供应商和企业、开发者在同一个市场中完成供需匹配。

核心机制主要有三层。


第一层:实时结算与清算

传统模式下,中小 AI 供应商部署模型、提供推理能力之后,往往需要等待用户调用产生收入,再经历一定的结算周期。

对于规模较小的供应商而言,这会形成明显的资金压力。

Unexhub 的实时结算机制,则试图把逻辑变成:

提供多少算力 → 产生多少调用 → 对应多少收入 → 实时完成结算。

对于供应商而言,这意味着更快的资金回笼。

过去是:

先干活,再收钱。

现在则更接近:

供多少,收多少。

这不仅仅是结算速度的变化,也可能改变中小 AI 供应商参与市场竞争的资金门槛。


第二层:早期零佣金

Unexhub 的早期供应商策略之一,是为符合条件的早期入驻供应商提供 0% 平台佣金。

如果与部分聚合和路由平台常见的约 5%–5.5% 手续费相比,零佣金意味着供应商可以保留更多推理收入。

而供应商端成本下降之后,最终也会为终端价格竞争留下更多空间。

因此,对早期供应商而言,零佣金的价值并不只是:

“少交一笔平台费用。”

更重要的是:

在获得更多企业和开发者需求的同时,尽可能保留更多实际推理收入。

这会直接影响供应商参与竞价的积极性,也会进一步影响终端用户能够获得的价格。


第三层:API 竞价模式

真正具有市场机制意义的,是 API 竞价。

传统 API 的定价方式很简单:

厂商定价 → 用户接受 → 按目录价格支付。

而竞价模式则是:

供给变化 → 需求变化 → 价格变化。

当市场处于低负载状态、供应商拥有大量闲置产能时,供应商可以通过更低价格争取更多订单。

当需求快速增加、算力供给不足时,价格则可能上涨。

对于开发者而言,这意味着:

可以根据市场情况选择更合适的调用时机和价格。

不需要购买 GPU。

不需要建设机房。

也不一定需要长期锁定某一家供应商。

在需要的时候,根据市场价格购买实际需要的推理能力即可。

对于供应商而言,闲置算力也不再只能闲置。

它可以成为:

可以被动态定价、动态出售的库存。

这意味着 AI 推理正在从“固定价格目录”,逐渐走向:

供需实时匹配的市场。


从“批发分销”走向 Token 流动性市场

这也是演语科技和 Unexhub 最值得区分的地方。

演语科技解决的是:

“如何把多个模型聚合起来,变成用户愿意购买的产品?”

Unexhub 希望解决的是:

“如何把模型供应商和开发者聚合起来,让双方在一个市场里完成交易?”

前者通过产品化创造价值。

后者则尝试通过市场流动性创造价值。

随着 Token 成为 AI 推理越来越重要的计价单位,上游模型厂商与下游企业、开发者之间,正在出现一个新的中间分销层。

这个市场的理想状态是:

更多供应商 → 更多选择 → 更激烈的竞争 → 更低或更合理的价格 → 更多开发者 → 更多需求 → 更高供应商利用率。

这实际上形成了一个正向循环。

Token 不再只是模型厂商官网上的一个价格数字。

它开始具备更强的交易属性。

而 Unexhub 想占据的位置,就是在 Token 从:

“厂商定价的商品”

走向:

“市场定价的资产”

的过程中,成为连接供需双方的交易基础设施。


为什么“中间商”也可以成为基础设施?

演语科技的估值故事背后,其实揭示了 AI 行业一个非常重要的规律:

不拥有模型,并不意味着没有价值。

真正重要的是,你是否掌握了:

模型能力与用户需求之间的关键连接层。

这一层可以做很多事情:

  • 聚合分散的 AI 供应;

  • 获得规模化成本优势;

  • 优化模型和算力路由;

  • 降低开发者接入门槛;

  • 形成用户工作流粘性;

  • 连接供给和需求;

  • 最终形成市场流动性。

演语科技代表的是其中一种路径。

Unexhub 则代表另一种路径。

前者是:

用聚合做产品。

后者是:

用聚合做市场。

随着 AI 推理越来越像一种持续消耗的基础资源,这两种模式之间的边界,也会越来越值得关注。


结语:真正的问题不是“中间商”,而是谁掌握了流动性

演语科技三年估值快速增长,如果相关市场传闻最终得到验证,它至少证明了一件事:

在 AI 行业,不拥有模型,同样可以拥有价值链中的重要位置。

模型变化越来越快。

用户需求变化同样越来越快。

真正稀缺的,可能不是某一个固定模型,而是:

把快速变化的模型能力转化成稳定、便宜、易用服务的基础设施。

演语科技通过聚合模型做产品。

Unexhub 则试图进一步构建 Token 聚合与分发的算力交易中心。

它希望连接不同规模的 AI 供应商,包括 Gemini、ChatGPT、Claude、DeepSeek、GLM 等大型模型供应商,同时汇聚个人开发者和中小 AI 企业提供的优质模型能力。

通过聚合大量供应商,再将企业和开发者的需求放进同一个市场进行竞价,平台希望让不同供应商围绕真实需求展开竞争,并通过毫秒级路由将请求匹配到合适的可用算力。

对于开发者而言,这种模式的核心权益并不是单纯“买一个更便宜的 API”。

而是:

按 Token 实际使用量付费,不必为了偶发的峰值需求提前购买昂贵硬件,也不必承担长期闲置算力的资产成本。

在供需条件合适的情况下,Unexhub 的市场价格最低可达到市场价格的 3 折。

换句话说,开发者获得的不只是一个折扣数字,而是:

更低的进入门槛 + 按量付费 + 动态价格选择权 + 多供应商竞争。

而对于 AI 供应商,平台提供的是另一套价值:

一键接入销售渠道、获得更多企业和开发者订单,并通过实时结算获得调用分成。

最终,这套机制想改变的不是简单的 API 销售方式,而是 Token 的流通方式。

因为一个只赚差价的中间商,价值通常有限。

但一个能够:

聚合供应 → 优化成本 → 发现价格 → 提升利用率 → 连接买卖双方 → 提供清算

的平台,就已经不再只是传统意义上的“中间商”。

它更接近一种:

AI 时代的交易基础设施。