推理吃掉 AI 计算 80-90%!Token 交易市场如何用"按效果付费"重构算力账单?

AI 真正開始賺錢之後,最大的算力賬單可能並不是“把模型訓練出來”,而是——讓數以億計的用戶不斷調用它。


推理吃掉 AI 算力 80%-90%:Token 交易市场如何重算算力账单?

当训练不再是 AI 最大的算力消耗,推理正在成为新的“成本黑洞”。而 Token,也正在从一个计量单位,变成一项可以被重新定价、重新分配甚至重新交易的数字资源。

2026 年,AI 算力的叙事正在彻底改写。

三年前,衡量一家 AI 公司实力的标准,很大程度上还是“拥有多少 H100”。训练大模型是算力投入的核心,谁拥有更多 GPU,谁就拥有更大的模型竞争优势。

但到了 2026 年,这个逻辑正在反转。

联想集团董事长兼 CEO 杨元庆在 Q1 业绩交流会上表示,此前市场上的 AI 算力投入大约 80% 用于训练、20% 用于推理;如今这一比例已经逐渐变为 50%:50%。更值得关注的是,他预计未来更合理的算力投入结构可能变成 80% 用于 AI 推理、20% 用于训练。

训练与推理,正在完成一次算力角色互换。

TrendForce 的数据也印证了这一趋势:2026 年北美五大 CSP 的 AI 训练算力预计增长 56% 以上,而 AI 推理算力预计增长接近 122%,增速超过训练算力的两倍。

换句话说,AI 真正开始赚钱之后,最大的算力账单可能并不是“把模型训练出来”,而是——让数以亿计的用户不断调用它。


一、推理为什么正在成为 AI 最大的成本黑洞?

答案并不是单次调用有多贵,而是——推理从来不会停止。

训练更像一次性资本投入。

模型训练完成之后,大部分训练成本已经发生。

推理却完全不同。

每一次用户提问、每一次 Agent 调用、每一次代码生成、每一次工具执行、每一个 API 请求,都会持续产生新的计算成本。

因此,当 AI 应用真正进入规模化阶段,推理成本会随着用户量和调用频率持续增长。

SambaNova 等 AI 芯片厂商曾估算,在模型进入大规模生产阶段之后,推理可能占据模型全生命周期算力成本的 80%-90%。

Gartner 的预测同样值得关注:2026 年全球 AI 推理支出预计达到 233 亿美元,首次超过训练支出的 190 亿美元,占 AI 优化型 IaaS 支出的 55%;到 2029 年,这一比例还可能进一步超过 65%。

真正麻烦的地方还在于一个经典的经济学现象——杰文斯悖论(Jevons Paradox)。

当 Token 越来越便宜,企业往往不会因此少用 AI,反而会更加大胆地使用。

贝恩数据显示,从 2024 年底到 2025 年底,平均 Token 成本下降约一半,但 Token 消耗量却增长了 4.5 倍。

国内 Token 日均调用量也从 2024 年初的约 1000 亿,增长到 2026 年 3 月的约 140 万亿。

于是出现了一个非常有意思的现象:

Token 单价在下降,但总算力账单却在上涨。

AI 越便宜,使用越频繁;使用越频繁,推理需求越高;推理需求越高,算力资源越紧张。

这正是未来 AI 基础设施最难解决的矛盾之一。


二、按 Token 付费,真的就是最公平的方式吗?

表面上看,Token 定价非常公平。

模型处理多少 Token,用户就支付多少费用。

简单、透明、容易计算。

但随着 AI Agent 越来越复杂,一个新的问题开始暴露:

用户真正购买的,到底是 Token,还是任务结果?

举个简单例子。

一个法律检索任务:

  • 模型 A 消耗 5000 Token,完成任务;

  • 模型 B 只消耗 2000 Token,也完成同样的任务。

如果完全按照 Token 计价,模型 A 反而需要支付更多。

也就是说:

更低效的模型可能赚得更多,更高效的模型反而被“惩罚”。

这就是当前 Token 经济中的结构性矛盾。

上游模型供应商按照计算资源收费,而下游企业真正愿意支付的,却往往是结果、效率和 ROI。

因此,Token 只是价值传递过程中的中间单位。

真正的终点,是任务产生的价值。


三、Token 交易市场:把“算力资源”变成“可交易商品”

这也是 Unexhub 正在尝试解决的问题。

Unexhub 的核心定位,并不是简单再做一个 AI API 聚合平台,而是尝试建立一个 Token 聚合、分发与交易的算力交易中心。

它希望改变传统 AI 算力市场中“供应商定价、用户被动购买”的单向关系。

① Token 聚合:让更多算力进入同一个市场

Unexhub 汇聚不同层级的 AI 模型供应商,包括 Gemini、ChatGPT、Claude、DeepSeek、GLM 等大型模型,同时也接入个人开发者和中小型企业提供的 AI 模型资源。

这样做的意义在于:

用户不必绑定某一家模型供应商,而可以在不同模型和不同价格之间进行选择。

对于企业和开发者而言,这意味着更多选择,也意味着更强的价格发现能力。


② 毫秒级路由:让请求自动寻找更合适的资源

传统 API 调用往往是:

选择一个模型 → 绑定一个 API → 按官方价格持续支付。

Unexhub 想做的则是进一步:

需求进入 → 市场匹配 → 毫秒级路由 → 调用最合适的模型资源。

当某一供应商资源紧张时,可以切换其他可用资源;当某类模型价格更具优势时,也能够根据实际需求进行调度。

这实际上是在给 AI 推理市场增加一层“交通调度系统”。


③ API 竞价:让价格真正反映供需

这可能是 Token 交易市场最值得关注的一环。

传统 API 通常采用固定价格:

高峰期一个价,低谷期还是一个价。

但算力资源并不是永远处于同样的供需状态。

高峰期,GPU 紧张;

低谷期,大量算力可能处于闲置状态。

如果价格始终固定,市场就缺少一个最重要的东西:

价格信号。

Unexhub 尝试通过 API 动态竞价,让价格随着市场供需变化。

需求旺盛时,供应商获得更高收益;

资源充足时,开发者获得更低成本。

这与传统模型厂商单独调整价格不同。

前者是:

厂商决定价格。

后者更接近:

市场决定价格。


四、从“按 Token 付费”走向“按价值付费”

这可能才是 Token 交易市场真正值得关注的长期方向。

未来 AI 定价可能逐渐出现三种模式:

第一阶段:按资源收费

你用了多少 Token,就支付多少钱。

第二阶段:按能力收费

你调用什么模型、什么能力,就支付对应价格。

第三阶段:按结果收费

AI 最终帮你完成了什么任务,创造了多少价值。

这意味着 AI 定价体系可能逐渐从:

Token → Intelligence → Outcome

发生迁移。

当 AI Agent 能够独立完成搜索、编程、分析、运营、客服甚至复杂业务流程时,仅仅计算 Token 数量已经越来越难准确衡量真实价值。

未来真正重要的问题可能变成:

“这个 AI 花了多少 Token?”

逐渐变成:

“它用这些 Token 为我创造了多少价值?”


五、Unexhub 想抢的,不只是低价

如果只是“把 AI API 做得更便宜”,这个市场很容易陷入新的价格战。

Unexhub 真正值得关注的地方,在于它试图同时解决 价格、供给、流动性和结算 四个问题。

对于开发者而言,平台通过聚合不同供应商、动态路由和竞价机制,让开发者拥有更大的选择空间。

目前 Unexhub 提供的市场权益包括:

  • 按 Token 实际使用量付费,减少固定套餐带来的浪费;

  • 多模型聚合,一次接入即可使用 Gemini、ChatGPT、Claude、DeepSeek、GLM 等不同模型资源;

  • 毫秒级智能路由,根据实际需求匹配可用模型;

  • 动态竞价机制,让供需关系更直接地反映到价格;

  • 最低可享受约市场价格 3 折的 Token 使用成本,尤其适合高频调用、批量任务和 AI Agent 场景;

  • 对供应商而言,则可以通过 一键接入渠道、实时结算调用分成,降低获客和资金周转压力。

换句话说,开发者获得的不只是一个“更便宜的 API”。

更重要的是:

同样的预算,可以获得更多 Token;同样的 Token,可以通过更多模型完成任务;同样的模型需求,也拥有更多价格选择。

对于每天消耗数百万甚至数十亿 Token 的企业而言,这种差异最终可能直接反映在基础设施成本和 AI 项目 ROI 上。


六、推理吃掉 80% 算力之后,谁来重新算这笔账?

AI 行业正在进入一个非常特殊的阶段。

模型越来越强。

Token 越来越便宜。

但 Token 使用量增长得更快。

于是,真正决定 AI 应用能否规模化的,可能已经不再是:

“谁拥有最多 GPU?”

而是:

“谁能最高效地调度这些 GPU?”

推理正在吞噬 80%-90% 的 AI 全生命周期算力成本。

这意味着,算力正在从一种单纯的基础设施,逐渐变成一种需要被定价、交易、调度和流动的数字资源。

Unexhub 正是在这个方向上进行尝试:重塑全球 Token 逻辑,建立 Token 聚合与分发的算力交易中心。

如果说过去 AI 行业竞争的是模型参数,那么接下来竞争的可能是:

谁能用更低的成本,获得更多的有效智能。

当 Token 从“计量单位”变成“交易主体”,当算力从“固定资源”变成“流动商品”,AI 推理市场才真正开始拥有自己的价格发现机制。

推理吃掉的是 80% 的算力。

而下一场竞争,可能是:

谁能把这 80% 的算力用得最值。


English Version | Bilingual Article — English Readers, Please Continue Below

When Inference Consumes 80%-90% of AI Compute: How Can a Token Trading Market Recalculate the Compute Bill?

Training is no longer the biggest consumer of AI compute. Inference is becoming the new cost black hole. At the same time, tokens are evolving from simple measurement units into digital resources that can be repriced, redistributed, and potentially traded.

In 2026, the economics of AI compute are being rewritten.

Three years ago, one of the simplest ways to measure an AI company's strength was to ask how many H100 GPUs it controlled. Training large models was at the center of the AI infrastructure story.

That logic is now reversing.

At a Q1 earnings discussion, Lenovo Chairman and CEO Yang Yuanqing said that AI compute allocation had previously been roughly 80% for training and 20% for inference. Today, the ratio is approaching 50% for training and 50% for inference. More importantly, he expects a more reasonable long-term structure to be around 80% inference and 20% training.

Training and inference are effectively switching roles.

TrendForce data points in the same direction. In 2026, AI training compute among the five largest North American CSPs is expected to grow by more than 56%, while AI inference compute could grow by nearly 122%.

In other words, once AI enters mass-market deployment, the biggest compute bill may no longer come from training the model.

It may come from letting millions of users continuously call it.


1. Why Is Inference Becoming AI's Biggest Cost Black Hole?

The problem is not necessarily the cost of a single inference request.

The problem is that inference never stops.

Training resembles a one-time capital investment.

Once a model has been trained, most of the training expense has already been incurred.

Inference is different.

Every user query, every Agent step, every code-generation request, every tool call, and every API request creates another compute expense.

SambaNova and other AI-chip companies have estimated that once models enter large-scale production, inference could account for 80%-90% of their lifetime compute costs.

Gartner has made a similar projection: global AI inference spending could reach $23.3 billion in 2026, exceeding the estimated $19 billion spent on training for the first time and representing 55% of AI-optimized IaaS spending. By 2029, inference could account for more than 65%.

There is another complication: Jevons' Paradox.

When tokens become cheaper, companies often do not use less AI.

They use much more.

Bain data shows that average token costs roughly halved from the end of 2024 to the end of 2025, while token consumption increased by 4.5×.

China's daily token usage has also reportedly increased from roughly 100 billion tokens in early 2024 to approximately 140 trillion tokens by March 2026.

The result is counterintuitive:

The price per token falls, while the total compute bill keeps rising.

Cheaper AI encourages more usage. More usage drives inference demand. Higher inference demand creates pressure on compute resources.

That is becoming one of the central infrastructure challenges of the AI economy.


2. Is Paying by Token Really the Fairest Pricing Model?

On the surface, token pricing looks extremely fair.

The model processes a certain number of tokens, and the user pays accordingly.

Simple.

Transparent.

Measurable.

But as AI Agents become more sophisticated, a deeper question emerges:

Are users really buying tokens—or are they buying results?

Consider a simple legal research task.

Model A uses 5,000 tokens to complete it.

Model B uses only 2,000 tokens and produces an equivalent result.

Under pure token-based pricing, Model A costs more.

In other words:

The less efficient model can effectively generate more billable usage, while the more efficient model is penalized for using fewer tokens.

That is a structural mismatch in today's token economy.

Model providers charge for computational consumption, while enterprise customers ultimately care about results, efficiency, and ROI.

Tokens are therefore only an intermediary unit of value.

The real destination is the value created by the task.


3. The Token Trading Market: Turning Compute Into a Tradable Commodity

This is where Unexhub is attempting to address the mismatch.

Unexhub is not simply positioning itself as another AI API aggregator. Its broader goal is to build a Token aggregation, distribution, and compute trading center.

The idea is to move the AI compute market away from a one-directional model where providers set prices and users simply accept them.

① Token Aggregation: Bringing More Compute Into One Market

Unexhub aggregates AI model resources across different tiers, including major models such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, as well as resources offered by individual developers and smaller companies.

The goal is straightforward:

Users do not have to be locked into a single model provider.

They can compare different models, resources, and pricing options within a broader marketplace.

For enterprises and developers, that means more choice—and potentially better price discovery.


② Millisecond Routing: Finding the Right Resource Automatically

Traditional API usage generally looks like this:

Choose a model → connect an API → continuously pay the provider's official price.

Unexhub aims to go one step further:

Demand arrives → market matching → millisecond routing → appropriate model resource.

When one provider becomes congested, other available resources can potentially be used. When another model offers a better price-performance ratio, routing can respond to the actual requirement.

In effect, this adds a traffic-management layer to the AI inference market.


③ API Bidding: Letting Supply and Demand Influence Price

This may be the most interesting component of a Token marketplace.

Traditional APIs typically use fixed pricing.

The same price applies during peak congestion and during periods when compute capacity is sitting idle.

But compute resources do not have constant supply and demand.

During peak periods, GPUs become scarce.

During off-peak periods, significant capacity may remain unused.

A fixed price therefore removes one of the most important mechanisms in any market:

the price signal.

Unexhub is experimenting with dynamic API bidding, allowing prices to respond to market supply and demand.

When demand is strong, providers can earn more.

When capacity is abundant, developers can potentially access lower-cost compute.

The difference is important.

The traditional model is:

The provider determines the price.

The marketplace model aims toward:

Supply and demand determine the price.


4. From Pay-Per-Token to Pay-for-Value

This may ultimately be the more important direction.

AI pricing could evolve through three stages.

Stage 1: Pay for resources

You consume a certain number of tokens and pay accordingly.

Stage 2: Pay for capabilities

You pay according to the model and capability you use.

Stage 3: Pay for outcomes

You pay according to the task the AI completes and the value it creates.

That suggests a potential transition:

Token → Intelligence → Outcome

As AI Agents become capable of independently handling research, coding, analysis, operations, customer service, and increasingly complex workflows, token count becomes a less complete measure of economic value.

The important question may gradually shift from:

“How many tokens did the AI use?”

to:

“How much value did those tokens create?”


5. Unexhub Is Not Just Chasing Lower Prices

If the only objective were to make AI APIs cheaper, the market could quickly turn into another price war.

What makes Unexhub's proposition more interesting is its attempt to address four issues simultaneously:

price, supply, liquidity, and settlement.

For developers, aggregating multiple providers, dynamic routing, and competitive pricing can expand the range of available choices.

Unexhub currently highlights several marketplace benefits:

  • Pay according to actual Token usage, reducing waste associated with fixed packages;

  • Multi-model aggregation, providing access to resources from Gemini, ChatGPT, Claude, DeepSeek, GLM, and other models;

  • Millisecond-level intelligent routing, matching requests with available model resources;

  • Dynamic bidding, allowing supply and demand to influence pricing;

  • Potential Token costs as low as roughly 30% of market prices, particularly attractive for high-frequency API calls, batch workloads, and AI Agent applications;

  • For suppliers, one-click channel integration and real-time settlement of usage revenue, reducing customer-acquisition and cash-flow pressure.

In other words, developers are not simply getting a “cheaper API.”

The bigger advantage is potentially this:

The same budget can purchase more tokens; the same tokens can be routed across more models; and the same AI requirement can have more pricing options.

For enterprises consuming millions or even billions of tokens every day, those differences can eventually translate directly into infrastructure costs and AI project ROI.


6. After Inference Consumes 80% of Compute, Who Recalculates the Bill?

The AI industry is entering a very unusual phase.

Models are becoming more capable.

Tokens are becoming cheaper.

But token consumption is growing even faster.

The question determining whether AI applications can scale may therefore no longer be:

“Who owns the most GPUs?”

It may increasingly become:

“Who can allocate those GPUs most efficiently?”

Inference is moving toward 80%-90% of lifetime AI compute costs.

That means compute is gradually becoming more than infrastructure.

It is becoming a digital resource that needs to be priced, traded, routed, and made liquid.

That is the market Unexhub is attempting to build:

a global Token aggregation and distribution center designed around a compute trading marketplace.

If the previous generation of AI competition was about model parameters, the next phase may be about something more fundamental:

Who can obtain more useful intelligence at a lower cost?

When tokens move from measurement units toward tradable assets, and compute moves from fixed infrastructure toward a liquid commodity, AI inference finally begins to develop a genuine mechanism for price discovery.

Inference may consume 80% of the compute.

But the next battle may be:

Who can make that 80% of compute deliver the most value?