从"永远降价"到峰谷翻倍:Token 交易市场这一刀下去,模型供应商的成本账全重算了

DeepSeek repeatedly lowered the price of its flagship models. Xiaomi’s MiMo series announced reductions of up to 99%. Some Tencent Cloud model services cut prices by more than 90%. GPT-5.6 Luna brought input pricing down to $1 per million tokens.

The Era of Permanent API Price Cuts Is Ending: Peak Pricing and Token Markets Are Rewriting the Economics of AI Compute

Over the past two years, large-model API pricing appeared to follow only one direction:

Downward.

DeepSeek repeatedly lowered the price of its flagship models. Xiaomi’s MiMo series announced reductions of up to 99%. Some Tencent Cloud model services cut prices by more than 90%. GPT-5.6 Luna brought input pricing down to $1 per million tokens.

Across China’s mainstream model market, price reductions exceeding 90% became increasingly common. JPMorgan has even projected that average token prices could decline by more than 40% over the next 12 months.

For a time, “permanent price cuts” seemed to have become an industry law:

Models would become more capable, while every token would become cheaper.

This formula was repeated in product launches, investor presentations, and industry reports.

Then, in the summer of 2026, the formula began to break.


From Permanent Discounts to Peak-Hour Doubling

On June 29, DeepSeek notified developers that the official release of V4 would introduce peak-and-off-peak pricing.

During peak periods, API prices would rise to twice the standard rate.

The designated peak windows were:

  • 9:00 a.m. to 12:00 p.m.;

  • 2:00 p.m. to 6:00 p.m. Beijing Time.

These hours cover most of the normal working day for domestic developers and enterprise users.

For DeepSeek V4 Pro, the price changes were significant:

  • Cache-hit input increased from RMB 0.025 to RMB 0.05 per million tokens;

  • Cache-miss input increased from RMB 3 to RMB 6;

  • Output increased from RMB 6 to RMB 12.

The price did not rise uniformly throughout the day. Instead, it doubled during the periods when demand and cluster pressure were highest.

That distinction matters.

This was not simply a conventional price increase. It was the introduction of a time-based scheduling mechanism into model API billing.

A company that had built its market position around extreme cost efficiency was now adding a peak-and-off-peak switch to its API.

DeepSeek was not alone.

During the same period:

  • Zhipu AI adjusted core-model API prices several times, with some coding packages increasing by more than 30%;

  • Certain Tencent Cloud Hunyuan services recorded much larger increases;

  • MiniMax adjusted both paid subscriptions and API pricing;

  • Doubao and Tongyi Qianwen announced changes or discontinuations affecting some agent capabilities.

Some model prices continued to decline, while others increased.

The API market was no longer following a single downward curve. It was splitting into a more complex pricing structure based on:

  • Usage time;

  • Model capability;

  • Service priority;

  • Context length;

  • Tool and agent features;

  • Capacity availability;

  • Business scenario.

The age of one universal token price was beginning to disappear.


Price Increases Are Not Simply Greed—They Reveal Compute Scarcity

DeepSeek explained that peak-and-off-peak pricing was intended to allocate resources more efficiently and improve service stability.

In engineering terms, the message was straightforward:

Compute capacity is limited, so prices must encourage non-urgent workloads to move away from peak hours.

Peak pricing is therefore not only a revenue mechanism.

It is also a standardized scheduling tool.

Tasks that can be delayed or processed in batches—such as large-scale document analysis, offline embedding generation, code indexing, and non-real-time data processing—can be shifted into lower-demand windows.

That preserves daytime capacity for workloads with stronger time requirements, including:

  • Real-time coding agents;

  • Financial analysis;

  • Online customer service;

  • Interactive search;

  • Production automation;

  • Time-sensitive enterprise applications.

Price becomes a traffic-control signal.

Instead of rejecting every request when the cluster is full, the provider uses economic incentives to reshape demand.


Compute Costs Have Not Disappeared

Large-model inference costs have fallen dramatically over the past several years.

Quantization, speculative decoding, caching, improved chips, better serving frameworks, and more efficient model architectures have reduced the cost of generating each token.

But the underlying physical expenses remain real:

  • Electricity;

  • GPU depreciation;

  • Data-center rent;

  • Network bandwidth;

  • Cooling;

  • Storage;

  • Redundancy;

  • Operations and maintenance.

The thousandfold decline in inference costs was largely a dividend from technological progress.

The problem begins when demand grows faster than efficiency improves.

If the cost of producing a token falls by 50%, but the number of tokens requested increases tenfold, total compute pressure still rises sharply.

That is exactly what is now happening.

China’s daily token-call volume has reportedly expanded by more than a thousandfold in two years, reaching enormous levels by March 2026. Inference has also become the dominant source of compute consumption, accounting for a growing share of total AI workloads.

The physical limit of a cluster cannot be negotiated away by a lower API price.

Kimi K3 offered a particularly visible example. Demand surged so quickly after launch that new consumer subscriptions had to be paused as the existing cluster approached its capacity limit.

The model itself had not failed.

The infrastructure had simply been asked to serve more demand than it could immediately support.

This reveals the weakness behind the “permanent price-cut” narrative:

The formula assumes that improvements in compute efficiency will always outpace the growth of demand.

There is no guarantee that they will.


The Token Marketplace Cuts Into Every Existing Cost Ledger

In the same period that DeepSeek introduced peak-and-off-peak pricing, Unexhub proposed a more fundamental restructuring of the token market.

According to its public positioning, Unexhub aims to build a compute-trading center for token aggregation and distribution.

Its model is built around three mechanisms:

  1. Real-time settlement;

  2. Zero-commission incentives for early suppliers;

  3. API bidding and dynamic pricing.

Together, these mechanisms attempt to transform tokens from billing units defined by model providers into market-priced resources that can be distributed and traded across a broader supply network.


Layer One: Real-Time Settlement

Under the traditional model, small and medium-sized suppliers often face a difficult cash-flow cycle.

They may need to pay first for:

  • GPUs;

  • Cloud instances;

  • Data-center capacity;

  • Power and bandwidth;

  • Model deployment;

  • API integration;

  • Customer support.

After supplying the service, they may still need to wait for usage calculation, invoicing, and payment.

The infrastructure has already been purchased and the compute has already been consumed, but the revenue remains locked in accounts receivable.

As the settlement period becomes longer, the supplier’s financial pressure increases.

More demand can paradoxically create more cash-flow risk:

The more capacity the supplier provides, the more capital it must temporarily finance.

Unexhub is attempting to shorten this cycle through real-time or near-real-time settlement.

The intended logic is simple:

Supply capacity, record actual usage, and settle the corresponding revenue without a prolonged billing delay.

For smaller suppliers, the value is not merely faster payment.

It can also mean:

  • Lower receivables pressure;

  • Faster capital turnover;

  • More predictable cash flow;

  • Quicker reinvestment into additional capacity;

  • Greater ability to respond to sudden demand.

This changes compute from a heavy asset waiting for repayment into a resource capable of circulating more efficiently.

Real-time settlement cannot eliminate hardware or operating costs. But it can reduce the opportunity cost created when capital remains trapped after the service has already been delivered.


Layer Two: Zero-Commission Early Entry

Unexhub’s roadmap also includes zero-commission incentives for qualifying early suppliers.

In a rapidly developing token market, zero commission can provide more than a short-term discount.

It gives early suppliers room to:

  • Complete technical integration at a lower cost;

  • Test actual customer demand;

  • Build usage and settlement records;

  • Establish trust with enterprise buyers;

  • Validate their unit economics;

  • Secure early distribution channels;

  • Capture market share before the ecosystem matures.

New suppliers typically face several costs simultaneously:

  • Infrastructure investment;

  • Technical integration;

  • Customer acquisition;

  • Service operations;

  • Settlement management;

  • Platform commissions.

Reducing commissions during the initial stage allows a supplier to complete its first commercial cycle with less pressure:

Connect, distribute, generate usage, settle revenue, and reinvest.

The true value of zero commission is therefore not only the fee saved.

It is the ability to test whether the business model works before taking on the full long-term cost structure.


Layer Three: API Bidding and Dynamic Pricing

The third mechanism is the most consequential:

Allow API prices to move dynamically with market supply and demand.

During off-peak periods:

  • Developers gain access to lower prices;

  • Idle capacity can be monetized;

  • Suppliers improve cluster utilization;

  • Batch workloads can be completed more economically.

During peak periods:

  • Prices can rise in response to scarcity;

  • Suppliers receive incentives to release additional capacity;

  • Priority workloads can receive faster access;

  • Resources move toward users who value immediacy most.

Dynamic pricing is not a new invention.

Cloud providers have long used spot instances, reserved capacity, priority tiers, and time-sensitive pricing.

What is new is the attempt to apply similar logic across an open market of token APIs and model suppliers.

Instead of prices being hidden inside one provider’s proprietary system, a broader marketplace could allow price discovery to reflect:

  • Real-time capacity;

  • Current demand;

  • Model quality;

  • Latency;

  • Reliability;

  • Context size;

  • Tool-use capability;

  • Service-level guarantees;

  • Regional availability.

In that structure, the market—not a single supplier—plays a larger role in determining price.


Tokens Are Moving From Provider-Priced Units to Tradable Assets

Taken together, real-time settlement, early-supplier incentives, and API bidding all point toward a deeper transformation.

Under the traditional model, a token is a billing unit.

The model provider defines:

  • The input price;

  • The output price;

  • Cache discounts;

  • Priority premiums;

  • Capacity limits.

The user either accepts the price or chooses another provider.

Under a marketplace model, token value is influenced by more variables:

  • Which model generated the token;

  • When the request was made;

  • How scarce capacity was;

  • Whether cached input was used;

  • Which region or route handled the request;

  • Whether priority service was required;

  • What quality guarantee the supplier provided.

The token begins to behave less like a static utility meter and more like a tradable resource.

Of course, model tokens are not fully interchangeable.

One million tokens from a lightweight extraction model are not equivalent to one million tokens from a frontier reasoning or coding model.

A functioning market would therefore need to price both quantity and quality.

Possible categories could include:

  • Standard and priority tokens;

  • Peak and off-peak capacity;

  • Cached and uncached input;

  • General-purpose and specialized models;

  • Guaranteed and best-effort throughput;

  • Low-latency and economy routes;

  • Domestic and international model supply.

This is not a market for identical commodities.

It is closer to a market for differentiated compute services measured in tokens.


The Cost-Plus Pricing Model Is Being Challenged

Historically, model API pricing followed a cost-plus structure:

  1. Calculate compute and operating costs;

  2. Add a target margin;

  3. Publish a fixed API price.

The process was largely one-way.

The supplier defined the price, while users responded through demand.

A marketplace introduces a different mechanism.

Prices can be discovered through real-time interaction between supply and demand:

  • High demand pushes prices upward;

  • Excess capacity pushes prices downward;

  • Suppliers compete on price, quality, latency, and reliability;

  • Developers choose according to workload requirements;

  • Platforms route traffic toward the best available combination.

Suppliers no longer need to estimate every market condition in advance and embed it into one static price.

The marketplace continuously recalculates the value of available capacity.


Instant Settlement Changes the Capital Cost

The most significant change may not be the visible API price.

It may be the cost of capital.

For small and medium-sized suppliers, the largest burden is often not the cost of one inference request. It is the time between paying for infrastructure and receiving revenue.

Consider the traditional sequence:

  1. Purchase or rent hardware;

  2. Build the service;

  3. Provide API capacity;

  4. Accumulate usage;

  5. Issue invoices;

  6. Wait for settlement;

  7. Receive payment;

  8. Reinvest in new capacity.

During steps four through seven, capital is locked.

If demand suddenly rises, the supplier may be unable to expand because the money needed for additional capacity remains trapped in receivables.

Shortening the settlement cycle improves more than liquidity.

It can also improve the speed at which the market responds to shortages.

When suppliers recover funds faster, they can:

  • Rent additional GPUs;

  • Expand routes;

  • Increase redundancy;

  • Improve service quality;

  • Support larger bursts of demand.

The settlement system therefore becomes part of the compute-supply system.


The Fundamental Logic of AI Cost Accounting Has Changed

DeepSeek’s peak pricing and Unexhub’s marketplace model point in the same direction:

Tokens are shifting from fixed-price units toward resources whose value fluctuates with supply and demand.

This has several consequences for model suppliers.


Revenue Moves From Linear to Dynamic

Under the old model:

API revenue = call volume × fixed unit price.

The calculation was simple.

Under dynamic pricing:

API revenue = call volume × variable unit price.

Revenue now depends on:

  • How much demand arrives during peak hours;

  • Whether users can be shifted into off-peak windows;

  • How much idle capacity is successfully sold;

  • Whether premium routes attract higher-value workloads;

  • How effectively the supplier participates in multiple markets.

Two suppliers with the same capacity may generate very different revenue depending on how well they manage traffic.


Costs Move From Fixed Toward Variable

Traditional compute is heavily capital-intensive.

Suppliers purchase GPUs, build clusters, and depreciate equipment over time.

The cost exists even when utilization is low.

In a more liquid token market, some capacity can be purchased or routed on demand.

A supplier may combine:

  • Owned hardware;

  • Cloud instances;

  • Third-party capacity;

  • Marketplace routes;

  • Reserved and spot resources.

Part of the cost base shifts from fixed capital expenditure toward variable operating expenditure.

This does not eliminate infrastructure ownership, especially for high-performance or sensitive workloads.

But it gives suppliers more flexibility in responding to demand.


Competition Moves From “Who Is Cheapest?” to “Who Schedules Best?”

During the permanent-price-cut era, the market rewarded whoever could publish the lowest price per million tokens.

In a dynamic market, the winning supplier may be the one that can:

  • Fill idle capacity;

  • Shift deferrable traffic to off-peak periods;

  • Maintain performance during peaks;

  • Route tasks to the most appropriate model;

  • Control retry and failure costs;

  • Use caching effectively;

  • Convert price volatility into additional margin.

Scheduling, routing, and utilization become as important as the model itself.

The question changes from:

“Who has the lowest token price?”

to:

“Who can deliver the required result at the lowest total task cost?”


From GPU Counts to Token Economics

At WAIC 2026, enterprise buyers increasingly shifted their questions.

Instead of asking only:

“How many GPUs do you own?”

They began asking:

  • What is the cost per million tokens?

  • What is the cost per completed task?

  • What is the average response latency?

  • Can throughput be guaranteed?

  • How does pricing change during peak demand?

  • Can requests fail over to another model?

  • How quickly can additional capacity be added?

This reflects a larger industry transition:

  • Compute is moving from pure capital expenditure toward operating expenditure;

  • AI demand is shifting from training toward inference;

  • Competition is moving from hardware ownership toward cost and scheduling efficiency;

  • Tokens are becoming a commercial settlement layer between models, capacity, and users.

The number of GPUs still matters.

But how efficiently those GPUs are converted into useful, reliable tokens matters more.


Unexhub’s Role in a Multi-Model Token Market

According to its positioning, Unexhub is building a global token aggregation and distribution network that connects different levels of AI suppliers.

The platform brings together access to major model ecosystems such as:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • Other models provided by individual developers and small or medium-sized companies.

For enterprise users and developers, the practical benefit is a more unified access layer.

Instead of separately maintaining accounts, SDKs, API keys, billing methods, and error-handling systems for every provider, users can connect through a common gateway.

Unexhub also emphasizes millisecond-level routing.

Requests can be routed according to factors such as:

  • Model capability;

  • Current price;

  • Availability;

  • Latency;

  • Workload type;

  • Region;

  • Traffic pressure;

  • Enterprise policy.

This allows different workloads to be assigned to different models rather than forcing every task through the same expensive route.

For example:

  • Lightweight extraction can be assigned to a low-cost model;

  • Complex reasoning can be routed to a stronger model;

  • Coding workloads can use a specialist route;

  • Batch tasks can wait for off-peak pricing;

  • Urgent enterprise requests can use priority capacity.

Developers pay according to actual token consumption. Depending on the model, route, and market conditions, some services may be available at prices as low as approximately 30% of standard market rates.

For high-volume users, that benefit can extend beyond a single discounted request.

It may reduce the cost of:

  • Long-context processing;

  • Coding agents;

  • Batch document analysis;

  • Customer-service automation;

  • Multi-model testing;

  • High-frequency production workloads.


Value for Suppliers: One Connection, Broader Distribution

For model and compute suppliers, Unexhub’s value proposition is different.

A supplier may have a strong model, regional capacity, or idle inference resources, but lack:

  • Enterprise sales channels;

  • Global distribution;

  • Billing infrastructure;

  • Traffic acquisition;

  • Settlement systems;

  • Multi-tenant API management.

Through one integration, the supplier can potentially gain access to a broader developer and enterprise market.

Combined with real-time settlement, this creates a more direct path from available capacity to revenue:

Connect once, receive routed traffic, and settle according to usage.

For smaller suppliers, that can lower both customer-acquisition and financial barriers.

For larger suppliers, it can provide an additional distribution channel and a way to monetize idle or off-peak capacity.


The Developer Benefit Is More Than a Lower Price

A token marketplace should not be evaluated only by whether it offers a cheaper API rate.

For developers, the more meaningful benefits may include:

  • Unified access to multiple models;

  • Less dependence on one provider;

  • Automatic switching when capacity is limited;

  • Flexible choices between price and latency;

  • Lower off-peak rates;

  • Centralized billing;

  • More transparent usage data;

  • Faster experimentation across models.

A price as low as 30% of standard market rates is attractive.

But the long-term value comes from combining that discount with:

  • Stable routing;

  • Predictable quality;

  • Clear billing;

  • High availability;

  • Enterprise support.

Cheap capacity that frequently fails is not truly cheap.

The relevant metric remains total cost per completed task.


The Marketplace Model Still Needs Real-World Proof

The direction of token trading is compelling, but implementation will determine whether it can become genuine infrastructure.

A functioning marketplace must answer difficult questions:

  • How is model quality measured?

  • How are suppliers ranked?

  • Who bears the cost of failed calls?

  • How are retries settled?

  • How is latency guaranteed?

  • How are privacy and compliance requirements enforced?

  • How are enterprise keys and data protected?

  • How is price manipulation prevented?

  • How are different providers’ service-level agreements standardized?

  • Can real-time settlement remain stable at high transaction volumes?

The market cannot depend on low prices alone.

It must provide:

  • Secure routing;

  • Transparent settlement;

  • Reliable capacity;

  • Consistent output quality;

  • Clear accountability;

  • Sustainable supplier economics.

Unexhub’s instant settlement, zero-commission onboarding, API bidding, and millisecond routing address real problems.

Their ultimate value will depend on how reliably they operate at scale.


Final Thoughts: The Permanent-Price-Cut Era Is Over

For two years, the large-model market trained developers to expect one thing:

Every new model would be stronger and cheaper than the last.

That trend has not disappeared. Technological progress will continue to reduce the underlying cost of intelligence.

But the pricing structure around that intelligence is becoming more complicated.

The next era will include:

  • Peak and off-peak pricing;

  • Priority routes;

  • Dynamic API bidding;

  • Cached-input discounts;

  • Marketplace settlement;

  • Multiple quality tiers;

  • Real-time model routing;

  • Tradeable compute capacity.

DeepSeek’s peak-hour doubling is not an isolated reversal.

It is a signal that fixed prices can no longer absorb unlimited demand.

Unexhub’s token marketplace represents the other side of the same transition: instead of allowing one supplier to define a static price, create a broader market in which compute can be aggregated, routed, priced, and settled dynamically.

Once that change occurs, every model supplier must recalculate its ledger:

  • Revenue is no longer purely linear;

  • Compute does not have to remain entirely fixed;

  • Capital turnover becomes part of infrastructure efficiency;

  • Scheduling becomes a competitive advantage;

  • Token pricing becomes a market function.

The age of “permanent price cuts” is ending.

The age of dynamic token economics is beginning.

And the companies that recalculate their costs, routes, and settlement systems fastest may hold the pricing power of the next AI era.


中文版本

大模型 API“永久降价”时代正在结束:峰谷定价与 Token 市场开始重写算力成本账

过去两年,大模型 API 的价格似乎只有一个方向:

持续向下。

DeepSeek 多次下调旗舰模型价格,小米 MiMo 系列最高降幅达到 99%,腾讯云部分模型服务降价超过 90%,GPT-5.6 Luna 的输入价格也降至每百万 Token 1 美元。

在国内主流模型市场中,超过 90% 的价格降幅逐渐变得并不罕见。摩根大通甚至预测,未来 12 个月 Token 平均单价还可能进一步下降超过 40%。

一段时间内,“永久降价”几乎成为行业铁律:

模型能力越来越强,每个 Token 的价格越来越低。

这套公式反复出现在新品发布会、投资者演示和行业报告中。

但到了 2026 年夏天,这条铁律开始被打破。


从“永久降价”转向“高峰时段翻倍”

6 月 29 日,DeepSeek 向开发者发送通知,宣布 V4 正式版将引入峰谷分时定价机制。

在高峰时段,API 价格将上调至普通时段的两倍。

高峰时间包括:

  • 北京时间 9:00—12:00;

  • 北京时间 14:00—18:00。

这几乎覆盖了国内开发者和企业用户最核心的工作时间。

以 DeepSeek V4 Pro 为例:

  • 缓存命中输入从每百万 Token 0.025 元提高至 0.05 元;

  • 未命中缓存输入从 3 元提高至 6 元;

  • 输出从 6 元提高至 12 元。

价格并不是全天统一上涨,而是在需求最旺盛、集群压力最大的时段翻倍。

这一区别非常重要。

它并非传统意义上的简单涨价,而是将一种基于时间的算力调度机制加入 API 计费体系。

曾经依靠极致性价比迅速占领市场的 DeepSeek,开始主动在 API 上安装“峰谷开关”。

DeepSeek 并不是唯一一家调整价格与服务结构的厂商。

同一时期:

  • 智谱 AI 年内多次调整核心模型 API 价格,部分编程套餐涨幅超过 30%;

  • 腾讯云部分混元服务版本出现更大幅度调整;

  • MiniMax 同步调整订阅套餐和 API 价格;

  • 豆包、通义千问也先后调整或停止部分智能体功能。

同一个市场中,有的模型仍在降价,有的服务却开始涨价。

大模型 API 已不再沿着单一的下降曲线前进,而是逐渐分化为按照以下因素定价的复杂结构:

  • 使用时段;

  • 模型能力;

  • 服务优先级;

  • 上下文长度;

  • 工具与智能体功能;

  • 实时算力供给;

  • 具体业务场景。

一个 Token 对应一个固定价格的时代,正在逐渐消失。


涨价不只是“贪心”,而是藏不住的算力稀缺

DeepSeek 对峰谷定价的官方解释,是更加合理地分配资源并提高服务稳定性。

翻译成工程语言就是:

算力有限,需要通过价格引导非紧急需求转移到低谷时段。

因此,高峰定价不只是增加收入的方式,也是标准化的算力调度工具。

一些可以延后或批量执行的任务,例如:

  • 大规模文档分析;

  • 离线向量生成;

  • 代码仓库索引;

  • 非实时数据处理;

  • 批量内容生成;

可以被转移到价格更低、负载更小的时间窗口。

这样便能够为以下高时效任务保留工作时段的算力:

  • 实时代码智能体;

  • 金融分析;

  • 在线客服;

  • 交互式搜索;

  • 生产自动化;

  • 对延迟敏感的企业应用。

价格开始承担流量调度信号的作用。

相比在集群满载时直接拒绝所有请求,模型厂商希望利用经济杠杆重新分配需求。


底层算力成本从未真正消失

过去几年,大模型推理成本确实出现了大幅下降。

量化、投机解码、缓存、新型芯片、推理框架优化以及更高效的模型架构,共同降低了生成每个 Token 的成本。

但底层物理支出依然真实存在:

  • 电力;

  • GPU 折旧;

  • 数据中心租金;

  • 网络带宽;

  • 散热;

  • 存储;

  • 容灾与冗余;

  • 日常运维。

推理成本数年内下降千倍,更多是技术进步带来的红利。

真正的问题出现在:需求增长速度超过效率提升速度。

假设生成一个 Token 的成本下降了 50%,但市场调用量增加了十倍,总体算力压力依然会显著上升。

这正是当前行业正在发生的事情。

国内日均 Token 调用量在两年内增长超过千倍,到 2026 年 3 月已达到极为庞大的规模。推理也逐渐成为算力消耗的核心来源,占据全部 AI 工作负载中越来越高的比例。

无论 API 价格多低,物理集群的容量上限都不会自动消失。

Kimi K3 就提供了一个非常直观的案例。模型上线后需求增长过快,现有集群逼近容量上限,平台不得不暂停 C 端新用户订阅。

这并不是产品出了问题。

只是基础设施在短时间内承受了超出预期的调用量。

这暴露了“永久降价”叙事背后的薄弱前提:

它假设算力效率的提升,永远可以跑赢需求增长。

但现实并不保证这一点。


Token 交易市场的一刀,正在重算所有成本账

就在 DeepSeek 推出峰谷定价的同一时期,Unexhub 提出了一套更具结构性的 Token 市场方案。

根据其公开定位,Unexhub 希望建设一个 Token 聚合与分发的算力交易中心。

其核心机制包括三层:

  1. 即时结算;

  2. 早期供应商零佣金激励;

  3. API 竞价与动态定价。

三项机制结合起来,试图将 Token 从模型厂商单方面定义的计费单位,转变为可以在更广泛供给网络中分发、定价与交易的资源。


第一层:即时结算

在传统模式下,中小型算力和模型供应商通常面临一条艰难的现金流链路。

他们需要先支付:

  • GPU 采购或租赁费用;

  • 云主机成本;

  • 数据中心资源费用;

  • 电力与带宽;

  • 模型部署成本;

  • API 接入费用;

  • 客户服务与运维成本。

服务已经提供、算力已经被消耗,但供应商仍要等待用量核算、开票和回款。

随着结算周期拉长,资金持续沉淀在应收账款中。

需求越多,甚至可能带来更大的现金流风险:

供应商提供得越多,需要临时垫付的资金也越多。

Unexhub 希望通过即时或准即时结算,缩短这一周期。

基本逻辑是:

供应算力、记录真实调用,并尽可能减少漫长账期,及时结算相应收益。

对中小供应商而言,价值不只是更快收到一笔款项,还包括:

  • 降低应收账款压力;

  • 加快资金周转;

  • 获得更可预测的现金流;

  • 更快将收益投入新增算力;

  • 提高应对突发需求的能力。

算力由等待回款的重资产,转变为可以更加高效循环的生产资源。

即时结算无法消除硬件和运营支出,却可以降低服务已经交付、资金仍被长期占用带来的机会成本。


第二层:早期零佣金入驻

根据 Unexhub 的规划,符合条件的早期供应商可享受零佣金激励。

在快速发展的 Token 市场中,零佣金不仅是一项短期优惠,还能帮助供应商以更低成本:

  • 完成技术接入;

  • 测试真实客户需求;

  • 建立调用和结算记录;

  • 积累企业用户信任;

  • 验证单位经济模型;

  • 提前锁定分发渠道;

  • 在生态成熟前获取市场份额。

新供应商在早期往往需要同时承担:

  • 基础设施投入;

  • 技术开发成本;

  • 获客成本;

  • 服务运营成本;

  • 结算管理成本;

  • 平台抽佣。

降低早期佣金,可以让供应商用更小压力跑通第一轮完整商业闭环:

接入、分发、产生调用、完成结算,再继续投入。

零佣金真正的价值,不只是省下一笔手续费。

而是帮助供应商在承担完整长期成本之前,先验证业务模式能否成立。


第三层:API 竞价与动态定价

第三项机制最值得关注:

让 API 价格根据市场供需动态变化。

在低谷时段:

  • 开发者能够获得更低价格;

  • 闲置算力可以被利用;

  • 供应商提高集群利用率;

  • 批量任务能够以更低成本完成。

在高峰时段:

  • 价格可以根据稀缺程度上浮;

  • 供应商获得释放更多算力的动力;

  • 优先级较高的任务能够更快获得资源;

  • 算力流向最看重实时性的用户。

动态价格并不是新发明。

云计算行业早已有竞价实例、预留容量、优先级服务与弹性价格。

新的地方在于:将类似逻辑应用到一个聚合多家模型和算力供应商的开放 Token 市场。

价格不再完全隐藏在某一家厂商的内部系统中,而可能更加直接地反映:

  • 实时可用算力;

  • 当前调用需求;

  • 模型质量;

  • 延迟;

  • 稳定性;

  • 上下文容量;

  • 工具调用能力;

  • SLA 服务等级;

  • 区域可用性。

价格发现的权力开始由单一供应商,部分转向更广泛的市场。


Token 正从模型定价单位转向可交易资产

将即时结算、早期激励与 API 竞价结合起来,可以看到一个更深层变化。

传统模式下,Token 只是计费单位。

模型厂商决定:

  • 输入价格;

  • 输出价格;

  • 缓存折扣;

  • 优先调用溢价;

  • 使用上限。

用户只能接受价格,或者转向其他模型。

在市场化模式下,Token 价值将受到更多因素影响:

  • 由哪个模型生成;

  • 在什么时间调用;

  • 当时算力是否紧张;

  • 输入是否命中缓存;

  • 由哪个区域或线路处理;

  • 是否要求优先级;

  • 供应商提供怎样的服务保障。

Token 开始从静态计量单位,转向能够被交易和动态定价的资源。

当然,不同模型的 Token 并不完全同质。

轻量信息抽取模型的 100 万 Token,与前沿推理或编程模型的 100 万 Token,显然不具备相同价值。

成熟市场必须同时为数量与质量定价。

未来可能出现:

  • 标准 Token 与优先 Token;

  • 高峰算力与低谷算力;

  • 缓存输入与未缓存输入;

  • 通用模型与专业模型;

  • 保证吞吐与尽力而为服务;

  • 低延迟线路与经济线路;

  • 国内模型与海外模型供给。

这并不是一个完全标准化的商品市场,而更像是以 Token 为计量方式的差异化算力服务市场。


传统成本加成定价正在被挑战

过去,模型 API 多采用成本加成逻辑:

  1. 计算算力和运营成本;

  2. 加入目标利润;

  3. 发布固定 API 价格。

这是一个单向过程。

供应商制定价格,用户通过调用量表达需求。

Token 市场则引入了另一种机制。

价格可以在真实供需互动中形成:

  • 需求旺盛时价格上浮;

  • 算力过剩时价格下降;

  • 供应商通过价格、质量、延迟和可靠性竞争;

  • 开发者根据任务需求进行选择;

  • 平台将流量路由到更合适的供给组合。

供应商不必提前预测所有市场情况,并将风险全部写进一个固定价格中。

市场会持续重新计算可用算力的价值。


即时结算真正改变的是资金成本

最重要的变化,未必是用户看到的 API 单价。

而可能是供应商的资金成本。

对中小型供应商来说,最大的压力往往不是某一次推理请求花费多少,而是从投入基础设施到真正收到收入需要多长时间。

传统流程通常是:

  1. 购买或租赁硬件;

  2. 建设服务;

  3. 提供 API 算力;

  4. 累积用户调用;

  5. 开具账单;

  6. 等待结算;

  7. 收到资金;

  8. 再投资新增容量。

在第四步到第七步之间,资金被锁住了。

如果需求突然增长,供应商可能无法立即扩容,因为用于新增算力的资金仍沉淀在应收账款中。

缩短结算周期,不仅能改善流动性,还能提高整个市场应对算力短缺的速度。

资金更快回流后,供应商可以更快:

  • 租赁更多 GPU;

  • 增加模型线路;

  • 完善冗余;

  • 提升服务质量;

  • 承接突发流量。

结算系统由此成为算力供给系统的一部分。


AI 成本核算的底层逻辑已经改变

DeepSeek 的峰谷定价与 Unexhub 的 Token 市场,都指向同一个方向:

Token 正从固定价格的计费单位,转变为随供需变化而波动的资源。

这将为模型供应商带来多项变化。


收入结构从线性变为动态

过去的计算方式是:

API 收入 = 调用量 × 固定单价。

简单而直接。

动态定价下则变成:

API 收入 = 调用量 × 浮动单价。

最终收入将取决于:

  • 高峰期有多少需求;

  • 能否将用户引导至低谷时段;

  • 闲置算力能否被有效出售;

  • 高价值任务是否愿意购买优先线路;

  • 供应商能否进入更多市场和分发渠道。

即使两家供应商拥有相同算力,其收入也可能因为调度效率不同而产生巨大差异。


成本结构从固定走向可变

传统算力具有明显的资本密集属性。

供应商采购 GPU、建设集群,并在较长周期内折旧。

即使利用率不高,成本也已经发生。

在流动性更高的 Token 市场中,部分算力可以按照需求动态采购或路由。

供应商可以组合使用:

  • 自有硬件;

  • 云端实例;

  • 第三方算力;

  • 市场化模型线路;

  • 预留与竞价资源。

部分成本会从固定资本支出,转变为更加灵活的运营支出。

这不会让所有企业停止自建集群,尤其是对性能、安全和合规要求较高的场景。

但它能够提高供应商应对需求变化的灵活性。


竞争从“谁更便宜”转向“谁调度得更好”

在永久降价阶段,市场更奖励每百万 Token 报价最低的模型。

动态市场中,真正占优的供应商可能是最擅长以下工作的企业:

  • 填满闲置算力;

  • 将可延后的流量转移至低谷;

  • 在高峰时维持性能;

  • 将任务路由到最合适的模型;

  • 控制失败和重试成本;

  • 提高缓存命中率;

  • 将价格波动转化为利润空间。

调度、路由和利用率,开始与模型能力本身同样重要。

市场问题从:

“谁的 Token 单价最低?”

转变为:

“谁能以最低的总任务成本,稳定交付结果?”


企业从关心 GPU 数量,转向关心 Token 经济性

在 WAIC 2026 上,企业客户的问题开始发生变化。

过去,他们更习惯问:

“你们有多少张卡?”

现在,他们越来越关心:

  • 每百万 Token 的真实成本是多少?

  • 每完成一个任务需要花多少钱?

  • 平均响应延迟是多少?

  • 能否保障吞吐量?

  • 高峰时段价格如何变化?

  • 一个模型不可用时能否自动切换?

  • 新增算力需要多长时间?

这反映了更大的产业变化:

  • 算力正在从纯资本支出走向运营支出;

  • AI 需求正在从训练转向推理;

  • 竞争正在从硬件数量转向成本与调度效率;

  • Token 正成为连接模型、算力和用户的商业结算层。

GPU 数量依然重要。

但如何将 GPU 高效转化为稳定、可用且价格合理的 Token,更加重要。


Unexhub 在多模型 Token 市场中的角色

根据其公开定位,Unexhub 正在建设面向全球的 Token 聚合与分发网络,连接不同量级的 AI 供应商。

平台汇聚的模型生态包括:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • 个人开发者和中小企业提供的其他优质模型。

对开发者和企业用户而言,最直接的价值是统一接入。

用户无需分别管理每一家厂商的:

  • 账户;

  • SDK;

  • API 密钥;

  • 计费方式;

  • 异常处理系统。

通过统一网关,即可接入不同模型服务。

Unexhub 同时强调毫秒级路由,请求可以根据以下因素被分配:

  • 模型能力;

  • 实时价格;

  • 可用性;

  • 延迟;

  • 任务类型;

  • 服务区域;

  • 流量压力;

  • 企业策略。

这意味着,不同任务可以使用不同模型,而不必强迫所有需求都走同一条昂贵线路。

例如:

  • 轻量信息抽取交给低成本模型;

  • 复杂推理交给更强模型;

  • 编程任务选择专业线路;

  • 批量任务等待低谷价格;

  • 企业紧急请求使用优先算力。

开发者根据真实 Token 用量付费。按照具体模型、线路与市场条件,部分服务价格最低可达到常规市场价格的约 3 折。

对高频用户来说,这项权益不只是降低一次调用的价格,还可能持续减少:

  • 长上下文处理成本;

  • 编程智能体费用;

  • 批量文档分析支出;

  • 客服自动化成本;

  • 多模型测试费用;

  • 高频生产调用预算。


对供应商的价值:一次接入,获得更多分发渠道

对模型和算力供应商而言,Unexhub 提供的是另一种价值。

一家供应商可能拥有优质模型、区域性算力或闲置推理资源,却缺少:

  • 企业销售渠道;

  • 全球分发能力;

  • 计费基础设施;

  • 用户流量;

  • 结算系统;

  • 多租户 API 管理能力。

通过一次接入,供应商有机会触达更广泛的开发者和企业市场。

再结合即时结算,可以形成一条更加直接的变现路径:

一次接入、接受平台路由,并按照真实调用及时获得结算。

对中小供应商而言,这能够同时降低获客与资金门槛。

对大型供应商而言,它则可以成为新的分发渠道,并帮助出售闲置或低谷算力。


开发者获得的不只是更低价格

Token 市场不能只用“API 是否更便宜”来衡量。

对开发者而言,更有价值的权益还可能包括:

  • 多模型统一接入;

  • 减少对单一供应商的依赖;

  • 算力紧张时自动切换;

  • 灵活选择价格与延迟;

  • 获得低谷优惠;

  • 统一账单;

  • 更透明的用量数据;

  • 更快速地测试不同模型。

低至市场价 3 折的价格确实具有吸引力。

但长期价值必须建立在以下能力之上:

  • 稳定路由;

  • 可预测的服务质量;

  • 清晰账单;

  • 高可用性;

  • 企业级支持。

一个频繁调用失败的低价 API,并不是真正的低成本服务。

最终需要衡量的,依然是每个成功任务的总成本。


Token 市场仍需接受真实运营检验

Token 交易的方向值得关注,但能否成为真正的基础设施,取决于实际执行。

一个成熟市场必须回答:

  • 如何衡量模型质量?

  • 如何评价和淘汰供应商?

  • 调用失败成本由谁承担?

  • 重试费用如何结算?

  • 延迟如何保障?

  • 隐私与合规要求如何落实?

  • 企业密钥与数据如何保护?

  • 如何防止价格操纵?

  • 不同供应商的 SLA 如何统一?

  • 高频交易下,即时结算能否保持稳定?

市场不能只依靠低价吸引用户。

它还必须提供:

  • 安全路由;

  • 透明结算;

  • 稳定算力;

  • 一致的输出质量;

  • 清晰的责任边界;

  • 可持续的供应商收益。

Unexhub 提出的即时结算、零佣金入驻、API 竞价和毫秒级路由,确实瞄准了真实问题。

其最终价值,仍取决于能否在大规模运行中保持可靠。


写在最后:永久降价时代已经结束

过去两年,大模型市场让开发者形成了一种预期:

每一代新模型都会比上一代更强,也更便宜。

这一趋势并不会完全消失。技术进步仍会继续降低智能生产的底层成本。

但围绕这种智能能力的价格结构,将变得越来越复杂。

下一阶段可能同时出现:

  • 峰谷分时定价;

  • 优先级线路;

  • API 动态竞价;

  • 缓存输入折扣;

  • 市场化结算;

  • 多种质量等级;

  • 实时模型路由;

  • 可交易算力容量。

DeepSeek 的高峰价格翻倍,并不是一次孤立的反转。

它说明固定价格已经难以承载无限增长的需求。

Unexhub 的 Token 市场则代表着同一变化的另一侧:不再完全由单一模型厂商设定静态价格,而是建立一个能够聚合、路由、定价和结算算力的更广泛市场。

一旦这套逻辑成立,所有模型供应商都必须重算成本账:

  • 收入不再完全线性;

  • 算力不必全部固定化;

  • 资金周转成为基础设施效率的一部分;

  • 调度能力成为竞争优势;

  • Token 价格逐渐成为市场函数。

“永久降价”的时代正在结束。

动态 Token 经济时代,正在开始。

而那些最早完成成本、路由与结算体系重构的企业,将更有机会掌握下一阶段 AI 市场的定价权。