DeepSeek 峰谷、腾讯云跟调、高盛定性价格战尾声:Token 交易市场正在逼模型供应商学电网定价

For the previous two years, the industry had grown accustomed to a simple narrative: models would become more powerful, while token prices would continue to fall.

The End of One-Price-Fits-All AI APIs: Peak Pricing and Token Markets Are Rewriting the Economics of Compute

In the summer of 2026, the pricing logic behind large-model APIs began to change fundamentally.

For the previous two years, the industry had grown accustomed to a simple narrative: models would become more powerful, while token prices would continue to fall.

That trend has not disappeared. However, a single downward pricing curve is no longer enough to describe the market.

AI API pricing is now becoming:

  • Time-based;

  • Tiered by service quality;

  • Differentiated by workload;

  • Influenced by real-time capacity;

  • Increasingly shaped by supply and demand.

DeepSeek’s peak-and-off-peak pricing was one of the clearest signals of this transition.


DeepSeek Added a Peak-and-Off-Peak Switch to Its API

On June 29, DeepSeek notified developers that the official V4 release would introduce time-based pricing.

During peak periods, API rates would rise to twice the standard level.

The peak windows covered:

  • 9:00 a.m. to 12:00 p.m.;

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

These hours include nearly the entire core working window for developers and enterprise users in China.

For V4 Pro:

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

  • Output increased from RMB 6 to RMB 12 per million tokens.

A company that had built its market reputation around extreme cost efficiency was now applying a peak-and-off-peak mechanism to its API.

This was more than a conventional price increase.

The price did not rise equally for every request. It changed according to when the request was made and how much pressure the infrastructure was under.

DeepSeek was effectively importing the logic of an electricity grid into the AI compute market:

Higher prices during periods of concentrated demand, lower prices when capacity is more available.


DeepSeek Was Not an Isolated Case

Other model and cloud providers were also adjusting their pricing structures.

Tencent Cloud’s Hunyuan models moved from free public testing toward commercial billing. Some versions experienced substantial price changes. HY2.0-Instruct input pricing reportedly increased from RMB 0.0008 to RMB 0.004505 per thousand tokens, an adjustment of approximately 463%.

During the same period:

  • Zhipu AI’s API pricing rose significantly month over month;

  • Some Alibaba Cloud and Baidu AI Cloud compute products increased by between 5% and 50%;

  • Model providers began revising packages, usage limits, and commercial service structures.

In the same market, some prices continued to fall while others moved sharply upward.

This apparent contradiction reveals the real direction of the industry:

Large-model API pricing is no longer moving through one universal discount channel. It is splitting into a more complex structure based on time, model capability, service level, and usage scenario.

A batch-processing task that can wait until midnight may receive one price.

A real-time coding agent running during office hours may receive another.

A standard route may be inexpensive, while guaranteed throughput and priority service carry a premium.

The idea that every token from the same model should cost the same at every moment is beginning to disappear.


The Aggressive Price War May Be Approaching Its Limit

Goldman Sachs offered a direct interpretation of this change.

A research note published on July 3 argued that DeepSeek’s peak-and-off-peak mechanism could indicate that the aggressive price competition seen since late April 2026 was beginning to reach its limit.

Earlier price wars had compressed margins across the industry. In some cases, API rates were pushed close to zero-profit—or even loss-making—levels.

That strategy made sense during the land-grab stage.

Low prices helped model providers:

  • Attract developers;

  • Increase API traffic;

  • Build ecosystem dependence;

  • Encourage experimentation;

  • Move enterprise customers away from competing platforms.

But traffic growth eventually collides with physical infrastructure.

As inference costs are reassessed, pricing must gradually return to levels that reflect:

  • Electricity;

  • Hardware depreciation;

  • Networking;

  • Data-center operations;

  • Service reliability;

  • Capacity reserves;

  • Commercial sustainability.

Competition is therefore shifting away from the simple question of who can publish the lowest token price.

The new contest is about:

  • Who can control inference costs;

  • Who can schedule workloads more effectively;

  • Who can maintain latency during peak periods;

  • Who can commercialize capacity without destroying margins.


Why Did the Pricing Logic Change So Quickly?

The supply-side bottleneck is more severe than many expected.

Public industry data indicates that compute demand increased several times over the past year, while supply expansion remained far slower. The gap between available capacity and potential demand has become substantial.

Power availability has also become a critical constraint.

Gartner has projected that global data-center electricity consumption could reach approximately 565 TWh in 2026. Electricity use by AI-optimized servers could rise from around 95 TWh to 175 TWh, an increase of roughly 84%.

Microsoft CEO Satya Nadella has summarized the challenge directly:

The central problem is no longer simply whether enough chips exist. It is whether electricity and supporting infrastructure can be built fast enough where those chips need to operate.

A GPU cannot generate tokens without:

  • Power;

  • Cooling;

  • Networking;

  • Memory;

  • Storage;

  • Physical rack space;

  • Redundant infrastructure.

Model efficiency continues to improve, but demand is growing at the same time.

When lower prices encourage more users to run longer contexts, more agents, and more automated workflows, the reduction in unit cost can be overwhelmed by the expansion in total usage.


Peak Pricing Is a Scheduling Tool Disguised as a Price Change

DeepSeek’s peak-and-off-peak mechanism is best understood as a compute-scheduling system.

The objective is not only to charge more during the day.

It is to move workloads that do not require immediate completion into lower-demand periods.

Tasks that can be scheduled overnight may include:

  • Batch document processing;

  • Repository indexing;

  • Large-scale translation;

  • Offline report generation;

  • Data labeling;

  • Embedding creation;

  • Non-urgent code analysis.

Meanwhile, daytime capacity can be preserved for:

  • Real-time coding;

  • Customer support;

  • Financial decision systems;

  • Interactive agents;

  • Online search;

  • Time-sensitive enterprise workflows.

Developers will therefore be encouraged to improve:

  • Prompt caching;

  • Batch scheduling;

  • Retry timing;

  • Asynchronous queues;

  • Workload prioritization.

Peak-and-off-peak pricing makes inefficient scheduling visible on the invoice.

A team that sends every task during the busiest hours may pay far more than a team that separates urgent work from deferrable work.


Time-Based Pricing Is Only the First Step

DeepSeek has applied electricity-grid logic inside its own API system.

But another proposal goes further.

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

Rather than allowing each provider to manage capacity and pricing entirely inside a closed system, Unexhub is exploring a broader marketplace where token supply can be aggregated, routed, priced, and settled.

Its proposed model centers on three mechanisms:

  1. Real-time settlement;

  2. Zero-commission incentives for early suppliers;

  3. API bidding and dynamic pricing.

Taken together, these mechanisms attempt to move tokens from provider-defined billing units toward market-priced compute resources.


Layer One: Real-Time Settlement

Under the traditional model, small and medium-sized compute suppliers often need to invest first and recover revenue later.

They may pay in advance for:

  • GPUs;

  • Cloud capacity;

  • Cluster construction;

  • Electricity and networking;

  • Model deployment;

  • API operations.

After providing the service, they still need to wait for usage reconciliation, invoicing, and payment.

The longer the settlement period, the more capital becomes trapped in accounts receivable.

This creates an uncomfortable paradox:

The more capacity a supplier provides, the greater its short-term cash-flow pressure can become.

A rapidly growing supplier may appear successful in terms of usage while simultaneously struggling to finance additional capacity.

Unexhub proposes real-time or near-real-time settlement.

Its intended logic is straightforward:

Capacity is provided, actual usage is recorded, and the corresponding revenue is settled with minimal delay.

For suppliers, faster settlement may help:

  • Reduce receivables;

  • Improve cash turnover;

  • Reinvest revenue into new capacity;

  • Respond more quickly to traffic spikes;

  • Reduce the capital cost of waiting for payment.

The platform cannot eliminate the cost of hardware, power, or operations.

However, it can attempt to reduce the financial friction between delivering compute and recovering funds.

For small suppliers, this may be the difference between being able to expand and being forced to stop accepting new traffic.


Layer Two: Zero-Commission Early Entry

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

At first glance, this may appear to be a simple promotional discount.

Its practical value is broader.

A supplier entering a new token marketplace needs to test several things:

  • Whether the technical connection is stable;

  • Whether genuine demand exists;

  • Which models customers prefer;

  • Whether pricing remains profitable;

  • How quickly revenue is settled;

  • Whether enterprise customers trust the route.

Zero commission during the early stage can lower the cost of completing this first commercial cycle.

Suppliers gain an opportunity to:

  • Integrate at lower cost;

  • Establish usage records;

  • Validate settlement;

  • Build customer trust;

  • Test unit economics;

  • Secure early market share.

The benefit is not merely avoiding one platform fee.

It is the ability to enter a developing market, prove that the workflow functions, and build distribution before the market becomes crowded.


Layer Three: API Bidding and Dynamic Pricing

The most significant part of the proposal is the API bidding mechanism.

Instead of keeping every route at a fixed price, Unexhub aims to allow prices to move according to market supply and demand.

During off-peak periods:

  • Developers may receive lower API prices;

  • Suppliers can monetize idle capacity;

  • Batch tasks can run more economically;

  • Cluster utilization improves.

During high-demand periods:

  • Prices may increase;

  • Suppliers receive incentives to release more capacity;

  • Urgent workloads can purchase priority;

  • Scarce resources move toward higher-value demand.

Dynamic pricing is already familiar in cloud computing through spot instances and priority capacity.

What changes in a token marketplace is the scope.

Instead of one provider adjusting prices inside a closed platform, multiple model and compute suppliers can potentially compete in an open distribution network.

Price discovery could reflect:

  • Real-time capacity;

  • Request volume;

  • Model quality;

  • Latency;

  • Reliability;

  • Context length;

  • Geographic route;

  • Tool-use capability;

  • SLA requirements.

The price is no longer only what one provider decides to publish.

It increasingly becomes a signal produced by the market.


Turning Tokens Into Market-Priced Compute Assets

When real-time settlement, early-entry incentives, and API bidding are combined, the underlying objective becomes clear:

Transform tokens from model-side billing units into market-priced, tradeable compute resources.

Under the traditional system, model suppliers determine:

  • Input prices;

  • Output prices;

  • Cache discounts;

  • Usage tiers;

  • Priority premiums.

Pricing is largely static and one-directional.

The supplier publishes a rate, and users decide whether to accept it.

Under a marketplace model, prices may respond continuously to:

  • Available capacity;

  • Current demand;

  • Usage time;

  • Route quality;

  • Model performance;

  • Settlement conditions;

  • Required service guarantees.

High demand pushes the price upward.

Excess capacity during quiet periods pushes it downward.

Suppliers no longer need to predict every situation and embed all risk into one fixed price. The market can recalculate the value of capacity in real time.


Tokens Are Not Identical Commodities

A token marketplace does not mean every token has the same value.

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

The price must reflect both quantity and quality.

A mature token market may eventually distinguish between:

  • Standard and priority routes;

  • Peak and off-peak capacity;

  • Cached and uncached input;

  • General-purpose and specialist models;

  • Guaranteed and best-effort throughput;

  • Domestic and international routes;

  • Low-latency and economy services.

In this sense, the token is not a standardized commodity like one unit of electricity.

It is closer to a unit of differentiated compute service whose value depends on the intelligence, speed, and reliability behind it.


The Supplier Cost Ledger Must Be Recalculated

If dynamic token markets become viable, the cost structure of model suppliers will change substantially.

Historically, API pricing followed a cost-plus formula:

  1. Calculate infrastructure and operating costs;

  2. Add a target margin;

  3. Publish a fixed price.

Under the Unexhub-style model, pricing becomes more market-driven.

The value of capacity is recalculated according to current conditions.

This changes both revenue and cost accounting.


Revenue Changes From Linear to Dynamic

Under the traditional model:

API revenue = call volume × fixed unit price.

Under a dynamic market:

API revenue = call volume × variable unit price.

Revenue depends on:

  • How much traffic arrives during peak periods;

  • Whether off-peak capacity can be sold;

  • How effectively urgent and flexible workloads are separated;

  • Whether the route can earn a premium for lower latency or stronger reliability;

  • How well the supplier participates in broader distribution.

Two suppliers with the same number of GPUs could generate very different revenue if one schedules capacity more effectively.


Competition Changes From “Who Is Cheaper?” to “Who Schedules Better?”

During the API price-war era, the headline metric was the lowest cost per million tokens.

In a dynamic market, the competitive advantage shifts toward operational efficiency.

Winning suppliers will be those that can:

  • Fill idle capacity;

  • Move batch traffic into low-demand windows;

  • Maintain stability during peaks;

  • Improve cache utilization;

  • Reduce retries and failed calls;

  • Route tasks to the most suitable models;

  • Convert price fluctuations into sustainable margins.

The question is no longer simply:

“Who has the cheapest Token?”

It becomes:

“Who can complete the required task with the best combination of cost, latency, and reliability?”


From Owning GPUs to Managing Token Economics

At WAIC 2026, enterprise customers reportedly began changing the way they evaluated AI suppliers.

The old question was:

“How many accelerator cards do you own?”

The new questions are:

  • What is your real token cost?

  • What is the cost per completed task?

  • What latency can you guarantee?

  • How does pricing change during peak periods?

  • Can workloads be shifted automatically?

  • Can traffic fall back to another model?

  • How quickly can new capacity be added?

The shift is significant.

Compute is gradually moving:

  • From capital expenditure toward operating expenditure;

  • From training-dominated demand toward inference;

  • From hardware capacity competition toward cost-efficiency competition;

  • From closed infrastructure toward routable token supply.

Owning GPUs remains important.

But the ability to convert those GPUs into stable, affordable, and efficiently scheduled token output is becoming more valuable than the hardware count alone.


Unexhub’s Position in the Emerging Token Market

According to its public positioning, Unexhub is building a global aggregation and distribution layer for AI models and token capacity.

The platform connects suppliers across different scales, including access to ecosystems such as:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • Other models offered by individual developers and small or medium-sized providers.

For developers and enterprises, the first benefit is unified access.

Instead of maintaining separate:

  • Accounts;

  • API keys;

  • SDKs;

  • Authentication systems;

  • Billing rules;

  • Error-handling logic;

users can access multiple model services through a more consistent gateway.

Unexhub also emphasizes millisecond-level routing.

Requests can be assigned according to:

  • Model capability;

  • Current price;

  • Availability;

  • Latency;

  • Workload type;

  • Regional requirements;

  • Traffic pressure;

  • Enterprise policy.

A lightweight extraction request can be sent to a lower-cost model.

A complex reasoning task can be routed to a stronger model.

A batch workload can wait for off-peak pricing.

An urgent enterprise request can use a priority route.

This allows users to pay for the capability they actually need instead of sending every task through the most expensive option.


Developer Benefits: Pay for Actual Usage and Choose the Right Route

Unexhub allows developers to pay according to actual token consumption.

Depending on the selected 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, the value extends beyond one discounted API request.

It may reduce spending on:

  • Long-context processing;

  • Coding agents;

  • Batch document analysis;

  • Automated customer service;

  • Multi-model evaluation;

  • High-frequency production workloads.

More importantly, developers may gain greater flexibility:

  • Choose lower-cost routes when latency is less important;

  • Use premium routes for urgent tasks;

  • Switch models without rebuilding the entire integration;

  • Reduce dependence on a single provider;

  • Compare price and performance more transparently;

  • Consolidate usage and billing.

A low price is useful.

A low price combined with stable routing, model flexibility, and transparent billing is far more valuable.


Supplier Benefits: One Integration, Broader Distribution

For model and compute suppliers, Unexhub offers a different value proposition.

A supplier may have:

  • A high-quality model;

  • Regional inference capacity;

  • Specialized domain performance;

  • Idle GPUs;

  • Competitive off-peak pricing.

But it may lack:

  • Enterprise sales channels;

  • International distribution;

  • Billing infrastructure;

  • Settlement systems;

  • Traffic acquisition;

  • Multi-tenant API management.

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

Real-time settlement then creates a more direct path from capacity to cash flow:

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

For smaller suppliers, this can reduce distribution and capital barriers.

For larger suppliers, it can provide an additional channel for monetizing idle or off-peak capacity.


Real-World Execution Will Determine Whether the Model Works

The direction is compelling, but a functioning token marketplace must solve difficult operational problems.

Key questions include:

  • How is model quality measured?

  • How are suppliers evaluated?

  • Who pays for failed requests?

  • How are retries settled?

  • How are latency and throughput guaranteed?

  • How are privacy and compliance enforced?

  • How are customer data and credentials protected?

  • How is price manipulation prevented?

  • Can real-time settlement remain stable under large transaction volumes?

  • How are SLAs standardized across different providers?

A marketplace cannot succeed on low prices alone.

It must also provide:

  • Reliable routing;

  • Predictable quality;

  • Secure data handling;

  • Transparent billing;

  • Stable settlement;

  • Clear accountability.

Unexhub’s real-time settlement, zero-commission onboarding, API bidding, and millisecond routing address real industry pain points.

Whether they can become foundational infrastructure will depend on performance under real production scale.


Final Thoughts: AI APIs Are Learning to Price Like the Electricity Grid

DeepSeek’s peak-and-off-peak pricing, cloud-provider adjustments, and Goldman Sachs’ assessment that the aggressive price war may be ending all send the same signal:

The era of one uniform token price is coming to an end.

AI API pricing is beginning to learn from electricity markets:

  • Charge more when capacity is scarce;

  • Offer discounts when capacity is idle;

  • Use price to move flexible demand;

  • Reserve premium resources for urgent workloads.

But Unexhub’s proposed token marketplace takes the idea one step further.

DeepSeek is applying grid-style pricing inside a provider-owned system.

Unexhub is attempting to turn that private grid into a broader market—one where multiple suppliers can compete, users can choose routes, and prices can emerge from supply and demand.

As the token market becomes more active, model suppliers will find it increasingly difficult to hide their appetite for compute.

When the appetite can no longer be hidden, it must be fed.

The traditional answer was:

“Buy more GPUs and build a larger private cluster.”

The emerging answer may become:

“Purchase capacity from the market, route it dynamically, and settle it according to actual use.”

After all, even the most powerful model still requires electricity, accelerator cards, networking, and clusters.

What is changing is not whether those resources are needed.

What is changing is how they are priced, distributed, and financed.

The next stage of competition will not belong only to the provider with the lowest API rate or the largest GPU inventory.

It may belong to the platforms that can make scarce compute circulate more efficiently—and turn every available unit of capacity into tokens that users can afford.


中文版本

大模型 API 告别“统一定价”:峰谷机制与 Token 市场正在重写算力成本账

2026 年盛夏,大模型 API 的定价逻辑正在发生根本性变化。

过去两年,行业已经习惯了一条清晰的叙事主线:

模型能力越来越强,Token 单价越来越低。

这一趋势并没有消失,但单一的降价曲线,已经不足以描述当前市场。

大模型 API 定价正在变得:

  • 分时;

  • 分层;

  • 分场景;

  • 与实时算力供给挂钩;

  • 越来越受到市场供需影响。

DeepSeek 推出的峰谷定价机制,正是这一变化最明确的信号之一。


DeepSeek 给 API 装上了“峰谷开关”

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

高峰时段的 API 价格,将提高至普通时段的两倍。

高峰窗口覆盖:

  • 北京时间 9:00—12:00;

  • 北京时间 14:00—18:00。

这几乎囊括了国内开发者和企业用户最核心的工作时段。

以 V4 Pro 为例:

  • 缓存未命中输入价格从每百万 Token 3 元上涨至 6 元;

  • 输出价格从每百万 Token 6 元上涨至 12 元。

这家曾经依靠“极致性价比”打开市场的厂商,主动给自己的 API 安装了峰谷开关。

这并不是一次普通的统一涨价。

并非所有请求都变贵,而是请求在什么时间发生、当时的集群压力有多大,开始影响最终价格。

DeepSeek 相当于将电网的定价逻辑搬进了算力市场:

需求集中时价格更高,资源空闲时价格更低。


这并不是 DeepSeek 的孤立动作

其他模型与云服务厂商也在调整价格和商业模式。

腾讯云混元系列逐步从免费公测走向商业收费,部分版本价格出现较大幅度变化。HY2.0-Instruct 输入价格据称从每千 Token 0.0008 元提高至 0.004505 元,调整幅度约为 463%。

同一时期:

  • 智谱 AI 的 API 价格环比明显上涨;

  • 阿里云、百度智能云的部分算力产品涨幅在 5%—50% 区间;

  • 多家模型厂商开始重新调整套餐、用量限制与商业服务结构。

在同一片市场中,有些模型仍在降价,有些服务却在明显涨价。

这看似矛盾,实际上揭示了行业的真正方向:

大模型 API 正从单一的降价通道,分化为按时段、模型能力、服务等级和使用场景定价的复杂体系。

可以等待到凌晨执行的批量任务,可能获得一种价格。

工作时间运行的实时代码智能体,可能对应另一种价格。

普通线路可以较便宜,但保证吞吐与优先调用需要支付溢价。

同一个模型的每个 Token,在任何时间都保持完全相同价格的时代,正在逐渐过去。


激进价格战可能正在接近极限

高盛对这一变化给出了直接判断。

7 月 3 日发布的研报指出,DeepSeek 的峰谷定价可能意味着,自 2026 年 4 月底以来持续的激进价格竞争正在接近尾声。

此前的价格战曾将行业利润大幅压缩,部分 API 甚至接近零毛利或亏损水平。

在抢占市场阶段,这种策略有其合理性。

低价能够帮助模型厂商:

  • 吸引开发者;

  • 扩大 API 流量;

  • 建立生态依赖;

  • 鼓励更多试验;

  • 争取企业用户迁移平台。

但调用量最终会与物理基础设施发生碰撞。

随着推理成本被重新评估,价格需要逐渐回到能够覆盖以下真实支出的水平:

  • 电力;

  • 硬件折旧;

  • 网络;

  • 数据中心运营;

  • 服务稳定性;

  • 冗余容量;

  • 长期商业可持续性。

竞争的重点,因此正从“谁能报出最低 Token 单价”,转向:

  • 谁能控制推理成本;

  • 谁能更高效地调度任务;

  • 谁能在高峰期维持低延迟;

  • 谁能在不牺牲利润的情况下将算力商业化。


为什么定价逻辑会如此剧烈地转向?

算力供给侧的瓶颈,比许多人预想的更加严重。

行业公开数据表明,过去一年算力需求增长数倍,而供给扩张速度明显落后,可用容量与潜在需求之间已经出现较大缺口。

电力也开始成为关键约束。

Gartner 预计,2026 年全球数据中心用电量可能达到约 565 TWh。其中,AI 优化服务器的耗电量可能从约 95 TWh 增长至 175 TWh,增幅约为 84%。

微软 CEO 萨提亚·纳德拉也直接指出:

真正的问题已不只是有没有足够的芯片,而是电力与配套设施,能否在需要部署这些芯片的地方及时建成。

一张 GPU 想要持续生成 Token,还需要:

  • 电力;

  • 散热;

  • 网络;

  • 内存与显存;

  • 存储;

  • 机架空间;

  • 容灾与冗余设施。

模型效率在持续提升,但市场需求也在同步增长。

当低价刺激用户使用更长上下文、更多智能体和更复杂的自动化工作流时,单位成本的下降,很容易被调用总量的增长所吞没。


峰谷定价,本质上是披着价格外衣的调度工具

理解 DeepSeek 峰谷机制的最佳方式,不是将其视为简单涨价,而是将其看成一套算力调度系统。

其目标不只是让白天调用更贵。

更重要的是,将不需要立即完成的任务,引导至负载更低的时间窗口。

适合安排在夜间的任务可能包括:

  • 批量文档处理;

  • 代码仓库索引;

  • 大规模翻译;

  • 离线报告生成;

  • 数据标注;

  • 向量生成;

  • 非紧急代码分析。

白天的算力则可以优先留给:

  • 实时代码智能体;

  • 在线客服;

  • 金融决策系统;

  • 交互式 Agent;

  • 在线搜索;

  • 对延迟敏感的企业工作流。

开发者将被倒逼优化:

  • 提示词缓存;

  • 批量任务调度;

  • 重试时间;

  • 异步队列;

  • 任务优先级。

峰谷定价会让低效调度直接反映在账单上。

把所有任务都集中在高峰期执行的团队,成本可能远高于能够区分紧急任务与可延后任务的团队。


分时定价只是第一步

DeepSeek 是在自有 API 体系中引入电网式定价。

但另一套方案走得更远。

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

它不再让每一家厂商完全在封闭体系中管理算力与定价,而是尝试建立更广泛的市场,让 Token 供给能够被聚合、路由、定价和结算。

Unexhub 的核心机制主要包括:

  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,试图让 Token 从模型厂商定义的计费单位,逐步转变为由市场定价的算力资源。


第一层:即时结算

在传统模式下,中小型算力供应商通常需要先投入、后回款。

他们可能需要提前支付:

  • GPU;

  • 云资源;

  • 集群建设费用;

  • 电力与网络;

  • 模型部署成本;

  • API 运维支出。

服务已经提供之后,仍要等待用量核算、开票与回款。

结算周期越长,沉淀在应收账款中的资金就越多。

由此形成一个矛盾:

供应商提供的算力越多,短期现金流压力反而可能越大。

一家调用量快速增长的供应商,业务数据看起来十分亮眼,却可能因为缺少周转资金而无法继续扩容。

Unexhub 尝试通过即时或准即时结算缩短这一周期。

其理想逻辑是:

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。

对供应商而言,更快的结算有助于:

  • 减少应收账款;

  • 加快资金周转;

  • 将收益重新投入新增算力;

  • 更及时地应对流量高峰;

  • 降低等待回款带来的资金成本。

平台无法消除硬件、电力与运维成本,却可以尝试减少“服务已经交付、资金却迟迟不能回流”的摩擦。

对中小供应商来说,这可能决定其能否继续扩容,还是不得不暂停接受新流量。


第二层:早期零佣金入驻

根据 Unexhub 的公开规划,符合条件的早期供应商可以获得零佣金激励。

表面上看,这只是一项费率优惠。

实际上,其价值更加具体。

供应商进入一个新兴 Token 市场时,需要验证:

  • 技术接入是否稳定;

  • 市场是否存在真实需求;

  • 用户更偏好哪些模型;

  • 当前价格能否盈利;

  • 回款速度是否符合预期;

  • 企业客户是否信任相关线路。

早期零佣金能够降低完成第一轮商业验证的成本。

供应商可以更低成本地:

  • 完成平台接入;

  • 建立调用记录;

  • 测试结算流程;

  • 积累客户信任;

  • 验证单位经济模型;

  • 提前获得市场份额。

其价值并不只是少支付一笔平台费用。

而是帮助供应商在市场竞争加剧前,先证明接入、供给、结算与再投入的完整流程能够运转。


第三层:API 竞价与动态定价

这套方案中最值得关注的,是 API 竞价机制。

与固定价格不同,Unexhub 希望让不同线路的价格随市场供需变化。

在低谷时段:

  • 开发者可能获得更低 API 价格;

  • 供应商能够出售闲置算力;

  • 批量任务以更低成本执行;

  • 集群利用率得到提高。

在需求高峰期:

  • 价格可以适度上浮;

  • 供应商获得释放更多算力的动力;

  • 紧急任务可以购买优先级;

  • 稀缺资源流向价值更高的需求。

动态定价在云计算行业并不陌生,竞价实例与优先容量早已存在。

Token 交易市场的不同之处,在于它可能聚合多家模型与算力供应商,而不是将定价完全锁在单一厂商的内部系统中。

价格发现可以更加直接地反映:

  • 实时可用容量;

  • 当前请求数量;

  • 模型质量;

  • 延迟;

  • 稳定性;

  • 上下文长度;

  • 地域线路;

  • 工具调用能力;

  • SLA 要求。

价格不再只是一家供应商公布的数字,而逐渐成为市场共同产生的信号。


将 Token 变成市场定价的算力资产

即时结算、早期激励与 API 竞价结合起来,目标已经十分清晰:

将 Token 从模型侧的计费单位,转变为由市场定价、可以交易的算力资源。

传统模式下,模型厂商单方面决定:

  • 输入价格;

  • 输出价格;

  • 缓存优惠;

  • 用量等级;

  • 优先调用溢价。

价格相对静态,并且是单向发布的。

供应商给出价格,用户选择是否接受。

在市场化模式中,价格可以持续根据以下条件变化:

  • 当前可用容量;

  • 实时调用需求;

  • 使用时间;

  • 路由质量;

  • 模型能力;

  • 结算条件;

  • 服务保障要求。

需求强时,价格上浮。

闲时供给过剩时,价格下探。

供应商不再需要提前预测所有场景,并将全部风险写进同一个固定价格中。市场可以持续重新计算算力容量的价值。


不同模型的 Token 并不是同质商品

Token 市场并不意味着所有 Token 都拥有相同价值。

轻量抽取模型的 100 万 Token,与前沿推理、编程或多模态模型的 100 万 Token,提供的能力明显不同。

因此,市场必须同时对数量与质量进行定价。

成熟的 Token 市场未来可能区分:

  • 标准线路与优先线路;

  • 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;

  • 国内线路与海外线路;

  • 低延迟服务与经济型服务。

Token 并不像一度电那样完全标准化。

它更像是一种差异化算力服务的计量单位,其价值取决于背后的智能能力、速度和稳定性。


模型供应商必须重新计算成本账

如果动态 Token 市场能够顺利运行,模型供应商的成本结构将发生明显变化。

过去,API 定价通常采用成本加成方式:

  1. 计算基础设施与运营成本;

  2. 加入目标利润;

  3. 发布固定价格。

Unexhub 式的市场模型,则会让价格更多由供需决定。

算力价值根据实时环境不断被重新计算,收入与成本核算都将发生改变。


收入从线性转向动态

过去的公式是:

API 收入 = 调用量 × 固定单价。

在动态市场中,公式变为:

API 收入 = 调用量 × 浮动单价。

收入将取决于:

  • 高峰时段有多少流量;

  • 低谷闲置算力能否售出;

  • 紧急任务与弹性任务能否有效分层;

  • 低延迟或高稳定性线路能否获得溢价;

  • 供应商能否进入更广泛的分发市场。

两家拥有相同 GPU 数量的供应商,可能因为调度效率不同,最终获得完全不同的收入。


竞争从“谁更便宜”转向“谁更会调度”

在 API 价格战时期,最受关注的指标是每百万 Token 的最低报价。

进入动态市场后,竞争优势将转向运营效率。

真正占优的供应商,需要能够:

  • 填满闲置算力;

  • 将批量流量转移至低谷;

  • 在高峰期维持稳定;

  • 提高缓存利用率;

  • 减少失败调用与重试;

  • 将任务路由到最合适的模型;

  • 将价格波动转化为可持续利润。

问题不再只是:

“谁的 Token 最便宜?”

而是:

“谁能用更好的成本、延迟与稳定性组合,把任务真正完成?”


企业从关心 GPU 数量,转向关心 Token 经济性

在 2026 年 WAIC 上,企业客户评估 AI 供应商的方式也开始发生变化。

过去的问题是:

“你们有多少张算力卡?”

现在的问题变成:

  • 真实 Token 成本是多少?

  • 完成一个任务要花多少钱?

  • 能保证怎样的响应延迟?

  • 高峰时段如何收费?

  • 是否能够自动调度任务?

  • 一个模型不可用时能否切换?

  • 新增算力需要多长时间?

这反映出一场更大的产业变化。

算力正在:

  • 从资本支出走向运营支出;

  • 从训练需求转向推理需求;

  • 从硬件规模竞争转向成本效率竞争;

  • 从封闭基础设施转向可路由的 Token 供给。

GPU 数量仍然重要。

但能否将 GPU 高效转化为稳定、价格可控并且调度合理的 Token,正在变得更加重要。


Unexhub 在新 Token 市场中的位置

根据公开定位,Unexhub 正在建设面向全球的 AI 模型与 Token 算力聚合分发层。

平台连接不同规模的 AI 供应商,包括:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • 个人开发者与中小企业提供的其他优质模型。

对开发者和企业用户而言,第一项价值是统一接入。

用户不必分别管理每一家模型厂商的:

  • 账户;

  • API 密钥;

  • SDK;

  • 鉴权系统;

  • 计费方式;

  • 异常处理逻辑。

通过统一网关,即可调用不同模型服务。

Unexhub 同时强调毫秒级路由,请求可以按照以下条件进行分配:

  • 模型能力;

  • 当前价格;

  • 可用性;

  • 响应延迟;

  • 任务类型;

  • 地域要求;

  • 流量压力;

  • 企业策略。

轻量信息抽取可以交给低成本模型。

复杂推理可以选择能力更强的模型。

批量任务可以等待低谷价格。

紧急企业请求则可以使用优先线路。

开发者无需让所有任务都通过最昂贵的模型,而是按照实际需要为能力付费。


开发者权益:按真实用量付费,灵活选择线路

Unexhub 支持开发者按照实际 Token 用量付费。

根据具体模型、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。

对高频调用用户而言,其价值并不只是一笔 API 折扣,还可能持续降低:

  • 长上下文处理成本;

  • 编程智能体费用;

  • 批量文档分析支出;

  • 自动化客服成本;

  • 多模型评测费用;

  • 高频生产调用预算。

更重要的是,开发者能够获得更大的灵活性:

  • 对延迟要求不高时选择低价线路;

  • 紧急任务选择优先算力;

  • 切换模型时不必重做整套集成;

  • 减少对单一厂商的依赖;

  • 更透明地比较价格与性能;

  • 统一查看用量与账单。

低价格确实重要。

但低价格如果同时配合稳定路由、模型选择和透明账单,价值会更加明显。


供应商权益:一次接入,获得更广泛的分发渠道

对模型与算力供应商来说,Unexhub 提供的是另一种价值。

一家供应商可能拥有:

  • 优质模型;

  • 区域算力;

  • 专业领域能力;

  • 闲置 GPU;

  • 有竞争力的低谷价格。

但它可能缺乏:

  • 企业销售渠道;

  • 海外分发能力;

  • 计费基础设施;

  • 结算系统;

  • 用户流量;

  • 多租户 API 管理能力。

通过一次接入,供应商有机会进入更广泛的开发者与企业市场。

配合即时结算,可以形成更直接的变现路径:

一次接入、接受平台路由,并按照真实调用及时获得收益。

对中小供应商来说,这可以降低获客与资金门槛。

对大型供应商来说,则可以增加一条分发渠道,并帮助消化低谷期的闲置算力。


能否真正成立,仍取决于实际运营

方向具有吸引力,但一个成熟的 Token 市场必须解决许多复杂问题:

  • 如何衡量模型质量?

  • 如何评价供应商?

  • 调用失败成本由谁承担?

  • 重试费用如何结算?

  • 延迟和吞吐如何保障?

  • 隐私与合规如何落实?

  • 客户数据与密钥如何保护?

  • 如何防止价格操纵?

  • 高频交易下,即时结算能否保持稳定?

  • 不同供应商的 SLA 如何统一?

一个市场不能只靠低价取得成功。

它还必须提供:

  • 稳定路由;

  • 可预测的质量;

  • 安全的数据处理;

  • 透明账单;

  • 可靠结算;

  • 清晰的责任机制。

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

最终能否成为基础设施,仍取决于其在真实生产规模下的表现。


写在最后:大模型 API 正在向电网学习定价

DeepSeek 的峰谷定价、云厂商的价格调整,以及高盛关于激进价格战可能进入尾声的判断,共同释放出一个清晰信号:

Token 统一定价的时代正在结束。

大模型 API 正开始向电力市场学习:

  • 算力紧张时提高价格;

  • 资源闲置时提供折扣;

  • 用价格引导弹性需求;

  • 将优先算力留给紧急工作负载。

但 Unexhub 的 Token 市场方案,又将这一思路向前推进了一步。

DeepSeek 是在厂商自有体系内实行电网式调度。

Unexhub 则试图将这张私有“电网”扩展成一个更开放的市场,让不同供应商参与竞争,让用户选择线路,让价格由供需共同形成。

Token 市场越活跃,模型供应商越难掩盖自身对算力的饥渴。

当这种需求无法再被隐藏,就必须寻找新的“喂养”方式。

过去的答案是:

“购买更多 GPU,建设更大的自有集群。”

未来的答案可能变成:

“从市场采购算力,动态完成路由,并按照真实调用进行结算。”

毕竟,再强的模型也需要电力、算力卡、网络与集群。

改变的并不是这些资源是否仍然必要。

而是它们将如何被定价、分发与融资。

下一阶段的竞争,不会只属于 Token 单价最低或 GPU 数量最多的厂商。

它更可能属于那些能够让稀缺算力高效流动,并将每一份可用容量转化为用户可以长期承担的 Token 服务的平台。