DeepSeek-V4 原厂直供上腾讯云 TokenHub,模型供应商在 AI 算力交易中心玩起峰谷定价

The official version of DeepSeek V4 would arrive on Tencent Cloud’s TokenHub model-service platform and agent-development platform in mid-July through a direct-from-provider supply model.

DeepSeek V4 Goes Direct to Tencent Cloud TokenHub: Model Providers Bring Peak-and-Off-Peak Pricing to AI Compute Marketplaces

Bilingual Article|English + Chinese
The complete English version appears first.
中文完整版位于文章后半部分,请继续向下阅读。

On July 3, 2026, Tencent Cloud released an announcement with implications far beyond the launch of another model endpoint.

The official version of DeepSeek V4 would arrive on Tencent Cloud’s TokenHub model-service platform and agent-development platform in mid-July through a direct-from-provider supply model.

At the same time, Tencent Cloud would synchronize DeepSeek’s newly introduced peak-and-off-peak pricing structure.

Peak periods were defined as:

  • 9:00 a.m. to 12:00 p.m.;

  • 2:00 p.m. to 6:00 p.m.;

  • Beijing time, every day.

This was not simply another model being added to a cloud marketplace.

It represented two important shifts:

  1. A leading Chinese foundation model was entering a major cloud platform through direct official supply rather than conventional channel resale;

  2. The time-of-use pricing logic long associated with electricity markets was being written directly into AI API billing.

The AI industry is beginning to treat Tokens not merely as usage units, but as resources that can be scheduled, priced, traded, and settled according to real-time demand.


From Channel Resale to Direct Provider Supply

The phrase “direct from the original provider” deserves closer attention.

Under the traditional channel model, a cloud platform acts as an intermediary:

  1. It obtains access to a model API;

  2. Repackages the service;

  3. Adds its own commercial layer;

  4. Resells access to developers and enterprises.

This structure can add another pricing layer, another settlement process, and another commercial relationship between the model provider and the end user.

When the DeepSeek V4 preview launched in April, Tencent Cloud TokenHub offered day-zero access at the same headline price as the official service.

But the relationship was still closer to distribution.

Direct supply changes the structure.

DeepSeek provides the model service itself, while Tencent Cloud increasingly acts as the infrastructure and marketplace layer, offering:

  • Unified access;

  • Traffic scheduling;

  • Usage metering;

  • Billing;

  • Governance;

  • Monitoring;

  • Enterprise delivery.

The platform provides the shelf, traffic, transaction system, and operational infrastructure.

The model provider retains a more direct connection to the product and pricing.

This is similar to the difference between a product sold through a general reseller and an official flagship store inside a large marketplace.


TokenHub Is Becoming a Multi-Model Supermarket

Tencent Cloud TokenHub reportedly processes more than five trillion Tokens per day and aggregates major models such as:

  • Tencent Hunyuan;

  • DeepSeek;

  • MiniMax;

  • Kimi;

  • GLM;

  • Other general-purpose and specialized models.

Its role is therefore no longer limited to exposing a collection of APIs.

It is increasingly becoming a multi-model marketplace where enterprises and developers can:

  • Access several models through one platform;

  • Compare pricing and performance;

  • Switch models according to the workload;

  • Use unified billing;

  • Apply common governance controls;

  • Manage traffic more flexibly.

For model providers, entering this type of marketplace can significantly reduce the burden of building every commercial layer independently.

They do not need to create a separate distribution network, billing system, enterprise-sales organization, and customer-acquisition channel before reaching large numbers of developers.

One platform can provide access to an existing user base.

That is the immediate value of direct supply:

The model provider focuses on building and serving the model, while the marketplace handles more of the distribution infrastructure.


Peak-and-Off-Peak Pricing: Time-of-Use Electricity Rates for AI

The second major change is the introduction of time-based pricing.

DeepSeek is not the first model provider to raise or adjust API prices.

Since the beginning of 2026, Alibaba Cloud, Tencent Cloud, Baidu AI Cloud, Zhipu AI, and other major providers have revised prices across model APIs and AI computing services.

What makes DeepSeek different is that it has turned time-of-use pricing into a formal API strategy.

Under the announced structure, DeepSeek V4 pricing changes according to the time of day.

For V4 Pro:

Usage typeOff-peak ratePeak rate
Cached inputRMB 0.025 / 1M TokensRMB 0.05 / 1M Tokens
Reasoning inputRMB 3 / 1M TokensRMB 6 / 1M Tokens
Reasoning outputRMB 6 / 1M TokensRMB 12 / 1M Tokens

V4 Flash follows a similar proportional structure.

The system contains approximately seven peak hours per day, while the remaining seventeen hours use the lower rate.

Some users may describe this simply as a price increase.

A more precise interpretation is:

AI API pricing is becoming more granular.


Why Time-Based Pricing Is More Than a Price Increase

The logic is similar to electricity pricing.

Power demand is not constant throughout the day.

During peak periods, resources become scarce and expensive.

During quieter periods, infrastructure is underused.

AI inference has the same problem.

At working-hour peaks, demand from services such as the following may surge simultaneously:

  • Financial analysis;

  • Coding agents;

  • Customer-service systems;

  • Enterprise copilots;

  • Real-time search;

  • Interactive AI applications.

Meanwhile, many batch and offline workloads do not need to run at a specific time:

  • Large-scale document processing;

  • Content classification;

  • Data cleaning;

  • Test generation;

  • Embedding computation;

  • Non-urgent report generation.

Time-based pricing uses economic signals to move flexible workloads away from congested periods.

Developers with urgent, real-time tasks can continue using peak capacity.

Teams with non-urgent work can schedule it during lower-cost hours.

The result can be:

  • Better peak-period service stability;

  • Higher off-peak utilization;

  • More predictable resource allocation;

  • Lower costs for flexible workloads;

  • Less pressure to build excess idle capacity.

DeepSeek has previously experienced service instability during periods of heavy demand.

The problem was not necessarily model intelligence.

It was that peak inference demand exceeded available capacity.

Peak-and-off-peak pricing tells users:

Use premium periods for work that must happen now. Move flexible batch work to periods when capacity is cheaper and more abundant.

This is not merely a pricing tactic.

It is a scheduling mechanism.


Why Model Providers Are Entering AI Compute Marketplaces

When direct supply and time-based pricing are viewed together, the broader trend becomes clearer.

Model providers face a structural contradiction:

  • Training is expensive;

  • Inference is expensive;

  • Peak demand overwhelms available capacity;

  • Off-peak infrastructure remains idle.

Many intelligent-computing centers in China reportedly operate at average cabinet-utilization rates of only 20% to 30%.

Some enterprise centers may fall below 10%.

This means the market can experience two problems at the same time:

  • Developers wait for capacity during peak periods;

  • Expensive GPUs sit idle during low-demand periods.

The industry does not simply lack computing resources.

It lacks sufficiently efficient distribution, scheduling, and pricing mechanisms.

AI compute marketplaces offer one possible solution.


Marketplace Growth Is Accelerating

Tencent Cloud TokenHub reportedly maintained rapid growth during its first three months, with monthly usage continuing to expand.

China’s national computing platform has connected 15 provincial and municipal sub-platforms and attracted more than 130 computing-service providers.

Suning’s Zhixin Token has gathered models such as:

  • Qwen;

  • DeepSeek;

  • Kimi;

  • GLM;

  • Doubao;

  • Seedance.

Across the industry, these platforms are increasingly offering a common set of capabilities:

  • Unified model access;

  • Elastic resource scheduling;

  • Usage-based billing;

  • Automated governance;

  • Centralized settlement;

  • Multi-provider routing.

Model providers are not entering these marketplaces because of industry sentiment.

They are doing so because independently carrying every cost has become increasingly difficult.

Building a model is expensive.

Building the complete sales, distribution, billing, support, and settlement infrastructure around that model is expensive as well.

A marketplace allows those costs to be shared.


The Survival Economics of Platform Distribution

For model providers, joining a Token marketplace can offer several direct advantages.

Existing distribution

The supplier can reach developers and enterprise customers already using the platform.

Unified metering and billing

The platform can measure Token consumption, generate bills, and manage collections.

Elastic capacity

Requests can be distributed across available resources instead of depending entirely on one fixed cluster.

Lower customer-acquisition costs

The supplier does not need to independently build every sales and marketing channel.

Better utilization

Idle capacity may be matched with workloads from other regions, industries, or time periods.

For smaller model companies, these advantages are not optional conveniences.

They may be critical to survival.


Unexhub Goes Beyond Fixed Peak-and-Off-Peak Pricing

If Tencent Cloud TokenHub and China’s national computing platform are building marketplace and infrastructure layers, Unexhub is attempting to redesign the broader global distribution logic for Tokens.

Unexhub is building a computing-power trading center focused on Token aggregation and distribution.

Its purpose is to match fragmented AI supply with demand from:

  • Developers;

  • Enterprises;

  • AI applications;

  • Agent platforms;

  • High-volume inference businesses.

The platform brings together suppliers of different scales, including channels for major models such as:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • Other domestic, international, and specialized models.

It also provides an entry point for high-quality models and computing resources supplied by startups, independent developers, and small or medium-sized companies.


Faster Settlement Helps Smaller Suppliers Protect Cash Flow

For smaller providers, one of the most serious problems is not demand.

It is how long they must wait to receive payment.

Capacity may be supplied and consumed immediately, while the corresponding revenue remains in a lengthy billing and settlement process.

During that time, the supplier must continue paying for:

  • GPUs;

  • Cloud infrastructure;

  • Electricity;

  • Bandwidth;

  • Operations;

  • Engineering teams.

A profitable service can still fail if cash returns too slowly.

Unexhub’s financial clearing and settlement system is intended to shorten that cycle:

When computing capacity is consumed, the corresponding supplier revenue can be settled more quickly.

This can provide suppliers with:

  • Lower working-capital pressure;

  • More predictable cash flow;

  • Faster recovery of infrastructure investment;

  • Greater ability to maintain capacity;

  • More room to expand supply.

Eligible early suppliers may also receive promotional platform commission rates as low as 0%.

The value is not simply “zero commission.”

It can help early participants:

  • Keep more of their initial usage revenue;

  • Reduce distribution costs;

  • Reach enterprise demand sooner;

  • Test market pricing with lower risk;

  • Reinvest more capital into models and infrastructure.


API Bidding Moves From Vendor Pricing to Market Pricing

Unexhub’s planned API-bidding mechanism goes beyond a provider-defined peak-and-off-peak schedule.

DeepSeek’s model can be summarized as:

The provider defines when the peak occurs and what the price will be.

Unexhub’s proposed model is closer to:

Suppliers and buyers interact through market supply and demand, allowing prices to change dynamically.

During off-peak periods, developers may access unused resources at lower prices.

During periods of high demand, suppliers may earn higher returns for providing scarce capacity.

Potential benefits for developers

  • Lower prices for flexible workloads;

  • More suppliers to choose from;

  • Reduced dependence on one model provider;

  • Better control of Token budgets;

  • More opportunities to route tasks by price and latency.

Potential benefits for suppliers

  • Higher utilization of idle resources;

  • Broader access to demand;

  • Higher returns during demand peaks;

  • Faster settlement;

  • Lower distribution costs.

This is more than a discount mechanism.

It attempts to return part of Token pricing to the market.


Two Different Pricing Philosophies

DeepSeek’s peak-and-off-peak model and Unexhub’s API bidding represent two stages of pricing evolution.

Provider-defined time-based pricing

The supplier defines:

  • Peak hours;

  • Off-peak hours;

  • Price differences;

  • Applicable models.

This improves scheduling while preserving centralized pricing authority.

Market-driven dynamic pricing

Multiple suppliers compete according to:

  • Available capacity;

  • Model performance;

  • Response latency;

  • Region;

  • Reliability;

  • Current demand;

  • Price.

The platform facilitates price discovery instead of defining one fixed answer.

The difference is subtle but important:

  • In the first model, the provider uses prices to manage demand;

  • In the second, the market continuously discovers the value of available capacity.


Millisecond-Level Routing Turns Price Into an Execution Decision

Dynamic pricing only becomes useful when the platform can route requests efficiently.

Unexhub emphasizes millisecond-level routing based on factors such as:

  • Model capability;

  • Price;

  • Latency;

  • Regional availability;

  • Current capacity;

  • Task type;

  • Service quality.

A simple classification task should not automatically use the same expensive resource as a complex reasoning request.

A non-urgent batch workload can be routed to a lower-cost off-peak channel.

A latency-sensitive agent can be sent to a faster premium route.

A high-risk result can be passed to a second model for independent verification.

A practical routing structure could look like this:

Task typeSuitable resource
Classification and extractionLightweight, low-cost model
Batch content processingOff-peak discounted capacity
General business tasksBalanced model
Coding and tool orchestrationCoding-oriented model
Complex reasoningPremium model
High-risk outputMulti-model verification

This transforms Token pricing from a passive bill into an active resource-allocation signal.


Pay Only for Actual Token Usage

Unexhub also supports usage-based payment according to actual Token consumption.

Developers and smaller enterprises do not need to pre-purchase large fixed packages from every model provider.

They can:

  • Begin with a smaller budget;

  • Test different models;

  • Scale according to demand;

  • Avoid unused prepaid capacity;

  • Switch channels more easily;

  • Match spending to actual business volume.

Under selected supply conditions and promotional resource plans, effective prices may be as low as approximately 30% of standard market rates.

The value of this benefit increases with scale.

For high-volume users, the same budget may support:

  • More customer requests;

  • More agent workflows;

  • Higher concurrency;

  • More model experiments;

  • Additional retries;

  • More independent verification.

The saving is not merely a few cents on one API call.

It can extend the operating runway of an entire AI product.


Direct Supply, Time-Based Pricing, and Market Bidding Are Part of the Same Shift

DeepSeek V4’s direct supply to Tencent Cloud TokenHub, its peak-and-off-peak pricing, and Unexhub’s proposed API-bidding model may appear to be separate developments.

They are actually part of the same transformation.

Direct supply changes distribution

The model provider reaches users through a marketplace without surrendering the product entirely to a reseller.

Peak-and-off-peak pricing changes scheduling

Prices encourage flexible workloads to move away from congested periods.

Market bidding changes price discovery

Multiple suppliers and buyers can influence the real-time value of Token capacity.

Together, these mechanisms turn Tokens from a simple usage unit into a resource that can be:

  • Distributed;

  • Scheduled;

  • Priced;

  • Routed;

  • Traded;

  • Settled.


Challenges Still Remain

A more market-driven Token economy also introduces new risks.

Dynamic pricing can reduce costs, but may create volatility.

More suppliers can improve capacity, but may introduce inconsistent service quality.

Fast settlement can support smaller providers, but requires strong auditing and risk controls.

A mature AI compute marketplace must address:

  • Supplier verification;

  • Model-quality measurement;

  • Price transparency;

  • Capacity authenticity;

  • Data security;

  • Cross-border compliance;

  • Service-level guarantees;

  • Billing disputes;

  • Failure recovery;

  • Settlement risk.

A Token marketplace cannot succeed by offering cheap APIs alone.

It must also offer trust.


The Bottom Line

DeepSeek V4’s direct launch on Tencent Cloud TokenHub and the introduction of peak-and-off-peak pricing are not isolated commercial announcements.

They are a snapshot of how model providers are searching for sustainable business models inside a new AI compute marketplace.

TokenHub provides the shelf and infrastructure.

Time-based pricing provides a scheduling signal.

Unexhub is attempting to build a deeper trading, routing, and price-discovery mechanism.

Together, these developments point in the same direction:

Tokens are evolving from simple usage units into products that can be priced, scheduled, routed, traded, and settled.

Model providers are entering compute marketplaces because independent distribution is expensive, demand is volatile, infrastructure is underutilized, and cash-flow pressure is real.

The companies that solve the complete loop—from computing supply and model access to Token routing, dynamic pricing, billing, and settlement—will have a stronger chance of surviving the next stage of the market.

Benchmarks are shown to the industry.

Infrastructure bills are paid by the provider.

The real winners will be those who can turn model capability into a liquid, sustainable, and continuously operating market.


中文版

DeepSeek V4 原厂直供腾讯云 TokenHub:模型供应商在 AI 算力交易中心玩起峰谷定价

2026 年 7 月 3 日,腾讯云发布了一则看似简单、实则指向行业深层变化的公告。

DeepSeek V4 正式版将于 7 月中旬以**“原厂直供”**模式,登陆腾讯云大模型服务平台 TokenHub 与智能体开发平台。

伴随模型正式上线,腾讯云还将同步 DeepSeek 官网的峰谷定价机制。

每日高峰时段为:

  • 9:00—12:00;

  • 14:00—18:00;

  • 以上均为北京时间。

这并不是一次普通的模型上架。

它至少代表了两项重要变化:

  1. 国产头部大模型开始以“原厂直供”方式进入主流云平台,而不仅是通过传统渠道代销;

  2. 电力行业使用多年的分时计价逻辑,开始被正式写进 AI API 的计费体系。

Token 不再只是记录模型调用量的单位。

它正在变成一种可以被调度、定价、交易和结算的资源。


从“渠道分销”升级为“原厂直供”

“原厂直供”这四个字,值得单独拆开来看。

在传统渠道模式中,云平台主要扮演中间商:

  1. 采购或接入模型 API;

  2. 将服务重新包装;

  3. 增加一层商业渠道;

  4. 再将能力销售给企业与开发者。

这种模式通常意味着,模型供应商与最终用户之间,还隔着一层定价、结算与商业关系。

4 月 DeepSeek V4 预览版发布时,腾讯云 TokenHub 曾实现 Day 0 同步上线,价格与官网保持一致。

但当时的模式,本质上仍更加接近渠道代销。

“原厂直供”则改变了关系。

模型服务由 DeepSeek 直接提供,腾讯云的角色逐步从传统中介,转向平台与基础设施服务商,主要负责:

  • 统一接入;

  • 流量调度;

  • 用量计量;

  • 账单管理;

  • 安全治理;

  • 监控运维;

  • 企业服务交付。

平台提供货架、流量、交易系统与基础设施。

模型供应商则与产品和定价保持更加直接的关系。

这就像普通代销柜台升级成了官方旗舰店。


TokenHub 正在成为多模型超市

腾讯云 TokenHub 日均 Token 消耗量据称已经突破 5 万亿,并汇聚了多款主流模型,包括:

  • 腾讯混元;

  • DeepSeek;

  • MiniMax;

  • Kimi;

  • GLM;

  • 其他通用与垂直模型。

它的角色已经不再只是展示一组 API。

而是逐渐成为一个多模型超市,让企业与开发者可以:

  • 通过一个平台调用多款模型;

  • 比较能力与价格;

  • 根据任务切换模型;

  • 使用统一账单;

  • 应用统一治理规则;

  • 更灵活地分配调用流量。

对模型供应商而言,进入这类平台可以显著减少重复建设商业体系的压力。

它们不必在触达大量开发者之前,先独立建设完整的销售渠道、计费平台、企业服务团队和客户获取体系。

一个平台,就可以连接已有用户群体。

这正是原厂直供的直接价值:

模型供应商专注做好模型,平台承担更多分发基础设施。


峰谷定价:AI 算力行业的“分时电价”

第二项更具行业意义的变化,是分时定价。

DeepSeek 并不是 2026 年第一个调整价格的模型供应商。

阿里云、腾讯云、百度智能云、智谱 AI 等头部厂商,均对大模型 API 和 AI 算力服务进行了不同程度的价格调整。

DeepSeek 的不同之处在于,它将分时计价正式写入 API 价格策略。

按照公布的方案,DeepSeek V4 会根据调用时间采用不同费率。

以 V4 Pro 为例:

使用类型平峰价格高峰价格
缓存命中输入0.025 元/百万 Token0.05 元/百万 Token
推理输入3 元/百万 Token6 元/百万 Token
推理输出6 元/百万 Token12 元/百万 Token

V4 Flash 也会按照类似比例执行峰谷价格。

每日高峰时段合计约 7 小时,剩余 17 小时执行较低费率。

有人将其简单理解为涨价。

更加准确的说法是:

AI API 正在进入精细化定价阶段。


峰谷定价,不只是价格翻倍

这种逻辑与分时电价非常相似。

电力需求并不是全天恒定。

在高峰期,资源更加稀缺,成本更高。

在低谷期,大量基础设施又可能处于闲置状态。

AI 推理同样如此。

工作时段内,以下业务可能同时涌入:

  • 金融分析;

  • 编码智能体;

  • 客服系统;

  • 企业 Copilot;

  • 实时搜索;

  • 交互式 AI 应用。

但许多批量和离线任务,并不要求在特定时间完成:

  • 大规模文档处理;

  • 内容分类;

  • 数据清洗;

  • 测试生成;

  • 向量计算;

  • 非紧急报告制作。

峰谷定价通过价格信号,把可以延后执行的任务从高峰引导至低谷。

实时、高优先级任务继续使用高峰容量。

非紧急任务则可安排在价格更低、算力更充足的时段。

这样可能带来:

  • 提升高峰期服务稳定性;

  • 提高闲时算力利用率;

  • 改善资源分配;

  • 降低弹性任务的成本;

  • 减少为了高峰需求长期保留大量闲置算力的压力。

DeepSeek 此前曾在调用高峰出现服务波动。

问题未必出在模型本身。

更可能是高峰期推理需求超过了可用算力。

峰谷价格向开发者传递的信号是:

必须即时完成的任务使用高峰资源,可以延后的批量工作尽量移至低价时段。

因此,它不只是一个价格方案。

也是一套调度工具。


模型供应商为什么扎堆 AI 算力交易中心?

将“原厂直供”与“峰谷定价”放在一起,行业图景就更加清晰了。

模型供应商正在面对一个结构性矛盾:

  • 训练成本高;

  • 推理成本高;

  • 高峰期算力不足;

  • 低谷期资源闲置。

国内大量智算中心的机柜平均利用率,据称只有 20% 至 30%。

部分企业级智算中心甚至可能低于 10%。

这意味着市场中同时存在两种现象:

  • 高峰时开发者排队等待算力;

  • 低谷时昂贵 GPU 在机房中闲置。

行业并非单纯缺少算力。

而是缺少更高效的分发、调度和定价体系。

AI 算力交易中心由此成为一种解决方案。


算力交易平台正在快速增长

腾讯云 TokenHub 上线后的前几个月保持较快增长,调用规模持续扩大。

中国算力平台已连接 15 个省市级子平台,并吸引超过 130 家算力服务商入驻。

苏宁“智新 Token”则集中上架了:

  • 通义千问;

  • DeepSeek;

  • Kimi;

  • GLM;

  • 豆包;

  • Seedance。

这些平台逐渐形成了一套类似的能力体系:

  • 统一模型接入;

  • 弹性算力调度;

  • 按量计费;

  • 自动治理;

  • 集中结算;

  • 多供应商路由。

模型供应商进入这些平台,并不是因为行业情怀。

而是因为独自承担所有成本,正变得越来越困难。

研发模型很贵。

围绕模型独立建设销售、分发、计费、客服和结算体系,同样昂贵。

平台模式让这些成本可以被共享。


平台分发背后的求生经济学

对模型供应商而言,进入 Token 市场可以获得几项直接权益。

现成分销渠道

供应商可以触达已经在平台上的开发者和企业客户。

统一计量与计费

平台负责统计 Token 消耗、生成账单并完成资金收取。

弹性算力

请求可以根据可用资源动态调度,而不必完全依赖固定集群。

降低获客成本

供应商不需要独立建设所有销售与营销渠道。

提高利用率

闲置资源可以匹配其他行业、地区或时间段的工作负载。

对中小模型公司来说,这些并不是锦上添花。

而可能是生存所需。


Unexhub:比固定峰谷价格再往前一步

如果说腾讯云 TokenHub 与中国算力平台正在搭建卖场和基础设施,那么 Unexhub 正尝试进一步重构全球 Token 的分发逻辑。

Unexhub 建设的是面向 Token 聚合与分发的算力交易中心。

平台希望把分散的 AI 供给,与以下需求进行匹配:

  • 开发者;

  • 企业;

  • AI 应用;

  • 智能体平台;

  • 高调用量推理业务。

平台聚合不同规模的供应商,并提供包括以下模型在内的接入渠道:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • 其他国内外通用及专业模型。

初创企业、独立开发者与中小供应商提供的优质模型和算力,也能够借助平台进入更广泛的市场。


更快结算,为中小供应商保护现金流

对中小供应商而言,最危险的问题往往不是没有需求。

而是回款太慢。

算力今天已经提供并产生调用,对应收益却可能长期停留在账单和结算流程中。

与此同时,供应商仍需要持续支付:

  • GPU 成本;

  • 云端基础设施;

  • 电力;

  • 带宽;

  • 运维;

  • 工程团队开支。

即使业务账面盈利,只要现金回流太慢,也可能无法持续。

Unexhub 的金融结算清算体系,目标是缩短这一周期:

算力被使用后,对应收入可以更快回到供应端。

这可能为供应商带来:

  • 缓解周转资金压力;

  • 提高现金流可预测性;

  • 更快回收基础设施投入;

  • 保持稳定供给;

  • 为扩容留下空间。

符合条件的早期供应商,在推广阶段还可享受最低 0 佣金的平台费率。

这项权益的价值并不只是“零佣金”。

它可以帮助早期参与者:

  • 保留更多初期调用收益;

  • 降低渠道成本;

  • 更快触达企业需求;

  • 以更低风险测试市场价格;

  • 将更多资金重新投入模型和算力。


API 竞价:从“厂商定价”走向“市场定价”

Unexhub 计划中的 API 竞价机制,比由厂商规定的峰谷时段再向前一步。

DeepSeek 的峰谷定价可以概括为:

厂商规定高峰是什么时候,以及高峰应该收多少钱。

Unexhub 的竞价思路则更接近:

多个供应商和用户根据供需关系参与市场,让价格动态变化。

在低谷期,开发者可以用更低价格获得闲置资源。

在需求高峰,供应商则可以因为提供稀缺容量获得更高收益。

对开发者而言

  • 弹性任务可获得更低价格;

  • 拥有更多供应商选择;

  • 减少对单一模型厂商的依赖;

  • 更容易控制 Token 预算;

  • 可综合价格与延迟进行路由。

对供应商而言

  • 提升闲置资源利用率;

  • 连接更广泛的需求;

  • 高峰期获得更高回报;

  • 更快完成收益结算;

  • 降低分发成本。

这不只是优惠机制。

而是尝试将部分 Token 定价权交还市场。


两种不同的定价哲学

DeepSeek 的峰谷模式与 Unexhub 的 API 竞价,代表了定价机制演进的两个阶段。

厂商主导的分时定价

供应商决定:

  • 哪些时段是高峰;

  • 哪些时段是低谷;

  • 价格相差多少;

  • 哪些模型适用。

它能改善资源调度,但定价权仍集中在供应商手中。

市场驱动的动态竞价

多家供应商根据以下因素参与竞争:

  • 可用算力;

  • 模型能力;

  • 响应延迟;

  • 地区;

  • 稳定性;

  • 当前需求;

  • 价格。

平台负责促进价格发现,而不是只给出一个固定答案。

两种方式的差异在于:

  • 前者由供应商用价格管理需求;

  • 后者由市场持续发现算力的真实价值。


毫秒级路由,让价格真正影响执行

动态价格只有与高效路由结合,才有实际意义。

Unexhub 强调根据以下因素进行毫秒级路由:

  • 模型能力;

  • 价格;

  • 延迟;

  • 地区可用性;

  • 当前容量;

  • 任务类型;

  • 服务质量。

简单分类任务,不必自动调用与复杂推理相同的高端资源。

非紧急批量工作,可以被发送至闲时低价通道。

对延迟敏感的智能体,则可以选择速度更快的优先线路。

高风险结果,还可以转交第二款模型独立复核。

一个合理的路由结构可以是:

任务类型更适合的资源
分类与提取轻量低成本模型
批量内容处理闲时折扣算力
通用企业任务均衡型模型
编程与工具编排代码型模型
复杂推理高端模型
高风险输出多模型复核

这样一来,Token 价格就不再只是被动出现在账单上。

而会成为主动分配资源的信号。


按实际 Token 用量付费

Unexhub 还支持按照实际 Token 消耗付费。

开发者与中小企业无需分别向每家模型供应商预购大额固定套餐。

它们可以:

  • 用更小预算开始;

  • 同时测试多款模型;

  • 根据需求逐步扩容;

  • 避免预付资源浪费;

  • 更方便地切换通道;

  • 让支出与真实业务规模匹配。

在部分供应条件和推广资源方案下,实际价格最低可达到市场标准价格的约 30%。

使用规模越大,这项权益的实际价值越明显。

同样预算可能支持:

  • 更多客户请求;

  • 更多智能体流程;

  • 更高业务并发;

  • 更多模型实验;

  • 更多失败重试;

  • 更多独立验证。

它节省的并不只是单次调用的几分钱。

而可能延长整个 AI 产品的运营周期。


原厂直供、峰谷定价与市场竞价,是同一场变化

DeepSeek V4 原厂直供腾讯云 TokenHub、同步采用峰谷定价,以及 Unexhub 计划中的 API 竞价,看起来是三件不同的事情。

实际上,它们属于同一场行业转型。

原厂直供改变分发方式

模型供应商通过平台直接连接用户,不再完全依赖传统渠道转售。

峰谷定价改变调度方式

价格信号推动弹性任务避开拥堵时段。

市场竞价改变价格发现方式

多个供应商和用户共同影响 Token 算力的实时价值。

这些机制正在把 Token 从简单的使用单位,转化为一种可以被:

  • 分发;

  • 调度;

  • 定价;

  • 路由;

  • 交易;

  • 结算的商品。


挑战依然存在

更市场化的 Token 经济,也会带来新的风险。

动态价格可以降低成本,也可能增加波动。

供应商增多可以提升算力可用性,也可能带来服务质量不一致。

快速结算能够帮助中小供应商,却也需要成熟的审计与风控体系。

真正成熟的 AI 算力交易中心,需要解决:

  • 供应商认证;

  • 模型质量评估;

  • 价格透明度;

  • 算力真实性;

  • 数据安全;

  • 跨境合规;

  • 服务等级保障;

  • 账单争议;

  • 故障恢复;

  • 结算风险。

Token 市场不能只靠低价 API 获胜。

它还必须建立信任。


写在最后

DeepSeek V4 原厂直供腾讯云 TokenHub,并引入峰谷定价,并不是孤立的商业动作。

它是模型供应商在 AI 算力交易中心这个新市场中,探索可持续商业模式的一个缩影。

TokenHub 提供货架与基础设施。

峰谷定价提供资源调度信号。

Unexhub 则试图建立更底层的交易、路由和价格发现机制。

三条线索指向同一个方向:

Token 正从简单的调用计量单位,转变为可以定价、调度、路由、交易和结算的生产资源。

模型供应商扎堆进入算力交易中心,是因为独立分发成本高、需求波动明显、基础设施利用率不足,现金流压力也真实存在。

谁能够率先打通从算力供给、模型接入到 Token 路由、动态定价、账单管理与收益结算的完整闭环,谁就更有机会活到下一阶段。

跑分是给行业看的。

基础设施账单,却要由供应商自己承担。

真正的赢家,将是那些能够把模型能力转化为流动、可持续并且持续运转的市场体系的参与者。