How Should a Compute Trading Center Price AI Workloads? GPU-Hours, Core-Hours, or Tokens?
Bilingual Article
The compute market is no longer priced by one simple metric.
On July 8, 2026, the China Academy of Information and Communications Technology and the Computing Power Industry Development Alliance launched the “Compute Supermarket + Compute Bank Ecosystem Co-Building Plan.” The plan aims to explore two new service models: the compute supermarket and the compute bank, with a focus on standardized pricing, online trading, automated delivery, credit-based usage, and more inclusive access for small and midsize enterprises.
Earlier in April 2026, China’s Ministry of Industry and Information Technology had already proposed innovative models such as “compute bank” and “compute supermarket,” while encouraging flexible billing methods including GPU-hour, CPU-core-hour, and Token-based pricing.
Put simply, enterprises now have three pricing choices:
GPU-hour, core-hour, or token.
But which one should a company choose?
The answer depends on the workload.
1. GPU-Hour: Best for Training and Dedicated GPU Workloads
GPU-hour, often described in Chinese as “card-hour”, is the most traditional pricing model for AI compute.
You rent a GPU card for a certain period of time. Whether you use 10% or 100% of the GPU, you pay for the time reserved.
This pricing model is easy to understand:
one GPU, one hour, one price.
It works well for workloads that require long, continuous, dedicated GPU usage, such as:
model training;
fine-tuning;
large-scale inference clusters;
batch jobs that fully occupy GPU resources;
research workloads requiring stable hardware allocation.
The advantage is clarity.
A supplier can list an A100, H800, or other GPU resource by the hour. Buyers know exactly what they are renting.
The weakness is rigidity.
If a company only needs occasional inference or lightweight API calls, GPU-hour pricing can become wasteful. The enterprise pays for reserved capacity even when the GPU is underused.
For training, GPU-hour is often reasonable.
For normal model API usage, it may be like hiring a truck to deliver one cup of coffee.
2. Core-Hour: Useful for General Computing, but Not Always AI-Native
Core-hour pricing comes from the cloud computing era.
It measures how many CPU cores are used and for how long.
This model still has value, especially for non-GPU or CPU-heavy workloads:
data cleaning;
format conversion;
ETL pipelines;
lightweight preprocessing;
feature extraction;
task orchestration;
auxiliary compute around model workflows.
For many AI projects, not every step needs a GPU. Before model inference or training, teams often spend time preparing data, converting files, deduplicating records, cleaning labels, or running lightweight scripts.
In these cases, core-hour pricing can be cheaper than GPU-hour pricing.
But core-hour has a deeper limitation:
it measures compute consumption, not intelligent output.
A simple text classification task and a complex reasoning request may consume similar CPU time in a surrounding workflow, but their business value can be completely different.
That is why core-hour works well for general computing, but less well as the main pricing unit for AI model services.
3. Token: The AI-Native Pricing Unit
Token pricing is the youngest of the three models, but it has the most imagination.
A People’s Daily report described the compute supermarket as a marketplace where users can buy compute products online with flexible billing methods such as GPU-hour, core-hour, and Token-based pricing.
Guangzhou also launched what local authorities described as the country’s first city-level integrated compute operation platform based on Token-level scheduling. The platform uses Tokens as a unified measurement benchmark and supports multiple flexible billing methods, including usage-based and cycle-based billing.
Token pricing changes the logic of AI spending.
Instead of asking:
How many GPUs did I rent?
The enterprise asks:
How much model intelligence did I actually consume?
That is why Token pricing is especially suitable for:
large model API calls;
chatbots;
RAG applications;
customer support automation;
text generation;
summarization;
classification;
agent workflows;
multimodal model calls.
For small and midsize enterprises, Token pricing lowers the entry barrier.
They do not need to buy GPUs.
They do not need to reserve instances.
They do not need to manage infrastructure.
They only pay for actual intelligent usage.
4. The Real Choice: Match Pricing to the Workload
The three models are not enemies.
They are tools for different scenarios.
| Pricing Model | Best For | Main Advantage | Main Risk |
|---|---|---|---|
| GPU-Hour | Training, fine-tuning, dedicated GPU workloads | Clear resource allocation | Wasteful if utilization is low |
| Core-Hour | Data preprocessing, general computing, lightweight tasks | Cheaper for non-GPU workloads | Does not measure AI output value |
| Token | Inference, API calls, RAG, agents, content generation | Pay-per-use and AI-native | Token counting differs by vendor |
A practical rule is simple:
Use GPU-hour for training.
Use core-hour for preprocessing.
Use Token for inference and API calls.
The expensive mistake is mixing them incorrectly.
Using GPU-hour pricing for lightweight inference can waste budget.
Using token pricing for long-running training makes no sense.
Using core-hour pricing to evaluate AI output may hide the real business value.
The smarter enterprise does not ask which pricing model is “best.”
It asks:
Which pricing model fits this exact workload?
5. The Hidden Cost of Token Pricing
Token pricing is powerful, but it is not automatically transparent.
Different model providers may tokenize the same text differently. The same paragraph may become different token counts across different models or tokenizers.
That means enterprises should not only compare the price per million Tokens.
They should compare:
how many Tokens the same workload consumes;
how input and output are priced;
whether cached input is discounted;
how retries are billed;
whether tool calls add extra cost;
whether long-context requests change pricing;
whether routing affects final cost.
In other words:
The cheapest Token price is not always the cheapest task price.
A good compute trading center should make these differences visible.
That is where platforms like Unexhub become important.
6. The Role of a Compute Trading Center: Give Choice Back to Enterprises
The official compute supermarket concept emphasizes flexible billing, online trading, automated delivery, and pay-as-you-go usage to reduce the upfront cost barrier for small and midsize enterprises.
The compute bank concept goes further. It explores a mechanism where compute resources can be deposited, metered, withdrawn, and even used with credit support. It also encourages idle compute resources to be contributed and scheduled across regions and time periods.
Unexhub’s compute trading center follows the same direction, but pushes it deeper into the Token and API layer.
Its core logic is:
aggregate global model and compute suppliers;
distribute Token-based capacity to developers and enterprises;
support pay-per-token usage;
route requests in milliseconds;
enable dynamic API bidding;
support real-time supplier settlement;
give early suppliers commission rates as low as 0%;
allow developers to access selected model routes and packages at costs as low as around 30% of market pricing.
This is not just about being cheaper.
It is about making compute more transparent, more liquid, and easier to access.
7. Why Dynamic Pricing Matters
Traditional cloud pricing is often fixed.
Whether a developer calls an API during peak hours or off-peak hours, the posted price may remain the same.
But real compute supply is not fixed.
Supply changes.
Demand changes.
Model availability changes.
Latency changes.
Idle capacity appears and disappears.
Dynamic pricing allows the market to reflect those changes.
For developers, this means:
non-urgent workloads can run during cheaper windows;
batch jobs can be scheduled for lower-cost periods;
low-risk tasks can be routed to cheaper suppliers;
premium models can be reserved for high-value requests.
For suppliers, this means:
idle capacity can be monetized;
peak-time demand can generate higher revenue;
small and midsize suppliers gain more pricing flexibility.
This is how compute shifts from a fixed asset to a liquid market resource.
8. Why Real-Time Settlement Matters
Long payment cycles are one of the biggest pain points for small compute suppliers.
A supplier may deliver compute today but wait 30, 60, or even more days to receive payment.
That creates cash-flow pressure.
It also makes it harder for smaller suppliers to compete with large cloud vendors.
Unexhub’s real-time settlement model changes this logic.
If a supplier provides usable Token capacity and users consume it, the supplier can receive revenue much faster.
That is not just a payment feature.
It is a supply-side growth engine.
The earlier a supplier connects, the sooner it can turn idle capacity into revenue. With early-stage 0% commission incentives, more of that revenue can stay with the supplier.
For small and midsize compute providers, this can make market participation much more realistic.
9. What This Means for Developers and Enterprises
For developers, Unexhub creates a more flexible access layer.
Instead of manually comparing dozens of providers, developers can route workloads through one marketplace and pay according to real usage.
For enterprises, the value is even clearer:
one access point for multiple model suppliers;
support for Gemini, ChatGPT, Claude, DeepSeek, GLM, and specialized smaller models;
pay-per-token billing;
millisecond-level routing;
dynamic pricing;
lower model access costs;
enterprise-grade customization;
clearer supplier settlement.
This helps companies avoid single-vendor lock-in.
A business can route complex reasoning to premium models, high-frequency classification to lower-cost models, local workloads to domestic models, and vertical tasks to specialized models.
That is how AI infrastructure becomes cost-aware.
10. Engineering Advice: How to Actually Save Money
Here are three practical rules.
1. Classify the workload first
Training uses GPU-hour.
Preprocessing uses core-hour.
Inference uses Token.
If you mix them blindly, your cost will rise.
2. Do not compare only unit prices
For Token pricing, compare the total task cost.
Check tokenization, input-output ratio, cache hit rate, retries, latency, and model quality.
3. Use dynamic pricing strategically
If your task can tolerate delay, schedule it during cheaper periods.
If it is user-facing and latency-sensitive, route it to faster models.
If it is low-risk and high-volume, use lower-cost suppliers.
This is how AI teams move from “using models” to “operating an AI cost system.”
Final Takeaway: Pricing Is Now an Architecture Decision
GPU-hour, core-hour, and Token are not just billing options.
They represent three different views of compute:
GPU-hour sells hardware time.
Core-hour sells general computation.
Token sells intelligent output.
For small and midsize enterprises, the most important shift is pay-as-you-go access.
They no longer need to invest heavily upfront just to use AI compute. They can buy what they need, when they need it, and pay based on actual usage.
Unexhub’s role is to make that choice easier:
aggregate supply, expose pricing, route demand, support Token settlement, enable dynamic bidding, and help suppliers get paid faster.
That is the real value of a next-generation compute trading center.
It does not only make AI cheaper.
It makes AI spending clearer, more flexible, and more controllable.
算力交易中心怎么计价?卡时、核时、Token 三选一,中小企业随用随付
算力市场不再只有一种计价方式。
2026 年 7 月 8 日,中国信息通信研究院联合算力产业发展方阵正式启动 “算力超银生态共建计划”,依托中国算力平台,共同探索 “算力超市” 和 “算力银行” 两大创新服务形态,重点包括标准化计价、在线交易、自动交付、信用支撑和降低中小企业用算门槛。
早在 2026 年 4 月,工信部就已提出探索“算力银行”“算力超市”等创新业务模式,并鼓励推行按 “卡时”“核时”及 Token 计费 等灵活付费模式。
也就是说,企业现在面对的是三种计价方式:
卡时、核时、Token。
但到底选哪个,不能拍脑袋。
要看任务类型。
一、卡时:适合训练和长期独占 GPU 的任务
卡时,是最传统的 AI 算力计价方式。
你租一张 GPU 卡,按使用时间付费。不管这张卡有没有跑满,只要占用了时间,就要付钱。
这种模式很容易理解:
一张卡,一小时,一个价格。
它适合这些场景:
模型训练;
模型微调;
大规模推理集群;
能长期跑满 GPU 的批量任务;
需要稳定硬件资源的科研任务。
卡时的优点是清楚。
供应商可以把 A100、H800 或其他 GPU 资源按小时标价,买方也很容易理解自己租到了什么。
但卡时的缺点也很明显:
僵硬。
如果企业只是偶尔调用模型,或者只是做轻量推理,按卡时付费就很容易浪费。你可能只用了 10% 的算力,却付了 100% 的钱。
所以,训练任务用卡时通常合理。
普通 API 推理任务用卡时,就像用卡车送一杯咖啡。
二、核时:适合通用计算,但不够 AI 原生
核时来自云计算时代。
它衡量的是 CPU 核心使用了多长时间。
到了 AI 时代,核时依然有价值,尤其适合不需要 GPU 或 GPU 利用率不高的任务:
数据清洗;
格式转换;
ETL 流程;
轻量预处理;
特征提取;
任务调度;
模型流程中的辅助计算。
很多 AI 项目并不是每一步都需要 GPU。
在训练或推理之前,企业经常需要整理数据、转换文件、去重、清洗标签、跑轻量脚本。
这些任务用核时计费,往往比卡时更划算。
但核时有一个根本问题:
它衡量的是计算消耗,不是智能产出。
一个简单文本分类任务和一个复杂推理任务,外围工作流消耗的 CPU 时间可能差不多,但业务价值完全不同。
所以,核时适合通用计算,却不适合作为 AI 模型服务的核心计价单位。
三、Token:AI 时代更原生的计价单位
Token 是三种计价方式里最年轻、也最有想象力的。
人民网关于“算力超市”的报道提到,算力超市类似电商平台的算力交易市场,可汇聚各类算力产品,并支持按“卡时”“核时”及 Token 计费等灵活付费方式进行在线交易。
广州也上线了全国首个基于 “词元(Token)”级调度 的城市综合算力运行服务平台。该平台以 Token 为统一计量基准,构建按量、按周期等多种灵活计费体系。
Token 计价改变的是 AI 支出的逻辑。
企业不再问:
我租了多少张卡?
而是问:
我实际消耗了多少智能服务?
所以,Token 特别适合这些场景:
大模型 API 调用;
Chatbot;
RAG 应用;
智能客服;
文本生成;
摘要;
分类;
Agent 工作流;
多模态模型调用。
对中小企业来说,Token 计价降低了入门门槛。
不用买卡。
不用租实例。
不用维护基础设施。
只为实际智能调用付费。
四、三种模式怎么选?看场景
三种计价方式不是互相替代,而是服务不同任务。
| 计价方式 | 最适合场景 | 核心优点 | 主要风险 |
|---|---|---|---|
| 卡时 | 训练、微调、长期独占 GPU 任务 | 资源明确,适合持续占用 | 利用率低时浪费严重 |
| 核时 | 数据预处理、通用计算、轻量任务 | 非 GPU 场景更便宜 | 不能衡量智能产出 |
| Token | 推理、API 调用、RAG、Agent、内容生成 | 按量付费,AI 原生 | 不同厂商 Token 规则不同 |
最简单的判断是:
训练用卡时。
预处理用核时。
推理和 API 调用用 Token。
最贵的错误,是把它们混着用错。
用卡时跑轻量推理,容易浪费预算。
用 Token 计价理解长时间训练,不合适。
用核时衡量 AI 输出价值,也会失真。
成熟企业不会问:
哪种计价方式最好?
而会问:
这个任务最适合哪种计价方式?
五、Token 计价也有隐藏成本
Token 计价更灵活,但不等于天然透明。
不同模型供应商对同一段文本的切分方式可能不同。同样一段内容,在不同模型里可能消耗不同数量的 Token。
所以,企业不能只比较“每百万 Token 多少钱”。
还要比较:
同一任务到底消耗多少 Token;
输入和输出如何分别计价;
缓存输入是否有折扣;
重试是否计费;
工具调用是否产生额外成本;
长上下文是否改变价格;
路由策略是否影响最终成本。
换句话说:
最便宜的 Token 单价,不一定等于最低的任务成本。
好的算力交易中心,应该把这些差异展示清楚。
这正是 Unexhub 这类平台的价值所在。
六、算力交易中心的角色:把选择权还给企业
官方“算力超市”强调灵活计费、在线交易、自动交付和随用随付,以降低中小企业一次性投入门槛。
“算力银行”则更进一步,探索算力资源 “存入—计费—支取—信贷” 机制,鼓励中小企业存入闲置算力,并通过跨区域、跨周期调度实现灵活取用。
Unexhub 正在做的算力交易中心,与这套方向高度一致,但更深入到 Token 和 API 层。
它的核心逻辑是:
聚合全球模型和算力供应商;
向开发者和企业分发 Token 能力;
支持按 Token 实付;
毫秒级路由请求;
支持 API 动态竞价;
支持供应商实时结算;
早期入驻供应商最低可享 0 抽佣;
开发者在部分路线和套餐中,最低可享受市场价格约 3 折的成本优势。
这不只是便宜。
而是让算力更透明、更灵活、更容易采购。
七、为什么动态定价重要?
传统云服务 API 定价通常是固定的。
不管开发者在高峰期调用,还是在闲时调用,标价往往一样。
但真实算力供应并不是固定的。
供给会变化。
需求会变化。
模型可用性会变化。
延迟会变化。
闲置资源会出现,也会消失。
动态定价,就是让市场反映这些变化。
对开发者来说:
非紧急任务可以放到低价时段;
批量任务可以在闲时运行;
低风险任务可以路由到更便宜供应商;
高价值请求再交给高能力模型。
对供应商来说:
闲置算力可以变现;
高峰期需求可以带来更高收益;
中小供应商拥有更多定价弹性。
这就是算力从固定资产变成流动资源的过程。
八、为什么实时结算重要?
长账期,是中小算力供应商最大的痛点之一。
供应商今天交付算力,但回款可能要等 30 天、60 天,甚至更久。
这会造成现金流压力,也让中小供应商很难和大云厂商竞争。
Unexhub 的实时结算机制,就是要改变这个逻辑。
供应商提供可用 Token 能力,用户产生调用,平台快速结算分成。
这不只是一个支付功能。
它是供给侧增长引擎。
供应商越早接入,越早把闲置算力变成收入。叠加早期 0 抽佣激励,更多收益可以留在供应商手里。
对中小算力供应商来说,这会让参与市场变得更现实。
九、这对开发者和企业意味着什么?
对开发者来说,Unexhub 提供的是更灵活的模型接入层。
开发者不用手动比较几十家供应商,可以通过一个市场入口完成调用,并按真实使用付费。
对企业来说,价值更清晰:
一个入口接入多类模型供应商;
支持 Gemini、ChatGPT、Claude、DeepSeek、GLM 及专业小模型;
按 Token 实付;
毫秒级路由;
动态价格;
降低模型接入成本;
企业级定制服务;
更清晰的供应商结算。
这能帮助企业避免被单一供应商锁定。
复杂推理走高能力模型,高频分类走低成本模型,本地化场景走国内模型,垂直任务走专业模型。
这才是 AI 基础设施的成本智能化。
十、工程师实操建议:怎么真正省钱?
给三条实操建议。
1. 先判断任务类型
训练用卡时。
预处理用核时。
推理用 Token。
任务类型判断错了,成本一定上升。
2. 不要只看单价
Token 计价要看整体任务成本。
要核算 Token 切分、输入输出比例、缓存命中率、重试、延迟和模型质量。
3. 善用动态定价
能容忍延迟的任务,放到更便宜的时段跑。
面向用户的实时任务,路由到更快模型。
低风险高频任务,交给低成本供应商。
这就是 AI 团队从“用模型”进化到“运营 AI 成本系统”的过程。
结语:计价方式,已经变成架构决策
卡时、核时、Token,不只是三种账单方式。
它们代表三种不同的算力视角:
卡时卖的是硬件时间。
核时卖的是通用计算。
Token 卖的是智能产出。
对中小企业来说,最大的变化是随用随付。
不需要一次性投入巨资买卡,不需要长期锁定实例,而是需要多少买多少,用多少付多少。
Unexhub 的角色,就是让这种选择更简单:
聚合供应、展示价格、路由需求、支持 Token 结算、推动动态竞价、帮助供应商更快回款。
这才是新一代算力交易中心的真正价值。
它不只是让 AI 更便宜。
它让 AI 支出更透明、更灵活、更可控。