从"按卡时出租"到"按 Token 交付":算力交易中心 大模型 Token 计费价格体系重构,无问芯穹×MiniMax 抱团压成本

For a long time, computing power rental followed a simple logic: charging by graphics card and hourly usage. This bare-metal rental model suited the early cloud computing era but reveals fatal flaws in the large-model AI era.


From “Card Rental” to “Token Delivery”: Rewriting the Pricing Rules of Computing Power Trading Centers

On August 5, 2026, Unexhub (Wuwen Xinqiong) officially signed a strategic cooperation agreement with MiniMax. Unexhub is a Tsinghua-backed AI infrastructure enterprise with a scheduling computing power of 37,000P, while MiniMax is a leading model vendor with self-developed large models and massive application traffic. Both Shanghai-based enterprises have launched in-depth cooperation in four core dimensions: joint optimization of model inference efficiency, co-construction of large model Token services and AI infrastructure, collaborative innovation of application paradigms, and scaled expansion of industrial intelligent scenarios.

Traditional Billing Model: The Endogenous Limitations of Card-Time Rental

For a long time, computing power rental followed a simple logic: charging by graphics card and hourly usage. This bare-metal rental model suited the early cloud computing era but reveals fatal flaws in the large-model AI era.

The core contradiction lies in the mismatch between costs and revenue. Computing power suppliers bear fixed hardware and operating costs, while their income is floating and uncertain. Whether clients fully utilize the rented GPUs or generate valid Token outputs has no impact on suppliers’ revenue, which is solely derived from hardware rental fees rather than actual service output.

This model brings severe dilemmas to small and medium-sized computing power providers. After delivering computing power resources, they often face settlement cycles of weeks or even months. Stagnant cash flow severely restricts business expansion and scaled operation.

As Liang Wenfeng pointed out in a closed-door investor exchange: China is still facing insufficient high-quality computing power supply. For large-model enterprises, it is worthwhile to convert all available funds into hardware resources within a short period. What enterprises truly purchase today is no longer abstract computing power resources, but the final effective Token output generated by computing power.

Token Factory: Industrial Upgrade from “Selling Resources” to “Selling Output”

Since 2026, the business logic of the AI industry has undergone a fundamental overhaul. The “Token Factory” concept proposed by Jensen Huang at GTC 2026 has moved from theoretical conception to large-scale industrial practice, transforming traditional data centers into standardized factories dedicated to Token production.

At WAIC 2026, Unexhub launched its strategic layout of “Front Store, Back Factory & Central Hub”, building a closed-loop industrial system with clear division of labor:

1. Central Hub: Distributed Computing Power Dispatching Center

It centrally manages and intelligently schedules heterogeneous computing power resources distributed across the country. At present, the platform has accessed 37,000P of high-quality computing power, covering 16 mainstream chip models, realizing full coverage and flexible allocation of domestic computing power resources.

2. Back Factory: Standardized Token Production Factory

Taking power and chip running time as basic production materials and full-stack software optimization as the core production formula, the platform realizes high-efficiency, large-scale standardized Token output. Through technological innovations such as cross-cluster heterogeneous PD separation architecture (PDD), Unexhub has achieved remarkable cost and latency optimization: the single-Token inference cost has been reduced by 37.5% within one year, the first-Token latency has dropped by 51.5%, and the overall reasoning cost has been cut to one-tenth of the original level.

3. Front Store: AI Productivity Service Store

It undertakes market demand docking, packages standardized Token outputs into usable industrial services, and converts technical resources into scalable AI productivity for all industries.

The three-layer architecture forms a self-driven industrial flywheel. Market demand from the front store feeds back to optimize Token production efficiency in the back factory, which further activates richer resource scheduling from the central hub. According to Xia Lixue, co-founder and CEO of Unexhub, the platform’s daily Token call volume has surged 40 times from last December to July this year, with over 95% of inference tasks driven by intelligent agent scenarios.

Unexhub’s Solution: Building a Standardized Trading Exchange for Tokens

While the Token Factory solves the industry’s production efficiency problem, the disordered circulation and uneven pricing of Tokens have long plagued industrial development. The same Token unit varies by thousands of times in price across different vendors, models and scenarios, reflecting the lack of unified trading rules and fair price discovery mechanisms in the immature Token market.

Xia Lixue has repeatedly proposed building a “computing power supermarket” at public events, aiming to standardize and commercialize computing power resources like water and electricity to serve thousands of industries. Unexhub’s core positioning is to build a professional trading exchange for Token, the “new oil” of the AI era, with a four-layer core operating mechanism to standardize industrial transactions:

1. Real-Time Fund Settlement

Suppliers receive synchronous fund settlement based on the actual number of Tokens supplied. The real-time clearing mechanism completely eliminates cash flow pressure and solves the long-standing pain point of long settlement cycles in the computing power industry.

2. Zero Commission for Early Settlers

Adopting the strategy of exchanging short-term profits for long-term industrial liquidity, the platform offers zero-commission benefits for early settled suppliers. Rich supplier resources and diversified Token categories continuously strengthen the platform’s network effect and ecological barriers.

3. Market-Driven API Bidding Mechanism

Breaking the traditional one-sided pricing model of manufacturers, API prices float dynamically according to real market supply and demand. Prices decline in off-peak periods to release suppressed long-tail demand, and rise moderately in peak periods to incentivize suppliers to expand production capacity, realizing a win-win balance for both sides.

4. One-Stop Full-Cycle Infrastructure

The platform provides complete supporting systems for merchant settlement, Token trading and fund clearing. Suppliers only need to focus on core Token production and supply, with zero additional operational costs, greatly lowering the industry entry threshold.

Pricing Unit Overhaul: Rewriting the Underlying Logic of the Computing Industry

The shift from “card rental” to “Token delivery” seems to be only a change in billing unit, but it completely subverts the industry’s economic logic.

Card-time rental sells idle hardware resources, charging regardless of actual usage and effective output. In contrast, Token-based delivery sells real industrial value, where users only pay for the actual Token resources consumed. As analyzed by Dongwu Securities, this transformation will significantly improve the profitability of computing power rental enterprises. Guolian Minsheng Securities estimates that under the new Token sharing model, the capital return cycle of computing power investment can be shortened to approximately 2 years.

The strategic cooperation between Unexhub and MiniMax essentially realizes the deep binding of the production end and consumption end of the Token economy. Unexhub optimizes MiniMax’s model inference efficiency to achieve lower costs and faster response, while MiniMax provides real commercial scenarios and massive Token consumption demand. Facing the dual pressure of high computing costs and difficult scenario implementation, such upstream and downstream ecological cooperation has become a mainstream development trend for domestic large-model enterprises.

Xia Lixue predicts that future AI services will further evolve from simple Token quantity billing to value-based billing linked to actual business benefits. Price stratification is only a superficial phenomenon, while value stratification represents the core future competition of the industry.

Engineer’s Perspective: The Historical Watershed of the Computing Industry

Tokens are evolving into the most standardized currency in the AI era, and any mature currency system requires a complete trading and clearing market.

Unexhub’s computing power dispatching center and Token factory, together with MiniMax’s massive Token consumption scenarios, have jointly overturned the traditional computing pricing system. The shift from resource-based card rental to value-based Token delivery marks a historic watershed for the industry.

Before 2026, computing power was defined as a scarce resource, billed by hardware and time. After 2026, computing power is defined as actionable output, billed by Token quantity and practical value. This tiny change in billing units reshapes the entire industrial underlying logic: every unit of computing power must be accountable for its Token output and the actual industrial value it creates.

This marks the official start of Token Economics, with the entire industrial clearing layer as the core competitive chip.

Unexhub Core Platform Advantages & Exclusive Benefits

Unexhub has rebuilt the global Token trading logic and built a leading aggregated computing power trading center. The platform gathers full-category AI suppliers worldwide, covering mainstream large models such as Gemini, ChatGPT, Claude, DeepSeek and GLM, as well as high-quality customized models from individual developers and small and medium-sized enterprises.

Relying on the multi-supplier bidding mechanism and millisecond-level intelligent routing technology, the platform delivers ultra-low-cost and high-efficiency AI services. For developers and enterprises, the actual Token-based payment cost is as low as 30% of the official market price. For AI suppliers, one-click platform access and real-time settlement of call sharing revenue are supported, with zero early commission and efficient capital turnover.


2026年8月5日,无问芯穹(Unexhub)与MiniMax正式达成战略合作。一方是拥有清华系技术背景、算力调度规模达37000P的头部AI基础设施企业,一方是自研大模型、坐拥海量落地场景与应用流量的顶尖模型厂商。两家扎根上海的科技企业,围绕四大核心维度展开深度合作:模型推理效能联合优化、大模型Token服务与AI基础设施共建、应用范式协同创新、行业智能化场景规模化拓展。

用工程逻辑通俗解读此次合作:无问芯穹负责做优Token供给,持续压低单Token生产成本、提升供给效率;MiniMax负责放大Token价值,依托海量真实场景拉升Token实际利用率与商业价值。前者深耕供给端效率,后者撬动需求端体量,两大国产大模型产业链核心玩家深度绑定,打通了国产大模型“训练-推理-商业化落地”的全链路闭环。

而这场上下游深度联手的背后,是整个算力产业正在经历的一场颠覆性变革:行业计价规则彻底迭代,从传统的“按卡、按时长出租”,全面迈向精细化、价值化的“按Token实际交付计费”。

传统计费模式:按卡出租的固有行业瓶颈

过去,算力租赁商业模式简单粗放,核心逻辑为“按卡按时长收费”。这种裸金属租赁模式适配早期云计算发展阶段,但在AI大模型规模化落地的当下,固有弊端彻底暴露。

其核心痛点是成本与收益严重错配:算力供应商承担硬件采购、设备运维、电力损耗等固定成本,但收入完全浮动,与客户实际使用效果无关。客户租赁显卡后,无论是否跑满算力、是否产出有效Token,供应商都只能收取固定硬件租金,无法共享模型落地的增值收益。

这一模式更是中小算力供应商的发展桎梏。商家完成算力供货后,回款周期普遍长达数周甚至数月,现金流长期承压,无法支撑业务规模化扩张,行业中小玩家难以做大做强。

正如梁文锋在闭门投资人交流中坦言:国内优质算力始终处于供给不足状态,大模型企业若短期内将资金全部投入算力建设,必然是最优选择。当下行业真正的采购核心,早已不是抽象的硬件算力资源,而是算力最终产出的、可落地、可复用的有效Token价值。

Token工厂:从“卖硬件资源”到“卖有效产出”

2026年,AI行业商业模式迎来根本性迭代。黄仁勋在GTC 2026大会上提出的“Token工厂”概念,正式从行业理论落地为规模化产业实践,传统数据中心加速转型为标准化、高效化的Token智能生产基地。

无问芯穹在WAIC 2026世界人工智能大会上发布“前店后厂一中心”核心战略,构建分工明确、闭环运转的全新产业体系,彻底重构算力变现逻辑:

一中心:全域算力集散调度中心

统一纳管全国分布式异构算力资源,依托智能调度算法实现算力精准分配。目前平台已整合37000P优质算力,兼容覆盖16类主流芯片型号,实现国内算力资源的全域覆盖、灵活调度与高效复用。

后厂:标准化Token智能生产工厂

以芯片算力、电力能耗为基础生产资料,依托全栈自研软件优化技术打造专属“生产配方”,实现Token规模化、低成本、高效率量产。通过跨集群异构PD分离架构(PDD)等核心技术突破,无问芯穹实现多项关键指标大幅优化:单Token推理成本一年内下降37.5%,首Token延迟降低51.5%,整体推理成本压缩至原先十分之一。

前店:AI生产力服务商店

聚焦市场需求对接与商业化落地,将标准化Token产出包装为各行各业可直接落地、高效复用的AI生产力服务,完成从技术资源到商业价值的最终转化。

三层架构形成自驱动正向飞轮:前端市场需求反向驱动后端Token工厂优化生产效率,进而调动算力中心全域优质资源,实现产业闭环持续迭代升级。据无问芯穹联合创始人兼CEO夏立雪介绍,平台单日Token调用量从去年12月至今年7月暴涨40倍,超95%的推理任务均来自智能体商业化场景,产业落地成效显著。

Unexhub全新解法:为Token经济搭建标准化交易枢纽

Token工厂解决了行业“高效生产”的核心难题,但Token流通环节的乱象始终存在:不同厂商、不同模型、不同场景下的同款Token,价格差距高达上千倍。无序定价、规则缺失,成为制约Token经济规模化发展的核心阻碍。

夏立雪多次公开提出,要打造算力行业的“生态超市”,让算力像水电一样标准化、普惠化、服务化,赋能千行百业智能化升级。无问芯穹Unexhub的核心价值,就是为AI时代的“新石油”Token搭建一座专业、公允、高效的算力交易交易所,通过四层核心机制规范行业交易规则:

1. 实时清算结算,零现金流压力

供应商按照实际Token供货量,同步回收对应收益,资金实时到账清算。彻底解决传统算力行业回款周期长、现金流卡顿的痛点,将供应商财务压力降至最低。

2. 早期入驻零佣金,让利生态共建者

平台以短期利润让利换取行业流动性,对早期入驻的供应商实行零佣金政策。入驻商家越多、Token品类越丰富,平台生态网络效应越强,持续构建行业竞争壁垒。

3. API动态竞价,市场化公允定价

打破厂商单方面垄断定价模式,API价格随市场供需关系动态浮动。需求低谷时段价格下行,释放被高价压制的长尾开发者需求;需求高峰时段价格合理上浮,激励供应商扩容增产,实现供需两端双向共赢。

4. 全链路基础设施,降低入局门槛

平台搭建涵盖商家入驻、Token交易、资金清算、数据统计的一站式完整体系,供应商无需投入额外运营成本,只需专注核心的Token生产与供给,大幅降低行业入局门槛。

计价单位革新,重塑全行业底层商业逻辑

从“按卡出租”到“按Token交付”,表面是计价单位的简单调整,本质是整个算力产业商业逻辑的颠覆性重构。

按卡出租,售卖的是闲置硬件资源,无论是否产生有效价值,均按时长收费;按Token交付,售卖的是真实产业产出,用户仅为实际消耗的有效Token付费,计费更精准、成本更可控、价值更匹配。东吴证券分析指出,这一模式转型将大幅提升算力租赁企业的盈利空间;国联民生证券测算,全新Token分成模式下,算力投资回本周期可缩短至2年左右,产业经济效益大幅提升。

无问芯穹与MiniMax的战略合作,本质是Token经济生产端与消费端的深度绑定、双向赋能:无问芯穹助力MiniMax模型推理更高效、成本更低,优化Token产出质量与效率;MiniMax为平台提供海量真实商业场景与规模化Token消耗需求。在当前算力成本高企、场景落地困难的行业现状下,这种上下游抱团共生、生态共建的模式,已成为国产大模型产业发展的主流趋势。

夏立雪预判,未来AI服务计费模式将持续迭代,从单一的“按Token数量计费”,升级为按Token创造的实际业务价值分层计费。价格分层只是表象,价值分层终将成为行业核心竞争壁垒。

工程师视角:2026,算力产业的时代分水岭

Token正在成为AI时代最具通用性、标准化的数字货币,而任何成熟的货币体系,都离不开完善的交易与清算市场。

无问芯穹的全域算力集散中心、标准化Token工厂,叠加MiniMax的规模化Token消费场景,三者深度联动,彻底改写了算力行业的定价规则。从资源导向的按卡计费,到价值导向的按Token计费,2026年已然成为算力产业的清晰分水岭。

在此之前,算力是稀缺硬件资源,计价依托设备与时长;在此之后,算力是可量化的智能产出,计价依托Token数量与商业价值。计价单位的微小变革,撬动的是整个产业的底层逻辑升级:每一份算力资源,都必须对自身的Token产出、以及Token创造的实际产业价值负责。

这正是Token经济学落地的真正起点,而整场产业变革的核心筹码,是覆盖全行业的Token经济清算体系。

Unexhub平台核心权益

Unexhub全面重构全球Token交易逻辑,打造行业顶尖的Token聚合分发与算力交易中心。平台汇聚全域AI供应商,涵盖Gemini、ChatGPT、Claude、DeepSeek、GLM等海内外主流大模型,以及个人、中小企业优质定制模型。依托海量供应商竞价机制+毫秒级智能路由技术,为企业与开发者提供极致低价的AI服务。

开发者、企业用户可按Token实际消耗量付费,最低享官方市场价3折,高强度批量调用也能极致控本;AI供应商可一键入驻平台,调用分成实时结算、早期入驻零佣金,资金周转高效无压力。