算力交易中心 Token 即服务 推理效能优化成核心竞争力,每瓦 Token 生产效率定生死

The industry competition logic has completely changed from simply expanding computing power cluster scale to improving per-watt Token production efficiency.


MaaS Exit, TaaS Takeover: The Token Economics of Computing Power Trading Centers

In the summer of 2026, academician Zheng Weimin from the Chinese Academy of Engineering repeatedly emphasized a core judgment at multiple industry forums: the core competition of the AI industry is shifting from MaaS (Model-as-a-Service) to TaaS (Token-as-a-Service). The industry competition logic has completely changed from simply expanding computing power cluster scale to improving per-watt Token production efficiency.

This is not a simple replacement of industry abbreviations, but a fundamental rewrite of the underlying rules for AI industrial development.

Academician Zheng Weimin gave a straightforward interpretation: As the smallest semantic unit for AI information processing, Token serves as the basic measurement unit for computing power consumption and the core pricing unit for AI services. The sharp rise in Token importance stems from the fundamental transformation of AI application forms. The generative AI era is dominated by simple dialogue requests, while the AI Agent era requires autonomous, long-cycle task execution, triggering a hundred-fold expansion in Token consumption.

Faced with a thousand-fold growth in Token consumption, traditional measurement, billing and transaction models can no longer adapt to industrial development demands.

Limitations of MaaS: Model Supply Does Not Equal Productivity Output

The core logic of MaaS (Model-as-a-Service) is simple: service providers deploy complete models, and users pay based on call times or usage duration. This model perfectly matched the early generative AI stage when model resources were scarce and basic availability was the core demand.

However, inherent flaws are fully exposed in the AI Agent era. Simple model integration can no longer support efficient industrial supply. What enterprises and developers truly need is not “available model resources”, but low-cost, high-speed, high-stable Token supply capacity.

Academician Zheng pointed out that current domestic AI computing power infrastructure is mainly oriented to large-model training rather than continuous Token production for large-scale AI application. Simply stacking open-source models and inference frameworks lacks in-depth hardware optimization and system-level intelligent scheduling. Expensive computing power is restricted by multi-link bottlenecks, eventually falling into an industrial dilemma of high energy consumption and low Token output.

Essentially, MaaS focuses on “whether the model can run normally”, while TaaS focuses on “whether Tokens can be produced stably, cheaply and efficiently at scale”.

Per-Watt Token Efficiency: The New Core Competitive Metric of the Industry

With the paradigm shift from MaaS to TaaS, the industry’s core competition benchmark has been completely updated.

At GTC 2026, Jensen Huang proposed the “Token Factory Economics”. Its core logic is clear and rigorous: the upper limit of physical power consumption is fixed. A 1GW computing cluster cannot be upgraded to 2GW simply by optimization. Under fixed power consumption constraints, vendors with higher per-watt Token throughput have lower production costs and stronger market competitiveness. The traditional evaluation standard focusing on peak computing power and hardware parameter stacking has been completely eliminated.

Currently, the core evaluation standard of AI infrastructure has officially shifted toToken cost efficiency — maximizing valid Token output per watt of power consumption while meeting business latency requirements. Per-watt Token production efficiency has become the lifeline of all computing power service providers.

PDD Architecture: From Theoretical Paradigm to Industrial Engineering Implementation

The PDD cross-cluster heterogeneous inference architecture released by Unexhub at WAIC 2026 is a typical landmark engineering practice under the TaaS paradigm.

Large-model inference consists of two completely different stages with differentiated resource attributes: computation-intensive Prefill stage and memory-bandwidth-intensive Decode stage. Traditional industrial solutions adopt a “universal hardware” idea, improving performance only by stacking computing power, video memory and bandwidth, resulting in serious resource waste.

Unexhub adopts a differentiated innovative idea: connecting geographically dispersed homogeneous clusters through low-cost wide-area Ethernet to realize cross-cluster heterogeneous separated inference. It creatively upgrades the traditional two-stage P-D separation into a three-tier P-RLD-MD separated inference architecture:

  • Prefill (P) Stage: Deployed on high-performance GPU clusters, responsible for computing-intensive input prompt pre-filling tasks.

  • RelayDecode (RLD) Stage: Serving as the core intermediate relay layer, it efficiently transmits KV Cache data between Prefill and MainDecode, completely breaking the bandwidth bottleneck of cross-cluster transmission.

  • MainDecode (MD) Stage: Deployed on high-bandwidth video memory hardware, specializing in memory-intensive iterative Token generation output.

All optimization data comes from real production environment verification. The architecture reduces Time-To-First-Token (TTFT) latency by 51.5% and cuts single-Token production costs by 37.5%. Supported by 37,000P nationwide accessible computing power and 16 types of mainstream heterogeneous chips, Unexhub has reduced overall inference costs to one-tenth of the original level in the past year. Meanwhile, the platform’s daily Token call volume has increased by approximately 40 times compared with December 2025.

Computing Power Trading Center: Standardized Infrastructure for Token Circulation

Technological innovation solves the problem of Token production efficiency, while a standardized trading platform is required to realize efficient, free and fair Token circulation.

Essentially, Unexhub builds a professional trading exchange for Tokens — the “new oil” of the AI era, completing the industrial closed loop from Token production to market circulation through standardized institutional mechanisms:

  • Real-Time Settlement: Suppliers receive synchronous fund returns based on actual Token supply volume, zeroing out cash flow pressure and long settlement cycles.

  • Zero Commission for Early Settlers: The platform sacrifices short-term profits to gather ecological liquidity, helping early suppliers rapidly expand market scale with zero operational cost.

  • Market-Driven API Bidding: Breaking the monopoly of one-sided vendor pricing, API prices float dynamically with real market supply and demand. Off-peak price declines release long-tail developer demand, while peak price increases incentivize suppliers to expand capacity, forming a healthy self-regulating market.

Academician Zheng Weimin summarized three core technical foundations for TaaS industrial development: full-system heterogeneous collaboration, storage-computation synergy, and SLO-oriented intelligent scheduling. Unexhub’s computing power trading center essentially realizes the landing of these underlying technologies at the business level, realizing standardized production, pricing, trading and settlement of Tokens.

Engineer’s Perspective: Value Restructuring of the AI Industry Chain

The shift from MaaS to TaaS is far more than a conceptual abbreviation upgrade; it is a comprehensive re-anchoring of the entire industrial value chain.

In the MaaS era, industrial value is anchored in models — vendors with superior model performance hold absolute pricing power. In the TaaS era, value is anchored in Token production capability — vendors with lower per-watt costs, lower latency and more stable throughput dominate the market.

Nebula Technology’s semi-annual report strongly validates this trend: its revenue reached 254 million yuan, a year-on-year increase of 497.11%, with a net profit growth of 540.15%, achieving a turnaround from loss to profit. The core transformation is upgrading the traditional computing power rental model to a Token consumption-based billing model — the change of pricing unit reshapes the entire business logic.

Tokens are evolving into the most standard digital currency in the AI era, and per-watt Token production efficiency has become the ultimate metric for measuring Token production costs. PDD architecture, Token factories and computing power trading centers jointly build a complete infrastructure system, making Token production and circulation as stable, efficient and accessible as water and electricity resources.

As Zheng Weimin stated at the Jiuxianqiao Forum: “In the past, we focused on building independent models; today, we need to build a complete intelligent supply system. The shift from Model-as-a-Service to Token-as-a-Service is an inevitable infrastructure upgrade for the Agent era.” This is the core essence of the TaaS era.

Unexhub Platform Core Ecological Advantages

Unexhub reshapes the global Token circulation and trading system, building an integrated computing power trading center for Token aggregation, distribution and transaction. The platform gathers full-spec AI ecological partners worldwide, covering top international large models such as Gemini, ChatGPT, Claude, DeepSeek and GLM, as well as high-quality customized model resources from individual developers and small and medium-sized enterprises.

Relying on multi-merchant bidding competition and millisecond-level intelligent routing technology, Unexhub delivers ultra-cost-effective AI services. Developers and enterprise users pay strictly based on actual Token consumption, with the lowest discount as low as30% of the official market price. AI suppliers can access the platform with one click, enjoy real-time settlement of call sharing revenue, and achieve efficient and stable profit realization with zero early commission cost.


2026年盛夏,中国工程院院士郑纬民在各大行业论坛持续抛出核心判断:AI产业的竞争核心正在从MaaS(模型即服务)全面跃迁到TaaS(Token即服务),行业比拼维度彻底从“算力集群规模堆砌”,转向每瓦Token生产效率的精细化竞争。

这绝非简单的行业概念替换,而是AI产业底层发展规则的全面重写。

Token崛起:AI产业全新核心计量单位

一组行业硬核数据印证了产业变革:2026年3月,国内Token日均调用量突破140万亿次,相较2024年初的100亿次,涨幅超1000倍。

郑纬民对此做出通俗解读:Token是AI处理信息的最小语义单元,是算力消耗的基础计量标尺,更是AI商业服务的核心计价单位。Token地位的跨越式提升,根源在于AI应用形态的迭代升级。生成式AI时代以单次简单对话消耗为主,而智能体(Agent)时代,AI可围绕目标自主完成长周期、多步骤任务,直接带动Token消耗量呈百倍级爆发增长。

当行业消耗以千倍速度激增,传统的算力计量、交易、计费模式,已然无法适配产业规模化发展需求。

MaaS模式瓶颈:有模型,不等于有生产力

MaaS(模型即服务)的核心逻辑简单直白:平台部署模型,用户按需调用,按调用次数、使用时长计费。在生成式AI早期、模型资源稀缺的阶段,这套模式完美适配市场需求,满足企业基础AI落地诉求。

但进入智能体时代,MaaS的底层短板彻底暴露。单纯的模型堆叠、接口调用,已经无法支撑产业高效供给。企业真正的核心需求,早已从“有模型可用”,升级为以最低成本、最快响应速度、最稳定质量获取足量Token产能。

郑纬民指出,国内现有AI算力基建大多聚焦大模型训练场景,而非智能体规模化落地所需的持续性Token生产。行业普遍存在“重模型部署、轻底层优化”的问题,单纯堆砌开源模型与推理框架,缺乏硬件深度适配与系统级智能调度,导致高价算力在多环节遭遇效能瓶颈,最终陷入高能耗、低产出的Token产能困境。

一句话总结两代模式差异:MaaS关注“模型能不能跑”,TaaS关注“Token能不能稳定、低价、高效、规模化生产”。

每瓦Token生产效率:行业全新核心竞争标尺

伴随MaaS到TaaS的范式切换,整个行业的竞争评判标准彻底重构。

黄仁勋在2026年GTC大会提出“Token工厂经济学”,核心逻辑直击行业本质:硬件物理功耗存在固定上限,1GW算力集群无法通过优化突破2GW物理极限。在固定功耗约束下,谁能实现更高的单瓦Token吞吐量,谁就能压低生产成本,抢占市场优势。过往比拼峰值算力、堆叠硬件参数的粗放式选型标准,已经彻底过时。

如今行业核心评估指标,正式聚焦Token成本效率——在满足业务延迟、稳定性要求的前提下,单位功耗产出的有效Token数量。每瓦Token生产效率,已然成为决定算力服务商核心竞争力的关键命脉。

PDD架构落地:TaaS范式的工业化工程实践

无问芯穹(Unexhub)在2026世界人工智能大会(WAIC)发布的PDD跨集群异构推理架构,是TaaS范式从理论走向落地的标杆级工程成果。

大模型推理分为两大特性完全对立的阶段:Prefill预填充(计算密集型)、Decode解码(访存密集型)。传统行业思路追求硬件全能适配,通过堆叠算力、显存、带宽弥补性能短板,造成严重的资源浪费与效能损耗。

无问芯穹采用差异化创新思路:依托低成本广域网以太网,串联全国多地已建成的同构算力集群,实现跨集群异构分离推理。同时将传统二段式P-D分离架构,迭代升级为P-RLD-MD三级分层推理架构,实现任务与硬件的精准匹配:

  • Prefill预填充层(P):部署于高性能GPU集群,专注承接计算密集型的输入预填充任务,最大化释放算力优势。

  • RelayDecode中继解码层(RLD):搭建专属中间传输枢纽,高效流转Prefill与MainDecode之间的KV Cache数据,彻底破解跨集群传输带宽瓶颈。

  • MainDecode主解码层(MD):部署于高显存带宽硬件集群,专注逐Token迭代生成输出,适配访存密集型场景需求。

所有优化数据均来自真实生产环境:该架构实现首Token延迟降低51.5%,单Token生产成本下降37.5%。依托全国37000P可调度算力、16类主流适配芯片,无问芯穹一年内将整体推理成本压缩至原先十分之一,平台单日Token调用量较2025年12月暴涨40倍。

算力交易中心:构建Token高效流通的产业基建

PDD架构解决了Token“高效生产”的问题,而产业规模化发展,还需要一套标准化体系解决Token“顺畅流通、公允定价”的核心难题。

无问芯穹Unexhub的核心价值,就是为AI时代的“新石油”Token搭建专属交易枢纽,通过三大核心机制,完善Token产业流通基建:

  • 实时清算结算:算力供应商按实际Token供货量同步回款,资金即时到账,彻底归零企业财务周转压力与长周期回款难题。

  • 早期入驻零佣金:平台以短期利润让利换取产业流动性,扶持早期入驻生态伙伴快速起量、零成本拓业。

  • 市场化API竞价:打破厂商单方面垄断定价,API价格随市场供需动态浮动。需求低谷降价释放中小开发者长尾需求,需求高峰涨价激励供应商扩容增产,形成良性产业循环。

郑纬民总结TaaS产业发展的三大技术底座:全系统异构协同、存算协同的存换算体系、面向SLO的智能调度。而Unexhub算力交易中心,正是三大底层技术的商业化落地载体,实现Token生产、定价、交易、结算的全流程标准化。

工程师视角:产业价值链的彻底重构

从MaaS到TaaS的迭代,绝非简单的概念更替,而是整个AI产业价值链的价值重锚定。

MaaS时代,产业价值锚定在“模型本身”,模型性能优劣决定定价权归属;TaaS时代,产业价值锚定在“Token产能”,谁能实现更低单瓦能耗、更低首Token延迟、更稳定的规模化产出,谁就能掌控市场核心话语权。

星云科技半年报数据,精准印证了这一产业趋势:企业营收2.54亿元,同比增长497.11%,净利润同比增长540.15%,成功扭亏为盈。其核心变革,正是将传统算力租赁模式,升级为Token实量计费模式——计价规则的革新,彻底重构了商业增长逻辑。

Token已然成为AI时代最具通用性的数字硬通货,而每瓦Token生产效率,就是衡量这份“数字资产”生产成本的终极标尺。PDD架构、Token工厂、算力交易中心,三大体系协同发力,共同让Token的生产与流通,变得如同水电一般稳定、高效、随取随用。

正如郑纬民在酒仙桥论坛所言:“过去我们专注搭建单一模型,如今我们需要搭建完整的智能供给体系。从模型服务走向Token服务,是智能体时代对产业基建的必然要求。”这正是TaaS新范式的核心内核。

Unexhub平台核心权益

Unexhub重构全球Token交易流通规则,打造集Token聚合、分发、交易于一体的全域算力交易中心。平台汇聚Gemini、ChatGPT、Claude、DeepSeek、GLM等海内外头部大模型厂商,同时吸纳个人开发者、中小企业优质定制模型资源,汇聚海量AI服务商竞价入驻。

依托毫秒级智能路由与市场化竞价机制,平台极致压缩服务成本。开发者、企业用户支持Token实量付费,最低可享市场原价3折,大幅降低AI研发与落地门槛;AI供应商可一键快速入驻,调用分成实时结算,早期入驻全程零佣金,高效实现商业变现。