九章云极发布“AI工厂”战略 定义智能规模化新基建
英文摘要
Jiuzhang Yunji (JZYJ) announced its 'AI Factory' strategy, pivoting the company to an infrastructure platform that measures compute input in DCUs and output in standardized professional Tokens. The AI Factory consists of a training factory for refining generic intelligence into task-specific professional models, and a token factory for encapsulating those models into scalable, measurable delivery units. The company targets a 100,000 P intelligent computing cluster, daily throughput of 10 trillion tokens, and a thousand-fold comprehensive cost reduction through full-stack in-house technology. Three technical paradigm shifts underpin the factory: a compute-storage-transmission integrated architecture with tenfold TPS improvement, a persistent kernel execution mechanism to eliminate idle cycles, and energy-defined computing for per-token energy accountability. JZYJ simultaneously launched an open plan to build a neutral, inclusive computing ecosystem with chip, model, and energy partners, and was recognized in an Analysys report as a leader in China's third-party neutral intelligent computing cloud market.
中文摘要
九章云极发布“AI工厂”战略，以DCU计量算力投入、专业Token计量智能产出，构建可度量的规模化交付体系。AI工厂包含两大引擎：训练工厂（专业模型诞生地）和Token工厂（标准化专业Token交付网络）。公司设定了10万P智能算力集群、单日10万亿Token流转承载力、千倍级综合降本三大目标，并依托全栈自研Alaya NeW智算底座完成系统架构、计算调度、能效架构三大范式重构。同时启动智算开放计划，联合产业链共建中立普惠算力共同体。易观分析报告显示九章云极在全国第三方普惠智算云市场位居前列。
关键要点
New AI Factory strategy uses DCU as compute input unit and professional Token as standardized output, industrializing AI delivery.
AI工厂战略以DCU计量算力投入、专业Token计量产出，实现AI规模化、标准化交付。
Two engines: training factory converts generic intelligence to professional models; token factory packages them for scalable, measurable delivery.
训练工厂将通用智能转化为专业模型，Token工厂封装为可规模化、可精确度量的专业Token。
Ambitious scale targets: 100,000 P computing cluster, daily 10 trillion tokens throughput, thousand-fold cost reduction via full-stack self-developed technology.
目标建设10万P智算集群，单日10万亿Token流转承载力，通过全栈自研实现千倍级降本。
Three technical paradigm shifts: integrated compute-storage-transmission architecture with tenfold TPS gain, persistent kernel scheduling to eliminate idle waste, and energy-defined computing with per-token energy tracking.
三大技术范式重构：算存传一体化架构带来十倍推理性能提升，持久化执行流机制消除算力空耗，能源定义计算架构实现专业Token能耗全程量化。
Open plan launched to build a neutral, inclusive intelligent computing community with chip, server, model, and energy partners.
启动智算开放计划，联合芯片、服务器、大模型、能源企业共建中立普惠算力共同体。
Analysys report places Jiuzhang Yunji as a leader in China’s third-party neutral intelligent computing cloud market.
易观分析报告显示九章云极在全国第三方普惠智算云市场位居前列，竞争逻辑从算力保有量转向交付价值。