您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：基本功的比拼，才是经营之道。
10月19日，一张全球AI新云（NeoCloud）产业图谱引发行业关注。
CoreWeave、Nebius以GPU云服务为核心；Lambda、Together AI与九章云极 DataCanvas则在算力供给之上，跨入“训练・推理・Token工厂”交叉赛道；AI 工程与AIDC运营，则作为底层底座贯穿整条产业链。在这场全球竞争中，九章云极是为数不多跻身第一梯队的中国厂商。
2026全球顶级AI新云NeoCloud图谱
这张图谱提供了观察行业的窗口：全球NeoCloud服务商正在告别单一算力售卖模式，走向商业模式分化、能力边界持续延伸的专业化赛道。
资本对头部标的的评估标尺也随之迭代，算力规模、硬件跑分等浅层指标逐步失效，行业核心评判标准收敛至四项：总体拥有成本（TCO）、训推交付稳定性、Token服务效率、AI基础设施工程成熟度。四项指标共同锚定技术底盘、履约能力、客户价值与长期经营质量，把短期算力红利沉淀为可复购、可持续的服务价值。
01
全球AI新云巨头分野
2026年，全球NeoCloud分层格局已经明朗。
Gartner高级总监分析师Enrique Castera提出，具备长期投资价值的新云厂商，除差异化技术能力之外，还需要稳定的客户结构、健康的财务模型与标准化企业级交付能力。
行业退去“有卡就能盈利”的幻想，集群供给、模型服务能力、工程运维水平、融资成本，共同划定企业的增长上限。
从业务路径划分，当前NeoCloud演化出三条并非互斥的赛道。
第一类，是以GPUaaS为核心，主打算力供给效率，代表厂商为 CoreWeave、Nebius。竞争围绕大规模GPU集群、AI原生云底座建设与算力交付效率展开，叠加融资成本决定单位算力盈利水平，客户以大模型研发企业与大型集团为主。
第二类为GPUaaS叠加MaaS，从算力向上延伸至模型服务，Lambda、Together AI、Crusoe、GroqCloud、九章云极DataCanvas均布局于此。其核心增量不是外挂一层模型API，而是打通算力与模型运行链路：训练团队需要弹性大规模集群，应用开发者需要低抖动、高吞吐的推理接口，政企客户看重合规可控、成本可预测的完整服务。
第三类将算力、模型服务统一纳入AI Engineering体系，依托优化物理编排与AIDC 运营为底座，逐层对外开放。这条路线在中国及算力主权市场都具备特殊现实意义。随着模型与集群规模扩张，算力能否释放理论性能，高度取决于软硬件协同。客户采购算力之后，环境适配、模型调优、运维排障会持续消耗研发资源；集群越大，网络、存储、散热、故障自愈越容易成为瓶颈。AI Engineering负责调度适配、训推优化与全链路运维；AIDC运营保障供电、网络、容灾稳定——硬件理论性能，只有在工程与基础设施协同之下，才能转化为稳定的生产算力。
九章云极DataCanvas是国内率先打通GPUaaS、MaaS、AI Engineering的新云厂商，同步参与多条赛道竞争。平台基于异构算力调度、训推编排、自动化运维搭建一体化底座，可帮助客户将TCO降低73.8%。
赛道边界持续消融，伴随硬件迭代与应用落地，AI新云正在发生训练、推理、Token 三重技术裂变与商业模式变化。竞争落脚点，最终是算力资源向标准化、可付费服务的转化能力。
02
国内市场：评估标尺回归业务本质
中国AI新云已进入训练与推理双轮驱动的增长周期。
IDC数据显示，2025年国内智算云基础设施（AI IaaS）市场规模486.7亿元，同比增长128%；同年公有云公有云MaaS模型调用量达到1944万亿Tokens，增幅约16倍，对应营收30.7亿元。IDC预测，2026全年Token消耗量将达到约4万万亿Tokens，公有云MaaS营收有望攀升至186亿元。
中国公有云MaaS市场进入高速增长期，Token经济从概念走向规模化实践
AI IaaS代表底层基建需求，MaaS代表模型商业化落地规模，二者不能简单加和，但共同印证增长同时发生在算力底座和模型服务两端。赛道热度持续吸引资本入场，2026年，花旗研究看好AI新云平台从算力供应商向完整AI平台演进，支撑逻辑包括GPU供需格局、企业AI应用渗透、底层软件栈升级。但资本市场已经告别无差别看多，融资成本、交付履约、细分增速、技术储备，都列为核心考察项。
落到厂商层面，具备可比性的评估指标回归业务本源：
总体拥有成本（TCO），衡量客户综合投入；训练与推理效率决定算力产出价值；Token 服务能力决定模型服务的经济性；AI Engineering工程能力决定AI落地效率。
这套指标串联技术与商业：对厂商，它决定红利能否转化为收入、现金流与关键场景；对客户，它决定采购的算力能否变成可用的AI能力。
03
九章智算云率先跑通全栈路径
在全球NeoCloud第一梯队图谱中，九章云极DataCanvas是少数具备完整全栈能力的中国厂商，在TCO、训练、推理、Token生产四项核心指标上达到国际领先水平，并被Forrester列为亚太地区普惠算力利用率的领先者之一。
2026全球AI新云NeoCloud细分业务能力可比性评估，九章云极位于第一梯队
区别于单纯供给算力资源，九章智算云依托“训推一致+AI工厂”闭环体系，打通异构算力调度、训推协同与底层基础设施运营，走出适配复杂硬件生态的差异化路线。
训练侧，九章智算云以长期有效算力利用率为核心，而非单纯比拼集群规模。灼识咨询30天满载压测显示：平台MFU（模型FLOP利用率）可达54.1%，日间波动率仅1.60%，优于行业常规区间。更高的算力利用率能够压缩训练周期、摊薄单位成本。平台推出独家DCU（1度算力）标准化计费模式，助力客户将总体拥有成本降低约 73.8%。其自研DART-GUI-7B拿下OS World 7B赛道全球第一，印证平台支撑高难度RL场景的工程实力。
九章智算云-训练算力利用效率
多方评估显示，九章智算云已成为国内科研机构、模型研发、物理AI开展大模型微调、强化学习训练的主流平台。
训推一致是这套体系的核心优势，落地难度远高于海外。海外NeoCould大多基于同构GPU集群搭建，而国内多芯片架构并存、软件生态割裂，极易出现“训得好、推不动”的断层。九章智算云的训练工厂与Token工厂共享Alaya OS+Alaya Atom Pod异构调度底座，训练阶段沉淀的算子、通信、集群优化策略可直接复用至推理，依托万卡级弹性集群支撑高吞吐推理服务。
面向Token规模化生产，九章智算云配套全套底层组件：自研DingoFS混合存储支持分钟级检查点，数百GB模型归档压缩至60秒内完成；叠加产品矩阵与DCU计量体系，将抽象算力转化为可计量、可交易的标准化Token服务，全量兼容国内主流开源大型，Token生产效率提升80%，已经在物理AI、科研等场景落地。
物理硬件编排构筑底层护城河。平台基于硬件拓扑感知就近分配任务，采用轨道优化互连与网络内计算技术，All-Reduce性能翻倍、流量减半；自动化容灾能力让千节点集群扩容耗时降低85%以上，节点配置提速7倍，把硬件纸面参数转化为稳定的生产算力。底层AIDC运营和上层MaaS服务互补，这也是它和算力租赁、API分销厂商最本质的区别。
九章智算云位居中国智算新云市场第一
九章智算云率先跑通全栈路径并实现高质量增长。灼识咨询口径，按2025年智算服务收入统计，九章智算云位居国内智算新云市场第一。
04
九章云极：异构环境下的中国方案
在全球NeoCloud能力图谱中，九章云极DataCanvas将算力供给、模型服务、AI 工程三层能力纳入同一框架，代表全球竞争语境下的中国方案。它的核心不是产品堆叠，而是打通异构算力、资源调度、训推交付的完整链路，直面算力本土化进程中长期存在的异构生态约束。
国内AI新云与生俱来的挑战，是多类芯片、驱动、编译框架长期共存，硬件无法互通。同一模型的运行表现，持续受算子适配、显存调度、通信框架、推理引擎制约。客户痛点早已不只是在算力资源，而是环境适配、算子调优、持续运维消耗大量研发资源。
九章云极依靠异构资源管理与云原生工程体系，覆盖异构算力、资源调度、训练推理、数据存储与性能观测，大幅降低AI落地摩擦。
算力供给层面，平台依托多容器集群适配多样化负载。面向训练场景，保障多卡多节点集群稳定运行，减少等待与任务中断；面向推理场景，按需弹性扩缩，规避高峰资源不足、低谷资源闲置，异构环境越复杂，调度价值越突出。推理链路中，PD分离、KV Cache复用、连续批处理等优化。这些技术的价值，最终需要在相同模型、相近硬件和一致工作负载下，以吞吐、时延、资源利用率及单位Token成本等指标验证。
AI工程层面，训练、推理、数据、调度是一体化闭环，完成从单项技术到工程协同。以具身智能研发为例，这类模型研发流程横跨数据采集、大规模并行训练、云端仿真、终端部署，算力需求波动明显。借助弹性能力，算力规模跟随研发阶段动态调整，削减闲置成本。
对九章云极而言，已经构建的异构生态工程能力，最终需真实业务的检验，体现在训练效率、推理性能、资源利用率与总体拥有成本等可量化指标上。
全球NeoCloud正沿着GPU供给，模型服务、数据中心基础设施和 AI 工程协同等不同路径演进，中国市场的增长为各类玩家提供了广阔空间。与此同时，驾驭复杂异构生态，既是中国AI新云厂商必须跨过的门槛，也可能成为构建差异化竞争优势的关键。
赛道竞争终将回归经营本质：能否稳定交付、控制成本，并持续降低客户的工程损耗，决定算力能否真正转化为可持续的服务价值。基本功的比拼，才是经营之道。
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: The basic skill is the way to operate.
On 19 October, a global NeoCloud industry map drew industry attention.
CoreWeave, Nebius, with GPU cloud services at its core; Lambda, Together AI, and Nine Clouds DataCanvas, on the other hand, have crossed the "Train-Judgery-Token Factory" cross-track over the calculus; AI works with AIDC, and runs as a bottom seat across the entire industrial chain. In this global competition, nine chapters of the clouds are the few Chinese manufacturers that are on the first stage.
2026 Top AI New Cloud NeoCloud Map
This map provides a window of observation for the industry: the global NeoClaud service provider is moving away from a single-calculable sales model towards a professional track that divides business models and continues to extend its capacity boundaries.
Capital assessments of headlines have been iterative, with shallow indicators of calculus size, hardware run-off, and core standards for industry being reduced to four: overall ownership costs (TCO), training delivery stability, Token service efficiency, and AI infrastructure engineering maturity. The four indicators together anchor the technology chassis, performance capacity, customer value and long-term operating quality, and sank short-term computing dividends as repurchaseable, sustainable service values.
01
Global AI New Cloud Headbreak
In 2026, the global NeoClaud hierarchy became clear.
Enrique Castera, Senior Inspector General of Gartner, suggested that new cloud manufacturers with long-term investment value needed, in addition to differentiated technological capabilities, stable client structures, healthy financial models and standardized enterprise-level delivery capabilities.
The industry retreats from the illusion of having a card to make a profit, with cluster supply, model service capacity, the level of engineering, financing costs, and a common ceiling for enterprise growth.
From the operational path, the current NeoCloud evolves into three tracks that are not mutually exclusive.
The first category, centred on GPUaAS, is intended to provide efficiency, representing the producers of CoreWeave and Nebius. Competition revolves around large-scale GPU clusters, the construction of the AI Nebula and the efficiency of calculator delivery, with co-financing costs determining the level of profitability per unit, with clients dominated by LLM R & D enterprises and large groups.
The second category, GPUAAAS superheavy Maas, extends from arithmetic to modelling services, where Lambda, Together AI, Crusoe, GroqClaud, and Nine Chapter Clouds DataCanvas are located. The core increment is not an outer layer of model API, but a functional and model running link: a training team requires flexible large-scale clusters, application developers require low-diverse, high-cost reasoning interfaces, and government clients value the full range of controlled and cost-predictable services.
The third category integrates computing and modelling services into the AI Engineering system, depending on the optimal physical organization and AIDC operations, and is open to the public from layer to layer. This route is particularly relevant in China as well as in the calculating sovereign market. As models and clusters expand, the ability of computing to release theoretical performance is highly dependent on the synergy of soft and hardware. After client purchasing power, environmental suitability, model optimization, and operational handicaps continue to consume R & D resources; the larger the clusters, the more easily the network, storage, heat dispersion, and failure itself become bottlenecks. AI Engineering is responsible for dispatch adaptation, training optimization and full-chain transport; AIDC operates to guarantee power supply, network, disaster-resilient-hardware theoretical performance, which can be translated into stable production capacity only in conjunction with engineering and infrastructure.
Nine chapters of the cloud pole, DataCanvas, are the first new cloud manufacturers in the country to reach GPUaas, Maas, AI Engineering and compete on multiple tracks simultaneously. The platform is based on an isomeric algorithm, training, automated transport, and integrated floors, which can help clients reduce TCO by 73.8 per cent.
The boundary of the track continues to melt, accompanied by an iterative and applied hardware, and the AI New Cloud is undergoing training, reasoning, token triple-technology fission and business model changes. The end point of competition is the ability of computing resources to convert to standardized paid services.
02
Domestic markets: assessing the return of the ruler to the nature of business
China ' s AI Xin Yun has entered a two-wheel-drive growth cycle of training and reasoning.
According to IDC, the market size of the domestic capital cloud infrastructure (AI IaaS) in 2025 was 48,670 million yuan, an increase of 128 per cent over the same period; the public cloud-based Maas model was called by 19,444 trillion Tokens in the same year, an increase of about 16 times, and its counterpart received 3.07 billion yuan. IDC predicts that Token consumption will reach approximately 4 trillion tokens throughout 2026, and that the revenues of the publicly owned Maas battalion are expected to climb to 18.6 billion yuan.
China's public cloud Maas market has entered a period of high growth, and the Token economy has moved from concept to scale.
AI Iaas represents the base infrastructure needs, and Maas represents the scale of commercialization of the model, which cannot be summed up, but co-proven growth occurs at both the base of computing and model services. The heat of the track continues to attract capital and, in 2026, Citigroup studies looked at the evolution of the AI New Cloud platform from a calculator supplier to a full AI platform, supported by logic, including the GPU supply and demand pattern, the enterprise AI application penetration and the upgrade of the bottom software stack. Capital markets, however, have been distinguished, and financing costs, delivery performance, subdivision rates of growth, and technology reserves are included as core items.
At the plant level, comparable assessment indicators revert to the source:
Overall cost of ownership (TCO) to measure integrated client inputs; training and reasoning efficiency to determine output value; Token service capacity to determine the economy of model services; AI Engineering capacity to determine AI landing efficiency.
This set of indicators links technology and business: for manufacturers, it determines whether dividends can be translated into income, cash flows and key scenarios; for customers, it determines whether the calculus of procurement can be turned into available AI capabilities.
03
Nine chapters of the Twilight cloud led the way to the inn.
In the global NeoClaud 1st Plateau, the nine-chapter cloud-top DataCanvas is a small group of Chinese manufacturers with full capacity, reaching international lead on four core indicators of TCO, training, reasoning, and Token production, and being one of the leading countries in the Asia-Pacific region in terms of the utilization of inclusive computing.
2026 Global AI New Cloud NeoClaud subdivision comparative assessment of operational capacity, with nine chapters of clouds in the first tier
Distinguished from purely supply-computing resources, nine chapters of the intellectual cloud rely on the “Teach for Convergence + AI Plant” closed loop system, which enables the interstellar power movement, training synergy and operation of the bottom infrastructure to move out of a differentiated path to fit complex hardware ecology.
On the side of training, the nine chapters focus on effective computing utilization over the long term rather than simply on cluster size. The 30-day full-load barometer of the calibration consultation showed that the platform MFU (model FLOP utilization) could reach 54.1 per cent, with only 1.60 per cent day-time fluctuations, which were better than the industry ' s conventional zone. Higher calculator utilization can compress the training cycle and minimize unit costs. The Platform has introduced a standardized costing model for the sole-source DCU (1 degree calculator), which helps clients reduce the overall cost of ownership by about 73.8 per cent. Its self-research DART-GUI-7B takes OS World 7B to the top of the world, and the certification platform supports the engineering power of the difficult RL scene.
Nine chapters of the book of minds -- training on efficiency in computing
Multi-stakeholder assessments have shown that nine chapters of the intellectual cloud have become the mainstream platform for national scientific institutions, model development, physical AI to fine-tune LLM and enhance learning and training.
Convergence is the core strength of the system, which is far more difficult to land than abroad. While overseas NeoCould is mostly based on co-constructed GPU clusters, domestic multi-chip structures coexist and software ecology is fragmented, with a high risk of “goodly trained and unmoved” faults. A nine-chapter training plant with the Token plant shares the Alaya OS+ Alaya Atom Pod isomer movement base, and the calculator, communication, cluster optimization strategy for the sedimentation of the training phase is directly re-engineered into reasoning and supports the high-tortment reasoning service with the Tovanca class flexible clusters.
For Token-scale production, nine chapters of the Smart Clouds package of submersible components: self-study DingoFS mixed storage support minute inspection points, with hundreds of GB models archived in 60 seconds; supersed product matrix and DCU measurement systems, transforming abstract algorithms into measurable, tradable and standardized Token services, full-size-compatible domestic mainstream open source, token production efficiency increased by 80 per cent, and physical AI, research, etc. have been set.
Physical hardware is structured to build the bottom moat. The platform is based on hardware leapfrogging awareness of near-allocation tasks, using orbit optimized interconnectivity and intra-network computing techniques, doubling All-Reduce performance and reducing traffic by half; and automating disaster capacity to reduce the capacity of the knob cluster by more than 85 per cent in time and increase the speed of nodal configuration by seven times, transforming hardware paper parameters into stable production capacity. The bottom AIDC operation and the upper level of Maas complement each other, which is also the difference between it and the essence of the API distributor.
I'm the best in China
Nine chapters have led the way to the whole warehouse and to high-quality growth. The calibration advice calibration, measured in 2025 by income from intellectual services, is the top of the domestic market.
04
Nine chapters of clouds: China ' s programme in an isomeric environment
In the global NeoClaud capability map, nine chapters of the clouds, DataCanvas, integrate the three tiers of computing, modelling services, AI engineering capabilities into the same framework, representing the Chinese programme in a global competitive context. Its core is not a stack of products, but a complete chain of isomeric algorithms, resource mobilization, and training delivery that confronts long-standing isomeric ecological constraints in the localization process.
The challenges inherent in the country ' s new AI cloud are the long-term coexistence of multiple types of chips, drivers, compilation frameworks and the inability of hardware to connect. The performance of the same model continues to be constrained by the calculator fit, VRAM movement, communication framework, and reasoning engine. Client pains have long been more than just computing resources, but rather environmentally appropriate, arithmetically well-adjusted, sustainable and resource-intensive.
Nine chapters of the cloud rely heavily on the management of isomeral resources and cloud bioengineering systems, covering isomeric algorithms, resource movement, training reasoning, data storage and performance observations, and significantly reducing AI landing friction.
At the power supply level, the platform is based on multi-container clusters adapted to diversified loads. A training-oriented scenario that ensures stable operation of the Docado nodes and reduces waiting and interruption of tasks; and a reasoning-oriented scenario that is flexible enough to scale up on demand, avoids peak under-resourced, low-lying resources, and the more complex the symmetrical environment, the greater the value of movement control. In the reasoning chain, PD separation, KV Cache reuse, batch processing etc. are optimized. The value of these technologies will ultimately need to be validated under the same models, similar hardware and consistent work loads for such indicators as ingestion, latency, resource utilization and unit Token cost.
At the AI engineering level, training, reasoning, data and movement control are integrated closed loops, from individual technology to engineering synergy. In the case of smart research, for example, the modelling process spans data collection, large-scale parallel training, cloud simulation, end-deployment, with significant fluctuations in computing needs. (c) To reduce idle costs by leveraging elasticity to calculate the size of power as it adjusts dynamically with the R&D phase.
For the nine chapters of the cloud, the constructed isomeric ecological engineering capacity will ultimately require real business testing, reflected in training efficiency, reasoning performance, resource utilization and overall ownership costs, among other indicators.
The global NeoClaud is evolving along different routes, such as GPU supply, modelling services, data centre infrastructure and AI engineering synergies, and the growth of the Chinese market has provided a wide range of players. At the same time, managing a complex isomeric ecology is both a threshold that China ' s AI New Cloud Manufacturers have to cross and may be key to building differential competitive advantages.
The race track competition will eventually revert to the nature of the business: the ability to stabilize delivery, control costs and continuously reduce client engineering losses and determine whether computing can truly be translated into sustainable service values. It's the basics that work.
For your account security, please verify the mailbox.
Check your mailbox.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.