2026 Global AI NeoCloud Landscape: DataCanvas Enters the Top Tier

On October 19, a global AI NeoCloud industry landscape map drew industry attention.

CoreWeave and Nebius center on GPU cloud services; Lambda, Together AI, and DataCanvas have moved beyond compute supply into the intersecting track of "training, inference, and Token factories"; AI engineering and AIDC operations run through the entire industry chain as the underlying foundation. In this global competition, DataCanvas is one of the few Chinese vendors to make it into the first tier.

2026 Global Top AI NeoCloud Landscape

This map offers a window into the industry: global NeoCloud service providers are moving away from the single compute-selling model toward specialized tracks with diverging business models and continuously extending capability boundaries.

Capital's yardstick for evaluating leading targets has also evolved accordingly. Shallow metrics such as compute scale and hardware benchmark scores are gradually losing relevance, and the industry's core evaluation criteria have converged on four items: total cost of ownership (TCO), training and inference delivery stability, Token service efficiency, and AI infrastructure engineering maturity. Together these four metrics anchor the technology foundation, delivery capability, customer value, and long-term operating quality, turning short-term compute dividends into repeatable, sustainable service value.

01

In 2026, the global NeoCloud tiered landscape has become clear.

Enrique Castera, Senior Director Analyst at Gartner, proposed that NeoCloud vendors with long-term investment value need, in addition to differentiated technical capabilities, a stable customer structure, a healthy financial model, and standardized enterprise-grade delivery capability.

The industry has shed the illusion that "having GPUs means profitability"; cluster supply, model service capability, engineering operations, and financing costs together define a company's growth ceiling.

By business path, NeoCloud has currently evolved three non-mutually-exclusive tracks.

The first category centers on GPUaaS, focusing on compute supply efficiency, represented by CoreWeave and Nebius. Competition revolves around large-scale GPU clusters, AI-native cloud foundation building, and compute delivery efficiency, with financing costs determining unit compute profitability; customers are mainly large-model R&D companies and large conglomerates.

The second category is GPUaaS plus MaaS, extending upward from compute to model services, where Lambda, Together AI, Crusoe, GroqCloud, and DataCanvas are all positioned. Its core increment is not bolting on a model API layer, but connecting compute and model operation links: training teams need elastic large-scale clusters, application developers need low-jitter, high-throughput inference interfaces, and government and enterprise customers value compliant, controllable, and cost-predictable full services.

The third category brings compute and model services unified into an AI Engineering system, relying on optimized physical orchestration and AIDC operations as the foundation and opening up layer by layer. This route has special practical significance in China and in compute-sovereignty markets. As models and cluster scale expand, whether compute can unleash theoretical performance depends heavily on software-hardware coordination. After customers purchase compute, environment adaptation, model tuning, and operations troubleshooting will continuously consume R&D resources; the larger the cluster, the more easily networking, storage, cooling, and fault self-healing become bottlenecks. AI Engineering handles scheduling adaptation, training and inference optimization, and full-link operations; AIDC operations ensure stable power supply, networking, and disaster recovery—hardware theoretical performance can only be converted into stable production compute under the coordination of engineering and infrastructure.

DataCanvas is the first NeoCloud vendor in China to connect GPUaaS, MaaS, and AI Engineering, competing simultaneously across multiple tracks. The platform builds an integrated foundation based on heterogeneous compute scheduling, training and inference orchestration, and automated operations, helping customers reduce TCO by 73.8%.

Track boundaries continue to blur, and with hardware iteration and application deployment, AI NeoCloud is undergoing a triple technological fission and business model change across training, inference, and Tokens. The ultimate competitive focus is the ability to convert compute resources into standardized, billable services.

IDC data shows that in 2025, China's intelligent computing cloud infrastructure (AI IaaS) market reached 48.67 billion yuan, up 128% year on year; in the same year, public cloud MaaS model invocation volume reached 1,944 trillion Tokens, an increase of about 16 times, corresponding to revenue of 3.07 billion yuan. IDC predicts that full-year 2026 Token consumption will reach about 40 quadrillion Tokens, and public cloud MaaS revenue is expected to climb to 18.6 billion yuan.

China's public cloud MaaS market enters a period of rapid growth, and the Token economy moves from concept to scaled practice

AI IaaS represents underlying infrastructure demand, and MaaS represents the scale of model commercialization; the two cannot simply be added together, but together they confirm that growth is occurring at both the compute foundation and model service ends. Track heat continues to attract capital. In 2026, Citi Research is bullish on AI NeoCloud platforms evolving from compute suppliers to full AI platforms, with supporting logic including the GPU supply-demand landscape, enterprise AI application penetration, and underlying software stack upgrades. But the capital market has moved past indiscriminate bullishness; financing costs, delivery fulfillment, segment growth rates, and technology reserves are all listed as core considerations.

At the vendor level, comparable evaluation metrics return to business fundamentals:

Total cost of ownership (TCO), measuring the customer's overall investment; training and inference efficiency determine the value of compute output; Token service capability determines the economics of model services; AI Engineering capability determines the efficiency of AI implementation.

03

Jiuzhang Intelligent Computing Cloud is the first to prove out the full-stack path

In the global NeoCloud first-tier landscape, Jiuzhang Cloud DataCanvas is one of the few Chinese vendors with complete full-stack capabilities, reaching internationally leading levels on the four core metrics of TCO, training, inference, and Token production, and was listed by Forrester as one of the leaders in inclusive compute utilization in the Asia-Pacific region.

2026 Global AI NeoCloud segmented business capability comparability assessment, Jiuzhang Cloud ranks in the first tier

On the training side, Jiuzhang Intelligent Computing Cloud centers on long-term effective compute utilization rather than simply competing on cluster scale. China Insights Consultancy's 30-day full-load stress test shows that the platform's MFU (model FLOP utilization) can reach 54.1%, with a daytime fluctuation rate of only 1.60%, better than the industry's typical range. Higher compute utilization can compress training cycles and dilute unit costs. The platform launched an exclusive DCU (1 unit of compute) standardized billing model, helping customers reduce total cost of ownership by approximately 73.8%. Its self-developed DART-GUI-7B took first place globally in the OS World 7B track, confirming the platform's engineering strength in supporting highly difficult RL scenarios.

Jiuzhang Intelligent Computing Cloud - training compute utilization efficiency

Training-inference consistency is the core advantage of this system, and it is far harder to implement than overseas. Most overseas NeoClouds are built on homogeneous GPU clusters, whereas in China multiple chip architectures coexist and software ecosystems are fragmented, making it very easy for a disconnect to emerge where "training works well but inference won't run." Jiuzhang Intelligent Computing Cloud's training factory and Token factory share the Alaya OS + Alaya Atom Pod heterogeneous scheduling foundation; the operators, communication, and cluster optimization strategies accumulated during training can be directly reused for inference, relying on a 10,000-card-level elastic cluster to support high-throughput inference services.

For large-scale Token production, Jiuzhang Intelligent Computing Cloud provides a full set of underlying components: the self-developed DingoFS hybrid storage supports minute-level checkpoints, archiving and compressing hundreds of GB of models within 60 seconds; combined with the product matrix and DCU metering system, it transforms abstract compute into measurable, tradable standardized Token services, fully compatible with mainstream domestic open-source large models, improving Token production efficiency by 80%, and has already been implemented in scenarios such as physical AI and scientific research.

Physical hardware orchestration builds the underlying moat. The platform allocates tasks nearby based on hardware topology awareness, adopts rail-optimized interconnect and in-network computing technology, doubling All-Reduce performance and halving traffic; automated disaster recovery capabilities reduce the expansion time of thousand-node clusters by more than 85% and speed up node configuration by 7x, turning paper hardware specifications into stable production compute. Underlying AIDC operations and upper-layer MaaS services complement each other, which is also the most essential difference between it and compute leasing and API distribution vendors.

Jiuzhang Intelligent Computing Cloud ranks first in China's intelligent computing new cloud market

Jiuzhang Intelligent Computing Cloud is the first to prove out the full-stack path and achieve high-quality growth. By China Insights Consultancy's methodology, based on 2025 intelligent computing service revenue, Jiuzhang Intelligent Computing Cloud ranks first in the domestic intelligent computing new cloud market.

04

Jiuzhang Cloud: a Chinese solution in heterogeneous environments

In the global NeoCloud capability landscape, Jiuzhang Cloud DataCanvas brings the three layers of compute supply, model services, and AI engineering into a single framework, representing a Chinese solution in the context of global competition. Its core is not product stacking, but connecting the complete chain of heterogeneous compute, resource scheduling, and training-inference delivery, directly confronting the long-standing heterogeneous ecosystem constraints in the process of compute localization.

The inherent challenge for domestic AI new clouds is that multiple types of chips, drivers, and compilation frameworks coexist over the long term, and hardware cannot interoperate. The runtime performance of the same model is continuously constrained by operator adaptation, memory scheduling, communication frameworks, and inference engines. Customer pain points are no longer just about compute resources, but about environment adaptation, operator tuning, and continuous operations consuming large amounts of R&D resources.

At the compute supply level, the platform relies on multi-container clusters to adapt to diverse workloads. For training scenarios, it ensures stable operation of multi-card, multi-node clusters, reducing waiting and task interruptions; for inference scenarios, it scales elastically on demand, avoiding resource shortages at peak times and idle resources at troughs; the more complex the heterogeneous environment, the more prominent the value of scheduling. In the inference chain, optimizations such as PD separation, KV Cache reuse, and continuous batching. The value of these technologies ultimately needs to be verified under the same model, similar hardware, and consistent workloads, using metrics such as throughput, latency, resource utilization, and cost per Token.

At the AI engineering level, training, inference, data, and scheduling form an integrated closed loop, completing the transition from individual technologies to engineering collaboration. Taking embodied intelligence R&D as an example, this type of model R&D process spans data collection, large-scale parallel training, cloud simulation, and terminal deployment, with clearly fluctuating compute demand. With elastic capabilities, compute scale adjusts dynamically with the R&D stage, cutting idle costs.

For Jiuzhang Cloud, the heterogeneous ecosystem engineering capabilities it has already built ultimately need to be tested by real business, reflected in quantifiable metrics such as training efficiency, inference performance, resource utilization, and total cost of ownership.

Global NeoCloud is evolving along different paths such as GPU supply, model services, data center infrastructure, and AI engineering collaboration, and growth in the Chinese market provides broad space for various players. At the same time, mastering complex heterogeneous ecosystems is both a threshold that Chinese AI new cloud vendors must cross and may also become the key to building differentiated competitive advantages.

Competition in the track will ultimately return to the essence of business: whether it can deliver stably, control costs, and continuously reduce customers' engineering losses determines whether compute can truly be transformed into sustainable service value. Competition in fundamentals is the way of business.