Zhi东西（WeChat official account: zhidxcom）
Author | Sanbei
Editor | Mo Ying
Recently, a staff member from a company providing intelligent computing cloud services told Zhi东西 that in the past half year, customer demand points have changed significantly: previously, they asked "how much is one card, how long can it be rented", but now the questions have become "how much does one million Tokens cost, can the task run stably?"
The change in questions reflects an important upgrade in the AI infrastructure. The global AI cloud market is still expanding rapidly. A well-known research institution, Gartner, predicts that by 2030, the AI cloud market will reach 267 billion U.S. dollars (about 1.78 trillion RMB), with new cloud (NeoCloud) vendors expected to capture about 20% of the market share. The latter is also a market worth about 350 billion RMB.
However, on the other hand, the efficiency of computing resources remains a critical issue that the industry urgently needs to address. A survey conducted by the well-known research institution IDC and China Communications Standards Association in April this year, targeting 250 companies, showed that nearly half of the surveyed companies have intelligent computing resource utilization rates between 51% and 70%, and in industries such as manufacturing and government, the proportion of companies with utilization rates exceeding 70% is less than 35%.
The next challenge for the computing power industry is already emerging: how to make the already built computing power resources used more fully, flexibly, and stably?
Over the past two years, NeoCloud has completed the initial validation of its business model through GPU procurement, cluster construction, and computing power leasing. Now, with the simultaneous changes in computing power architecture, AI workloads, and customer needs, the industry's competitive standards are also shifting.
The next phase of AI new cloud is beginning to shift from competing on computing power scale to competing on computing power orchestration and scheduling capabilities.
I. From GPU leasing to intelligent delivery, why must new cloud be upgraded?
Looking back at the first stage of NeoCloud, the core task was to address the shortage of computing power supply.
Focusing on GPU procurement, cluster construction, and computing power leasing, it aimed to solve the initial computing power supply shortage in the AI industry. Resource scale, hardware performance, and cluster delivery speed were the main competitive indicators; this was equivalent to who could get the cards and build clusters, who would have the say.
But the AI industry has not stopped at the training phase, the demand side is changing the rules of the game.
As models enter the application phase, inference gradually becomes a continuous workload. Especially after Agent starts executing multi-step tasks, a single user request may trigger multiple rounds of model inference, tool calls, code execution, and data retrieval, and the task duration may also extend from a few seconds to several hours or even days.
This puts new requirements on the underlying infrastructure. A survey released by Google in July 2026 showed that among the 1,400+ enterprise IT executives surveyed, 83% believe that production-grade Agentic AI is urgently in need of infrastructure upgrades. Physical AI further expands the boundaries of compute demand. Microsoft Research's study on Physical AI in September indicates that some inference tasks can be offloaded from the robot's body to edge or cloud GPUs, making cloud-edge-device collaboration an important new area of research.
The policy side is also driving changes in the organization of compute power. In February of this year, the Chinese Ministry of Industry and Information Technology advanced the construction of the "1+M+N" national compute power interconnectivity node system, promoting standardized interconnection and efficient flow of compute resources across different regions, entities, and architectures, and comprehensively implementing automated monitoring of compute power capacity status. Compute power is shifting from dispersed construction to unified monitoring, interconnection scheduling, and resources becoming increasingly like a network that needs to be scheduled.
Looking horizontally at global new clouds, the upgrade actions are equally evident. The leading new cloud providers in the U.S., such as CoreWeave, have already split inference into three paths: Serverless Inference billed by token, Dedicated Inference billed by GPU hours, and self-managed Kubernetes inference. In Europe, new cloud provider Nebius offers enterprise-level hosted inference through its Token Factory, integrating it with the bare-metal compute power and serverless capabilities of Nebius AI Cloud to form an incremental hierarchy from compute power to inference services. Another U.S. new cloud provider Crusoe has also extended its business to hosted inference, model fine-tuning, and cluster automation management services.
An industry chain insider told Zhixing that the first half of the new cloud competition is about resource acquisition capabilities, and the second half is about resource organization capabilities.
Customers now need more than just GPU hours; they want to continuously, stably, and economically obtain intelligent outputs. New clouds are therefore shifting from being mere providers of compute resources to organizers of compute power and deliverers of intelligent services.
A competition over "controllable compute power" is emerging around the organization, scheduling, and delivery of compute resources. II. What is "controllable compute power"? The threshold has shifted from asset scale to three layers of capabilities.
所谓可控算力，核心是让算力资源从“可获得”，进一步走向“可组织、可调度、可计量、可交付”。Controllable compute power, at its core, aims to transform compute resources from being "available" to being "organizable, schedulable, measurable, and deliverable".
而可控算力与拥有算力的本质区别是什么？可以拆成三层能力：技术层的定制、调度和计量，商业层的资源控制与稳定供应，以及产业层面向客户交付算力服务。What is the fundamental difference between controllable compute power and owning compute power? It can be broken down into three layers of capabilities: technical customization, scheduling, and measurement; commercial resource control and stable supply; and industrial-level delivery of compute services to customers.
1、技术层：算力可控，先过调度和计量两道关1. Technical Layer: To achieve controllable compute power, one must first pass through the two gates of scheduling and measurement.
第一层是技术控制力，核心在于异构调度和统一计量。
The first layer is technical control capability, with the core focusing on heterogeneous scheduling and unified measurement.中国信通院《智能算力服务研究报告（2026年）》提出，智能算力服务体系可以分为资源层、互联互通层和应用层，分别承担资源汇聚与池化、跨域连接与统一调度，以及场景适配与服务交付等功能。换句话说，算力要真正成为可调用的服务，需要经过资源汇聚、池化管理、任务调度和持续运营等环节。
The China Academy of Information and Communications Technology's "Smart Compute Power Service Research Report (2026)" proposes that the smart compute power service system can be divided into three layers: resource layer, interconnectivity layer, and application layer, which respectively handle resource aggregation and pooling, cross-domain connectivity and unified scheduling, as well as scenario adaptation and service delivery. In other words, for compute power to truly become a callable service, it needs to go through stages such as resource aggregation, pooling management, task scheduling, and continuous operations. NVIDIA and Nebius joint testing in February provided a clear example: through GPU partitioning (such as MIG) or time slicing, half-card resources can achieve 77% of the throughput of a full card, and can support 86% of the concurrent users of a full card; when deploying multiple models together, the total capacity can approach three times that of deploying a single model. Converted, the throughput efficiency per unit resource increases by about 54%. This indicates that the same batch of computing power resources can form different effective computing powers through different partition granularities, memory allocation, and scheduling strategies.Unified measurement is another fundamental capability. Different chips have varying theoretical peak performance, actual performance, and applicable tasks, requiring a unified measurement standard. The industry is exploring using DCU, which stands for "one degree of computing power," as the unit for measuring computing power investment.
Taking the DCU (one degree of computing power) disclosed by Jiuzhang Cloud Extreme, a domestic intelligent computing cloud provider, as an example, it is defined as the continuous operation of 312 TFLOPS FP16 computing power for one hour, which is counted as one degree of computing power; thus, the computing power investment of different chips can be converted to a unified scale. Jiuzhang Cloud Extreme's growth also serves as an observation sample. According to its publicly disclosed prospectus, its controllable intelligent computing capacity has exceeded 29,000 PFLOPS, and the utilization rate of computing power resources has increased from 56.0% in 2025 to 79.6% in the first half of 2026. Under the continuous full-load operation of a 504-card cluster for 30 days, the MFU (model computing power utilization) reached 54.1%, with a daily fluctuation rate of 1.60%.
These data allow us to observe the relationship between software scheduling, resource management, and computing power utilization efficiency from one perspective, but they do not imply that a single indicator can fully represent the platform's overall capabilities.
2. Commercial layer: Controllable computing power does not necessarily require ownership as a prerequisite, and operations can realize capital efficiency.
The second layer is commercial control power, with the core being the stability of resource supply and capital efficiency.
Computing power operation is a business with significant capital intensity. GPU procurement, data center construction, and equipment upgrades all require substantial capital investment, and with the fast iteration of chips, equipment depreciation also affects long-term operating costs.
Therefore, new cloud providers need to answer the following question: Is it necessary to own all the underlying computing power assets to ensure continuous delivery to customers?
A clever approach is to not own all computing power assets, but to control the availability of computing power resources through long-term agreements. According to Jiuzhang Cloud Extreme's publicly disclosed prospectus, it obtains AIDC and the underlying computing power infrastructure through lease agreements with fixed terms of 3 to 8 years, without the option of delay or termination. This model places capital efficiency and resource scheduling capability on equal footing, and new cloud providers are more like organizers of computing power, focusing more on ensuring resource availability, scheduling rights, and service continuity.
Of course, long-term agreements also mean corresponding performance obligations and pressure for demand matching. The resource lock-in period, customer contract period, pricing mechanism, and actual utilization rate need to be incorporated into business management. Only when the resource supply matches customer demand can long-term agreements be transformed into stable service capabilities.
3. Industrial layer: From card-hour transactions to intelligent delivery, the service boundary is being rewritten.
The third layer is industrial control power, with the core being what is delivered.
Traditional computing power leasing usually uses the number of GPUs and usage duration as transaction units, giving customers the right to use a segment of computing resources. As the development of inference services and Agent applications progresses, customers are increasingly concerned about the final task completion cost, token output efficiency, and service stability.
This requires new cloud providers to extend their service boundaries from hardware resources to model deployment, inference optimization, task scheduling, and operational assurance. Billing methods may also gradually extend from card-hours to tokens or specific task outputs.
It is important to emphasize that the number of tokens itself cannot fully represent the value of intelligent services. Model capabilities, task complexity, response latency, accuracy, and service reliability all affect the final results customers obtain. Therefore, future intelligent delivery needs to establish a more comprehensive input-output evaluation system.
From resources to outputs, the three layers of capabilities—technology, business, and industry—are interrelated. Only the resource suppliers rely more on resource price differences for profit; service providers with some scheduling capabilities can improve resource allocation efficiency; and operators who can achieve resource customization, heterogeneous scheduling, unified measurement, long-term supply assurance, and continuous optimization delivery truly possess the complete capabilities of controllable computing power.
The core competitiveness of new clouds is extending from asset scale to scheduling, organization, and delivery capabilities.
III. Dual Migration of Customer Base and Workload, Controllable Computing Power is Reshaping the New CloudThe upgrading of computing power services is also reversing the customer base that the new cloud faces.
In the early days, a few large internet companies and model vendors were the main demanders of intelligent computing resources; today, a large number of AI-native enterprises, model vendors, and Agent teams are continuously emerging, focusing more on the flexibility, portability, cost, and actual output of computing power. New cloud providers offering the aforementioned "controllable computing power" services are now meeting this wave of new market demand. From the public offering prospectus of Jiuzhang Cloud Extreme, we can see the trend change: its intelligent computing cloud customers increased from 2 in 2024 to 84 by the first half of 2026. This data can serve as an observation sample for the growth of AI-native computing power demand, indicating that such demand is genuinely increasing.
For new clouds, changes in the customer structure mean that service methods need to be more refined. Different-sized enterprises may use different models, chips, and inference frameworks, and task types will continue to change with product iterations. Platforms need to configure resources based on customer needs and reduce the unit task cost through scheduling and optimization.
Not only are the customers changing, but the workloads are also changing.
From training to inference, to Agent and physical AI, each workload migration is expanding the service boundaries of new clouds.
In the training phase, the core challenge lies in the computing performance of large-scale clusters, network communication, and long-term stability. Once models enter the inference phase, the competition shifts towards latency, throughput, concurrency capacity, and unit Token cost. Agents further increase requirements for task state management, resource elastic allocation, and long-term operational assurance. By the physical AI phase, collaborative computing between cloud, edge, and robot entities brings latency, task priority, and cross-node scheduling to the forefront.
CoreWeave upgrades computing, storage, and network for Agentic AI, particularly emphasizing the sensitivity of multi-step Agent workflows to data access latency.
Returning to the industry itself, controllable computing power must connect resource perception, heterogeneous integration, dynamic scheduling, efficiency optimization, and service delivery to form a complete operational cycle. Only by seeing the computing power can it be managed; only by managing it can it be scheduled; only by scheduling it can it be optimized; ultimately, it can transform computing power into stable, sustainable AI services.
Conclusion: Transforming Computing Power into Sustainable Intelligent Capacity, AI New Clouds Enter a New Cycle
The competition for AI infrastructure is shifting from competing for scarce resources to improving resource organization efficiency. The upgrade direction of new clouds is moving from resource supply to intelligent delivery, from owning computing power to controllable computing power, from short-term supply to long-tail operations.
It can be predicted that in the future, truly competitive new clouds will not only be vendors with more computing power but also those capable of organizing dispersed, heterogeneous, and dynamically changing computing resources into stable, efficient, and sustainable AI services.
After the computing power arms race, the next threshold has already emerged. Transforming computing power into sustainable intelligent capacity might be the ticket for new clouds to enter a new cycle.
@@TAGn@@
