From 10,000 Cards to 100,000 Cards: Domestic Compute Is at the Turning Point of “System-Level Scale-Up”

Over the past week, the hottest topic in the compute sector has not been the benchmark score of a particular chip, nor a parameter race among model companies, but a broader and more structural signal: domestic compute is moving from “single-point breakthroughs” to “system-level scale-up.”

This judgment comes from two pieces of information that appeared almost simultaneously. One is from the policy side: the Ministry of Industry and Information Technology issued the "15th Five-Year Plan for the Development of the Information and Communications Industry," which explicitly proposes the orderly deployment of 10,000-card, 100,000-card and above intelligent computing clusters, greater efforts to adapt domestic compute chips, and a target of 9,800 EFLOPS of intelligent computing capacity by 2030. The other is from the industry side: JD Cloud announced at the 2026 JD Global Technology Explorer Conference that it plans to build a 100,000-card intelligent computing cluster based on Moore Threads’ full-function GPUs as the compute foundation, focusing on large-model training, inference and embodied intelligence, and opening the compute to the entire industry. Moore Threads explicitly stated that this marks the first time domestic GPUs have entered a top AI cloud company’s 100,000-card core intelligent computing cluster.

Policy planning and industrial implementation have formed a resonance within the same window. This is the most symbolic combination of events in the compute field in 2026. It means domestic compute is no longer just a “policy encouragement direction,” but has begun to enter the core business-carrying layer of top cloud vendors. It is also no longer just a story of “single-card replacement,” but a systemic competition integrating clusters, networks, storage, scheduling, frameworks, models and services.

I. Policy Direction: 100,000-Card Clusters and Domestic Chip Adaptation Enter Top-Level Design

The weight of the "15th Five-Year Plan for the Development of the Information and Communications Industry" lies in the fact that it no longer stops at “encouraging development,” but gives clear technical paths, scale targets and supporting requirements. The plan proposes building a multi-level computing facility system of “hub–region–edge,” deploying inference compute facilities on demand for scenarios, and “increasing efforts to adapt domestic compute chips.” In terms of compute transmission channels, the plan deploys 400G and above high-speed transmission technology and promotes the sinking of 100G optical network equipment. In green compute, it requires that by 2030, the PUE of newly built large and super-large computing facilities drop below 1.2.

At the same time, the Cyberspace Administration of China and six other departments jointly issued the "Implementation Plan for Promoting the Coordinated Transformation and Development of Digitalization and Greening (2026–2030)," elevating the greening of computing facilities to a position as important as the expansion of compute scale. It explicitly proposes strengthening green design for computing facilities, supporting technological innovation in liquid cooling, waste heat recovery and low-power chips, exploring an 800V high-voltage direct current power supply architecture, and promoting the deployment of racks above 100 kW.

Taken together, these two documents send a very clear policy signal: domestic compute is no longer merely a “backup option,” but is to become the main supply of computing infrastructure. Moreover, policy attention is no longer limited to “how much compute to build,” but also includes “how compute is connected, used and operated greenly.” This means that competition in compute construction over the next few years will not simply be a competition over the number of GPUs, but a comprehensive competition in power, cooling, networking, scheduling, software ecosystems and operational efficiency.

The shift from “10,000 cards” to “100,000 cards” is not a simple numerical addition. A 10,000-card cluster already places high demands on network topology, parallel training, fault tolerance mechanisms and storage throughput; a 100,000-card cluster magnifies these problems by an order of magnitude. Failures will become normal, communication bottlenecks will become more prominent, energy consumption and heat dissipation pressure will rise sharply, and scheduling and O&M complexity will increase exponentially. By writing 100,000-card and above clusters into the plan, policy is in effect forcing the entire industrial chain to shift from “selling hardware” to “delivering systems.”

II. Industry Implementation: Domestic GPUs Receive a “Vote of Confidence” from Top Cloud Vendors

The policy direction has been set, but what truly excites the industry is the cooperation between JD Cloud and Moore Threads.

The 100,000-card full-function GPU cluster uses Moore Threads GPUs as its compute foundation, serves large-model training, inference and embodied intelligence, and opens compute to the entire industry. Cao Peng, chairman of the JD Group Technology Committee and president of JD Cloud, positioned it as “a self-reliant and controllable ultra-large-scale compute foundation in the domestic AI cloud field.” The key words in this statement are two: “self-reliant and controllable,” and “ultra-large-scale.” The former points to supply chain security; the latter points to engineering capability. Whether domestic GPUs can enter the core intelligent computing clusters of leading AI cloud enterprises has long been a question mark. Now, that question mark is being straightened out.

More noteworthy is technological maturity. The joint solution from Qujing Technology and Moore Threads has been put into production, handling real token traffic from the official business of leading model vendors. In production operating data, the MTT S5000 and the domestic PD heterogeneous solution run cooperatively, with average token generation speed exceeding 50 TPS, KV cache hit rate exceeding 90%, and system stability reaching 99.9%. This means domestic GPUs are no longer merely “usable” in the laboratory, but “good to use” in production environments. For cloud vendors and model vendors, production data is far more convincing than any peak metric at a launch event.

Cluster scale is also ramping rapidly. A 10,000-card cluster based on the MTT S5000 has already provided services for ultra-large-parameter large-model training, and inference services for token factories are being launched and operated one after another. From 10,000 cards to 100,000 cards, the engineering capability of domestic compute is undergoing a real scale test. A 100,000-card cluster is not simply multiplying 10,000 cards by ten. It involves a series of system-level issues such as high-speed interconnection, lossless networks, parallel file systems, checkpoint recovery, fault isolation, task scheduling and energy management. If any link has a shortcoming, the effective compute of the entire cluster will be greatly discounted.

Therefore, the significance of JD Cloud’s move is not just “how many domestic GPUs it bought,” but that domestic GPUs have begun to be incorporated into the core compute planning of top cloud vendors. This is equivalent to a “vote of confidence”: domestic compute is no longer only edge business, pilot business or backup business, but can serve as one of the main foundations carrying key scenarios such as large-model training, inference and embodied intelligence.

III. Demand Side: Model Iteration and Compute Price Increases Create a Double Pull

Behind the accelerated scale-up of domestic compute is strong demand-side pull.

In early September, OpenAI released its new-generation flagship model GPT-6 Astra, with a context window of 1.05 million tokens and mature implementation in Computer-Use autonomous operation and multi-step end-to-end tasks. The generational leap in model capabilities directly brings a structural increase in compute consumption. In long-horizon tasks, a single call involves million-level token context and multi-turn tool calls, and the computation, VRAM usage and communication overhead in the inference stage all increase significantly. In the past, many applications only needed a short Q&A; now, agents must continuously plan, call tools, read long documents and execute multi-step operations. Token consumption and compute usage are no longer on the same order of magnitude.

Domestically, the DeepSeek V4.1 Flash model was officially released, comprehensively surpassing the previous V4 Pro in core indicators such as performance and response speed, while announcing price cuts across the Flash series, with the price for input cache misses reduced by 60%. Model cost reduction will further drive up API call volume, and inference compute demand will expand accordingly. Price cuts will encourage more developers to connect, and more application connections will bring more token calls, ultimately feeding back to the compute side as continuously growing inference demand.

Demand-side expansion has already been reflected in prices. Since the beginning of this year, Tencent Cloud, Alibaba Cloud, Baidu AI Cloud and other vendors have密集 raised prices for compute products, with some services increasing by more than 30% at the highest. Rental prices for Nvidia-series GPUs have also mostly risen, with B200 rental prices up 11.65% monthly and RTX5090 up 28.57% monthly. As of March this year, China’s average daily token usage had exceeded 140 trillion, an increase of more than 40% in three months compared with 100 trillion at the end of 2025.

The supply-demand gap is still widening. Chen Weiliang, founder of MetaX, pointed out at the 2026 China Unicom Partner Conference that large-model parameters have moved toward the trillion and even ten-trillion level, compute consumption for training and inference is growing exponentially, and compute supply is facing a sustained gap. This judgment explains why compute prices are rising, rentals are rising, cloud vendors are expanding, and policy is pushing. Demand is not a short-term pulse, but a long-term expansion brought by the superposition of improved model capabilities, falling inference costs and the spread of application scenarios.

IV. Why This Round of “Domestic Substitution” Is Different

In recent years, the advancement of domestic compute has often been criticized as having “policies but no scenarios.” Single-card performance was catching up, but the software ecosystem was weak, migration costs were high, and customers did not dare to put core business on it. Many projects remained at local intelligent computing centers, laboratory validation and small-scale pilot stages, making it difficult to form a commercial loop.

But the change in this round is that from policy goals to industrial implementation, and from model adaptation to production validation, every link has made substantive progress within the same time window.

The "15th Five-Year Plan" writes 100,000-card clusters and domestic chip adaptation into top-level design. The cooperation between JD Cloud and Moore Threads proves that domestic GPUs are already capable of undertaking the core business of top cloud vendors. Production data from Qujing Technology and Moore Threads shows that the competitiveness of domestic compute is extending from single-card performance to system-level token production efficiency. At the same time, China Unicom announced that during the "15th Five-Year Plan" period it expects to leverage 100-billion-yuan-level investment, strengthen compute in the Beijing-Tianjin-Hebei region, the Yangtze River Delta and the Guangdong-Hong Kong-Macao Greater Bay Area, build gigawatt-level parks, and “enable more domestic large models to use high-quality domestic compute.”

Behind this, a more complete consortium is taking shape: telecom operators provide networks, parks, energy and capital; cloud vendors provide platforms, scheduling and customers; GPU vendors provide the compute foundation; model vendors provide scenarios and traffic; and system integrators provide heterogeneous collaboration and engineering optimization. Domestic compute is no longer a one-company fight, but a systemic project jointly advanced by upstream and downstream players in the industrial chain.

The so-called “system-level scale-up” contains at least four layers of meaning. First, the core of competition shifts from single-card performance to cluster efficiency, with interconnection, storage, scheduling, fault tolerance and energy efficiency becoming key indicators. Second, buyers expand from laboratories and local intelligent computing centers to top cloud vendors, telecom operators and large model companies. Third, applications expand from pilot demos to real token traffic in production environments. Fourth, policy shifts from simple subsidies to planning, standards, green development, transmission channels and industrial collaboration. Only when these four layers advance simultaneously can it be called “system-level scale-up.”

V. What a 100,000-Card Cluster Means

A 100,000-card cluster first pushes energy and cooling issues to the fore. Gigawatt-level parks, racks above 100 kW, 800V high-voltage direct current, liquid cooling and waste heat recovery—these links that once seemed like “supporting technologies” are becoming decisive factors in whether a compute cluster can be deployed. A PUE below 1.2 is not the icing on the cake, but a hard constraint. Without green compute, there can be no sustainable compute scale.

Second, it means networks and storage must be upgraded in tandem. 400G and above high-speed transmission and the sinking of 100G optical network equipment are intended to form efficient channels between compute hubs, within clusters and between edge nodes. 100,000-card training requires high-bandwidth, low-latency, lossless networks and high-throughput parallel file systems to support checkpoint writing and reading. If networks and storage cannot keep up, no matter how many GPUs there are, they will be underfed.

Third, it means the importance of the software ecosystem is further amplified. The CUDA ecosystem has deep accumulation. For domestic GPUs to be used at scale in production environments, they must solve issues such as operator coverage, framework compatibility, compilation optimization and developer habits. Compatibility layers can reduce migration costs, but in the long run, they still need to form their own toolchains, communities and standards. Only when the software ecosystem matures can the system-level advantages of domestic compute be truly released.

Finally, it means the business model must withstand the test. A 100,000-card cluster requires huge investment, and power, land, equipment and O&M costs are high. How to improve utilization, how to serve model vendors well, and how to form a sustainable compute leasing and token factory model are all questions that must be answered. Compute is not over once construction is complete; operational efficiency is the long-term competitiveness.

VI. Challenges and Risks Remain

Domestic compute standing at a turning point does not mean challenges have disappeared. On the supply chain side, there are still constraints in advanced process nodes, HBM, advanced packaging and EDA tools. On the ecosystem side, developer habits, operator coverage and framework compatibility still need time. On the energy side, green compute requirements will raise construction costs and operational thresholds. On the market side, it is necessary to guard against redundant construction and idle capacity, and avoid “building for building’s sake.” On the commercial side, there is tension between rising compute prices and falling model costs, and the profit margins of cloud vendors and model vendors need to be balanced.

In addition, the actual delivery progress of 100,000-card clusters, the stability of domestic GPUs in ultra-large-scale training, the scheduling efficiency of heterogeneous solutions, and the performance loss after model migration all need to be verified by real production data. Policy and industry resonance has gotten off to a good start, but the real test lies in the delivery and operations stages.

VII. Key Observation Windows Ahead

In the period ahead, several indicators deserve close attention. First, the actual delivery progress and rollout pace of JD Cloud’s 100,000-card cluster. Second, whether more leading AI cloud vendors will follow suit with large-scale deployment of domestic GPUs. Third, whether the proportion of domestic large models completing training and inference on domestic compute continues to rise. Fourth, the maturity of the domestic GPU software ecosystem, including operator libraries, framework adaptation and developer feedback. Fifth, whether green compute indicators can meet standards while scale expands. Sixth, whether compute utilization and service prices form a virtuous cycle.

If these indicators continue to improve, then the system-level scale-up of domestic compute will not just be policy language, but will become industrial reality. Over the past week, policy and industry have formed a resonance within the same window. Domestic compute is standing at a turning point from 10,000 cards to 100,000 cards, from pilots to production, and from single-point breakthroughs to system-level scale-up. A 100,000-card cluster is not the endpoint, but the starting point for domestic compute to become a main force in computing infrastructure.