AI computing power is in short supply and structurally imbalanced; China's AI stands at a crossroads.
In July 2026, Kimi K3 suspended new user subscriptions. Price hikes and purchase limits by leading model companies are becoming the norm. This is no coincidence but a concentrated reflection of the structural contradictions in China's AI computing power infrastructure. Despite the massive total scale of China's intelligent computing power, the dilemma of supply falling short of demand for high-end computing power and oversupply for low- to mid-end computing power is pushing the AI industry to a crossroads. Against this backdrop, the "supernode" is highly anticipated, and 2026 is widely recognized as the first year of domestic supernodes. But is it the key to breaking through the bottleneck, or merely a stopgap measure?
Considerable Total Computing Power, Yet Structural Imbalance
Currently, China's AI computing power infrastructure is in a critical transition period from scale leadership to quality and efficiency improvement, which is the initial stage of expansion for large-scale construction of native intelligent computing infrastructure. Che Kai from China Unicom Digital Technology Co., Ltd. stated that in recent years, with the advancement of national projects such as computing node coordination and integrated computing power networks, the construction of computing power infrastructure will maintain strong momentum. Newly built intelligent computing centers generally adopt the native AIDC architecture, achieving comprehensive improvements in performance, green operations, and security.
As of March 2026, the total scale of China's intelligent computing power has reached 1882 EFlops, with 14.45 million standard racks in use at computing centers, and over 70 major computing power corridors have been built. The computing power network has been officially included in the national "Six Networks" strategic infrastructure array. Gao Wen, an academician of the Chinese Academy of Engineering, revealed that the first phase of China's computing power network, "Computing Power Resource Aggregation," is 70% complete, with 70% of the country's computing power resources incorporated into a unified database.
Lianjie Su, Principal AI Analyst at Omdia, pointed out that China's AI computing power is currently advancing towards effective, diversified computing power and actual utilization rates. In the era of AI agents, AI computing power must not only meet the continuously growing demands for model training but also take on more low-latency, high-throughput model inference and efficient token generation workloads.
However, the other side of computing power is the cold, hard structural contradiction.
Zheng Weimin, an academician of the Chinese Academy of Engineering, previously stated that high-end computing power is currently in short supply, while low- to mid-end computing power is oversupplied. He pointed out that the supply shortage of high-end computing power directly stems from U.S. export controls on AI chips to China; whereas in the low- to mid-end market, the production capacity of domestic computing power chips is not low. The real problem is the immature software ecosystem, which makes it difficult to efficiently mobilize a large amount of hardware resources.
This means that China's computing power dilemma is not simply a matter of insufficient total volume. On the one hand, high-end GPUs are choked off by bottlenecks; on the other hand, the utilization rate of deployed domestic computing power is far below design expectations due to the lack of a mature software stack. Tang Zhehui, Co-Managing Partner of the Audit Services Market for Greater China at EY, stated that the so-called "computing power oversupply" currently mainly refers to the oversupply of low-end general-purpose computing power, invalid intelligent computing center capacity lacking actual workloads, and the phased idleness of some previous-generation training clusters after completing their iterations. High-end training computing power and inference computing power for C-end/enterprise implementation remain in a continuous state of shortage.
"Considerable total volume and structural imbalance" has indeed become the most realistic developmental backdrop of China's AI computing power infrastructure at this stage.
Tight Supply and Demand for Computing Power, Industry Returns to Rationality
On July 19, 48 hours after the launch of Kimi K3, Moonshot AI announced the suspension of membership subscriptions for new users due to user request volumes exceeding expectations and existing computing power approaching the cluster capacity limit. For ordinary users, the supply and demand relationship of computing power, originally hidden behind the models, has thus become tangible and perceptible.
Recently, several leading model companies have successively launched more advanced models, but price hikes and purchase limits are becoming the norm, with computing power shortage being one of the main triggers. Che Kai analyzed that as AI application scenarios continue to expand, the demand for computing power keeps climbing, and the supply-demand gap is gradually highlighting. However, the deeper reason lies in the fact that the AI industry is shifting from "burning money for development" to "rational operation." The shortage of computing power is forcing the computing power market into a healthy development stage, pushing the industry from extensive expansion to intensive cultivation.
Against the backdrop of computing power shortage, redundant computing has become one of the core pain points in the current AI industry. Reducing computing power scheduling and consumption by expanding cache and improving hit rates has become an industry consensus. It is also a key shift for AI enterprises to cope with the computing power shortage—from competing on computing power to competing on efficiency, which once again proves that the AI industry is moving towards maturity and rationality.
Facing the dilemma of computing power shortage, Lianjie Su pointed out that AI cloud providers are actively promoting full-stack optimization to make up for the computing power shortage with systematic capabilities, covering areas such as centralized scheduling and management of computing, storage, and networks, MaaS capability enhancement, and intelligent collaboration of heterogeneous computing power. Major chip manufacturers are also accelerating their layouts, with Moore Threads and others vigorously promoting large-scale training and inference clusters.
He pointed out that low- to mid-end computing power will gradually lean towards inference services in the medium to long term, especially distributed and heterogeneous edge-side inference systems, but this relies on deep collaboration across the entire industry. Meanwhile, various domestic computing power scheduling platforms are constantly emerging, effectively promoting cross-regional and cross-architecture computing power scheduling, and delivering computing power to the demand side.
Supernodes: Becoming a Phased Breakthrough Direction?
Computing power has become a key variable in the development of the AI era. To make up for the computing power shortfall, besides revitalizing the idle capacity of existing intelligent computing centers, computing power expansion is an important path. However, multiple factors such as chip supply, data center construction, power capacity, and network layout continue to constrain the expansion speed of the computing power industry. Facing the increasingly prominent computing power bottleneck, the answer given by the domestic industry is the "supernode."
The so-called supernode, at its core, relies on high-speed interconnect technology to deeply integrate a large number of GPUs or accelerator cards from "multiple servers" into a system that logically acts like a single "supercomputer." This is to solve the bandwidth and latency bottlenecks of cross-machine communication in traditional clusters, thereby significantly improving the utilization rate of GPUs or accelerator cards and reducing the cost per token.
2026 is widely recognized by the industry as the "first year of domestic supernodes." Huawei's Ascend 950 supernode achieves a scale of 1,024 cards and 1 EFLOPS FP8 computing power, and its 384-card supernode launched last year has been commercially deployed in over 750 sets. Alibaba Cloud officially released the Lingjun Zhenwu M890 supernode instance. Sugon released the country's first 100,000-card AI super cluster, accelerating the formation of the domestic supernode industry ecosystem.
However, the industry's attitude towards whether supernodes can fundamentally solve the computing power shortage problem tends to be cautious. Che Kai believes that supernodes are a path to make up for the lack of high-end computing power and promote the domestication of computing power, and are currently moving from proof of concept to large-scale commercialization. But in essence, supernodes still rely on the stacking of low-end computing power, making up for the performance shortcomings of a single card through cluster-level system tuning, which is not a long-term solution to the computing power shortage.
He also pointed out that to achieve the same scale of computing power, the procurement cost and operating cost (power) of mainstream domestic supernodes are basically about twice those of mainstream foreign products, and the cabinet occupation is even up to 10 times that of foreign products. Although this "swarm tactic" can temporarily alleviate the urgent need for computing power, the more long-term problem lies in: when domestic high-end computing power equipment is truly introduced in the later stage, the infrastructure (power, cooling, etc.) previously built to support supernodes will be difficult to retrofit and match, potentially causing greater resource waste.
Lianjie Su also expressed that supernodes can alleviate the computing power shortage in the short term, but cannot cure the root cause. After all, they are essentially engineered products aimed at improving the efficiency of computing power clusters. The true computing power shortage still relies on the scaling of high-end chip manufacturing and breakthroughs in advanced packaging.
Five-Year Expectation: Domestic Computing Power Steps Out of Anxiety
Looking ahead to the next five years, Che Kai predicts that under the guidance of national policies and the joint efforts of industry synergy, domestic computing power may completely bid farewell to the involution of single-point chip parameters, and shift its focus to system efficiency, energy synergy, and resource scheduling, achieving a strategic upgrade from "single-point technology competition" to "overall system competition." The core of future computing power competition will focus on the implementation capabilities of the national integrated computing power network and computing-power-electricity synergy, competing for the lowest cost per token and optimal computing power efficiency.
"At that time, domestic computing power will step out of the anxiety of passive catching-up and stride into a systematic era that is more efficient, greener, more collaborative, and more characteristic of China," said Che Kai.
Lianjie Su added that over the next five years, domestic AI training and inference chips will continue to iterate and upgrade, and new industry development trends will accelerate. Domestic computing power manufacturers need to accelerate the pace of ecosystem co-construction, promote the commercialization of supernodes, continuously optimize the integrated synergy model of computing power and electricity, cultivate a comprehensive industry application and developer ecosystem, and at the same time increase investment and promotion efforts in open-source models and inference frameworks, comprehensively consolidating the foundation of the domestic computing power industry.
Editor-in-Charge / Layout: Zhu Wenfeng
Proofreader: Mei Yaxin
Supervisor: Liu Qicheng