The "Breakthrough Moment" for Domestic Computing Power: From Chips to Supernodes, China's Computing Foundation Leaps Forward Across the Board

If the surge in computing power rental prices was the topic that drew the most capital market attention in the computing power sector over the past week, then the collective breakthrough in domestic computing power infrastructure is another main thread with the greatest industrial depth this week. In just over ten days, Alibaba's T-Head released the Zhenwu V900 chip with triple the computing power, Huawei launched the world's first Ascend 960 supernode using NPO technology, Sugon released a fully self-developed 100,000-card AI supercomputing system, and China Mobile's Beijing node deployed the Ascend 950DT thousand-card supernode. At the same time, on September 28, the State-owned Assets Supervision and Administration Commission of the State Council clearly stated that during the "15th Five-Year Plan" period, central enterprises should continue to increase investment in the computing power network and deepen computing-power-electricity coordination and computing-network integration. From chips to supernodes, from clusters to policy, domestic computing power is undergoing a systemic breakthrough driven from the "roots." This breakthrough is not only about catching up in single-point performance, but also about the comprehensive restructuring of system architecture, software ecosystems, energy coordination, and industrial policy. As global AI computing power competition enters a white-hot stage, China's computing power industry is attempting to forge a differentiated, self-reliant, and systematic development path.

I. Chip Breakthrough: The "Awakening of Hard Power" of Zhenwu V900 and Ascend 950

The competitive landscape of domestic AI chips is undergoing a qualitative change. For a long time, domestic AI chips faced the dilemma of being "usable but not good enough": single-card performance lagged behind international leaders, the software ecosystem was incomplete, and cluster expansion capability was limited, causing large model training and inference to rely heavily on imported GPUs. However, this situation is being rapidly broken.

On September 22, at the 2026 Hangzhou Yunqi Conference, Alibaba's T-Head released Zhenwu V900, currently China's most powerful self-developed AI chip, with computing power tripled compared to the previous generation Zhenwu M890, a single cluster scalable to 500,000 cards, planned for mass production and market launch in the first quarter of 2027. Alibaba also announced that the Qwen4 large model based on the new-generation architecture is already in training, and future versions such as Qwen4.5 and Qwen5 will expand to 5 trillion to 10 trillion parameters, and proposed a goal for Alibaba Cloud to operate more than 20 GW of global data center capacity by 2032. This means Alibaba is not only pushing forward on the chip side, but has also formed a closed-loop layout of "chip-model-cloud" across the model and data center sides. The 500,000-card cluster expansion capability of Zhenwu V900 provides the hardware foundation for training future ten-trillion-parameter models; the 20 GW data center target demonstrates Alibaba's strategic judgment on long-term AI computing power demand.

On Huawei's side, Ascend 950 intelligent computing cluster services were commercialized domestically at the end of September and will expand to overseas markets on November 30. At the computing power level, the service is based on a full-link observable five-level rapid recovery mechanism, achieving stable cloud training for 40 days and fault recovery within 10 minutes. For large model training, stability and fault recovery capabilities are crucial. A training task often requires thousands of cards to run continuously for weeks or even months; any single point of failure can cause training interruption, progress rollback, and cost surges. Huawei's five-level rapid recovery mechanism compresses fault recovery time to within 10 minutes, significantly improving cluster availability and reducing training costs.

More notably, Alibaba Cloud's Lingjun Zhenwu M890 supernode instance GP9A has already begun providing services externally. This is China's first supernode-form computing power capable of stably running large models with over 2 trillion parameters; a single-node instance can support a 100,000-card cluster and has stably run frontier large models such as Qwen3.8-Max and Kimi K3. This means domestic supernodes have not only reached new heights in hardware specifications, but have also been verified in actual large model operation. Stably running a 2-trillion-parameter model places extremely high demands on memory bandwidth, interconnect latency, scheduling systems, and cooling capacity. Achieving this shows that domestic computing power system integration capability has reached a new level.

The successive deployment of two flagship chips marks a key leap for domestic AI chips from "usable" to "good to use." T-Head publicly disclosed for the first time the update plan for its Yitian series server CPUs, completing a full-stack self-developed product layout in computing, storage, and networking. This full-link tuning capability from chips to the system layer will significantly improve software-hardware synergy efficiency, reduce inference costs, and provide a solid foundation for computing power demand in the Agent era. In the past, domestic chip vendors often only provided chips, while system integration, software adaptation, and cluster O&M had to be solved by downstream vendors themselves, resulting in low overall efficiency. Now, leading vendors are beginning to provide full-stack solutions from chips, servers, and networks to cloud services, greatly deepening software-hardware synergy. This not only helps improve single-card utilization but also reduces total cost of ownership, making domestic computing power more attractive in commercial competition.

II. The Supernode Revolution: Cluster Leap from Thousand Cards to Million Cards

Supernodes are the most technologically rich breakthrough direction for domestic computing power this week. As large models move toward the ten-trillion scale, traditional server cluster architectures can no longer meet the needs of efficient training and inference. In traditional clusters, GPUs are interconnected via PCIe or Ethernet, with limited bandwidth and relatively high latency. When model scale expands to trillions of parameters, communication bottlenecks severely drag down training efficiency. Supernodes, through system-level architectural innovation, interconnect thousands or even tens of thousands of accelerator cards with extremely low latency, becoming an inevitable choice for ultra-large-scale AI infrastructure.

On September 17, at the Huawei Connect 2026 conference, Huawei released the world's first supernode using NPO (near-packaged optics) technology—the Ascend 960 supernode. This supernode adopts an orthogonal architecture and full liquid cooling design, achieves unified memory addressing based on Lingqu, can scale a single node to 4,096 cards, and provides 8 EFLOPS FP8 and 16 EFLOPS FP4 computing power. NPO technology places the optical engine near the packaging substrate, shortening electrical signal transmission distance and reducing power consumption and latency—a key technology direction for ultra-large-scale cluster interconnection. The full liquid cooling design solves the heat dissipation challenge brought by high-density computing power, enabling stable operation of a single node with 4,096 cards at high power density.

Huawei also launched Hi-ONE, the industry's first mass-produced NPO optical engine, with a total transmission capacity of 7.2T. It replaces the 48,000 800G optical modules required by traditional solutions with 5,500 Hi-ONE optical engines, reducing power consumption by over 550 kilowatts, doubling system fault-free operating time, and achieving 99.8% system availability. This comparison is striking: traditional solutions require 48,000 800G optical modules, while Hi-ONE requires only 5,500. This not only greatly reduces the number of optical modules but also lowers power consumption and failure points. For 10,000-card clusters, the sheer number of optical modules increases failure probability; reducing the number of optical modules means higher reliability and lower O&M costs. 99.8% system availability means less than 18 hours of unavailability per year, which is crucial for large model training and inference services.

More strategically significant is cluster expansion capability. Multiple Ascend 960 supernodes connected via Lingqu network or RoCE, using two-layer CLOS four-plane networking, can reach a maximum cluster scale of 512,000 cards; combined with multi-track topology technology, it can support up to a 1 million-card Ascend supernode cluster. This means domestic supernodes already possess the hardware foundation to support training of ten-trillion-parameter large models. What does a 1 million-card cluster mean? Currently, the world's most advanced AI clusters are mostly at the tens of thousands to hundreds of thousands of cards level. A million-card cluster would allow a single training run to cover a larger parameter scale and dataset, providing computing power assurance for next-generation foundation models. Of course, a million-card cluster also brings enormous challenges: network topology, fault isolation, task scheduling, power supply, and cooling systems all require systemic innovation. Huawei's multi-track topology and two-layer CLOS four-plane networking are precisely aimed at solving the communication bottlenecks and scalability issues of ultra-large-scale clusters.

The deployment of supernodes does not stop at product launches. On September 29, China Mobile Beijing Company, together with China Mobile Cloud Company, announced that the Huawei Ascend 950DT thousand-card supernode officially landed at the Mobile Cloud Beijing node in the Changping District Information Port Data Center. This thousand-card supernode is based on Huawei's supernode architecture, with a single node of 1,024 cards delivering 1 EFLOPS (FP8), supporting 256TB unified memory and microsecond-level latency, and capable of supporting high-level needs such as large model training and inference and agent development. This is an important sign of domestic supernodes moving from product launch to actual deployment. Operators possess enormous data center resources, network resources, and customer resources. Their entry into supernode deployment means domestic supernodes are beginning to enter the stage of large-scale commercialization. The landing at China Mobile's Beijing node not only verified the engineering capability of Ascend 950DT but also provided a reference case for subsequent procurement of domestic supernodes by more operators and industry customers.

III. Computing Power Ecosystem: The "Chinese Path" of Open Source and Full-Stack Synergy

The breakthrough in domestic computing power is not only about hardware performance improvement, but also about the reconstruction of the ecosystem. For a long time, NVIDIA's CUDA ecosystem constituted an extremely high migration barrier. Developers were accustomed to the CUDA toolchain, and model code, operator libraries, and tuning experience were all built around CUDA. Even if domestic chips approached in hardware performance, incomplete software ecosystems made it difficult to attract developers to migrate. Therefore, ecosystem building became the key to whether domestic computing power could truly establish itself.

Wang Tao, Huawei's Vice Chairman and Rotating Chairman, clearly stated at the Connect conference that the core of Huawei's AI strategy is computing power, insisting on building competitiveness around "supernode + cluster" through system architecture innovation, building a solid computing foundation for China, building an open-source computing ecosystem, and creating new choices for the world. This statement clearly conveys Huawei's strategic path: not pursuing simple surpassing of single-card performance, but forming differentiated competitive advantages through system-level innovation in supernodes and clusters, combined with an open-source ecosystem.

The key progress in ecosystem building lies in the comprehensive open-sourcing of the CANN computing framework. CANN has entered normalized open-source community operation. From "usable" to "easy to use," Ascend has crossed the ecosystem inflection point. The significance of this shift is that open source lowers developers' migration costs, enabling more mainstream large models to undergo native training on Ascend, thereby forming a positive cycle from chip to framework to model. In the past, developers migrating to Ascend needed extensive adaptation work. Now, with CANN open-sourced and community operations normalized, support for Ascend in mainstream frameworks such as PyTorch and TensorFlow has become more complete, greatly reducing model migration costs. As more models are natively trained on Ascend, the Ascend ecosystem's operator libraries, tuning tools, and best practices will also become increasingly rich, forming a positive cycle.

At the system level, Huawei introduced the supernode architecture into general computing systems, fully upgrading the Kunpeng supernode. Based on Lingqu all-optical networking, it supports up to 4,096 nodes and forms a 256TB unified memory pool, significantly improving Agent sandbox density, startup speed, and vector retrieval performance. It also released the OceanStor M900 cluster, an L3.5-layer PB-level KV cache solution based on the Lingqu UnifiedBus, supporting "one-hop direct connection" technology to optimize data access efficiency. These system-level innovations target the special needs of AI Agent and inference scenarios. Agent sandboxes require fast startup, isolated operation, and efficient retrieval; the Kunpeng supernode's unified memory pool and all-optical networking can significantly improve the efficiency of these operations. The PB-level KV cache and one-hop direct connection technology of the OceanStor M900 cluster target the KV Cache access bottleneck in large model inference, improving inference throughput and response speed.

On Alibaba's side, based on the pioneering fully modular design architecture CUBE 5.0, construction costs are reduced by more than 10%, and computing power density per unit area reaches 5 to 10 times that of the previous generation. The five major systems are manufactured in parallel in factories and pre-tuned into containers, achieving modular delivery of computing power infrastructure. The significance of modular delivery lies in shortening construction cycles, reducing on-site construction difficulty, and improving deployment consistency. Traditional data center construction has long cycles and complex on-site engineering, while containerized modular delivery can complete most manufacturing and pre-tuning work in factories, requiring only hoisting, water/electricity, and network connection on site before use. This is crucial for rapidly expanding computing power scale and responding to the AI demand explosion.

IV. Computing-Power-Electricity Coordination: The Underlying Constraint and Policy Support for Computing Power Expansion

The large-scale expansion of domestic computing power cannot be separated from the dual support of electricity and policy. Computing power infrastructure is a typical high-energy-consumption industry. A 100,000-card AI cluster can consume tens of megawatts or even hundreds of megawatts, equivalent to the electricity load of a small or medium-sized city. If power supply is unstable and electricity prices are too high, computing power operating costs will be difficult to control. Therefore, computing-power-electricity coordination has become key to the sustainable development of the computing power industry.

On September 28, Dai Xi, Director of the Planning and Development Bureau of SASAC, stated at a press conference of the State Council Information Office that during the "15th Five-Year Plan" period, central enterprises should continue to increase investment in the computing power network, rationally layout and orderly build eco-friendly computing centers; deepen computing-power-electricity coordination and computing-network integration, accelerate the implementation of more green power direct connection projects, and promote two-way empowerment between AI and energy. This policy statement conveys two key signals. First, the "computing power network" has been elevated to the infrastructure level alongside power grids and transportation networks, and central enterprises will become the main force in building the computing power network. Second, "computing-power-electricity coordination" has been clearly identified as the key direction for the next stage, meaning computing power expansion will no longer be simple GPU stacking, but will require systematic coordination with energy supply.

Continuous policy support provides institutional guarantees for the long-term expansion of domestic computing power. At the same time, China's first storage-computing integration project has officially started in Baotou, Inner Mongolia. Through coordinated construction of energy storage stations and distributed intelligent computing centers, it builds a green industrial closed loop of "input electricity, output tokens." Computing power infrastructure is evolving from a single computing center to a comprehensive system of "computing power + energy + network." The significance of the storage-computing integration project is that it attempts to solve the spatiotemporal mismatch between computing power and electricity. Wind power, photovoltaic and other new energy generation are intermittent, while AI training and inference require stable electricity. Through energy storage stations, electricity can be stored during low-price periods or peak new energy generation periods and released during computing power peaks; through distributed intelligent computing centers, green power can be consumed locally, reducing transmission losses. If this model can be replicated at scale, it will significantly improve the energy structure of computing centers, reducing carbon emissions and electricity costs.

V. Industrial Chain Resonance: Comprehensive Benefits from Chips to Liquid Cooling

The breakthrough in domestic computing power is driving resonance across the entire industrial chain. In the liquid cooling sector, CICC estimates that the global intelligent computing center liquid cooling market will be about 114.7 billion yuan in 2026, a year-on-year increase of 273%. As GPU power density continues to rise, traditional air cooling can no longer meet heat dissipation needs, and liquid cooling has become an inevitable choice for high-density computing centers. Liquid cooling not only improves heat dissipation efficiency but also reduces PUE and energy consumption. Sugon's released Sugon 8000 Dengfeng cluster is China's first fully self-controllable 100,000-card AI supercomputing system, using immersion phase-change liquid cooling technology, representing the frontier exploration of domestic supercomputing in cooling technology. Nanwang Digital's high-efficiency liquid cooling computing power series products have a data center immersion liquid cooling system PUE as low as 1.086, and a cold plate liquid cooling system PUE of 1.133, reducing energy consumption by more than 30% compared with water cooling/air cooling systems. The lower the PUE, the smaller the proportion of electricity used for cooling, and the higher the energy efficiency of the computing center. Immersion liquid cooling PUE as low as 1.086 is already close to the theoretical limit, demonstrating the maturity of domestic liquid cooling technology.

In the supercomputing field, "Bagua Furnace 2" from the High Performance Computing Institute of Tsinghua University's Department of Computer Science and Technology was awarded the Best Academic Paper at the 2026 China Computing Conference, demonstrating the continued exploration of domestic supercomputing at the academic frontier. Academic research is the source of industrial innovation. Breakthroughs in the academic field of domestic supercomputing will provide theoretical support in algorithms, scheduling, and fault tolerance for industrial-grade computing power systems.

From the investment side, large orders in the computing power industrial chain are frequent. Xingyun Technology's wholly-owned subsidiary signed a 5-year intelligent computing resource leasing service contract with a total tax-inclusive amount of 872 million yuan; Sinnet plans to add 1.265 billion yuan in investment to the Horinger Intelligent Computing Center project, increasing the total investment to 2.5 billion yuan; Xunce Technology's shareholders' meeting approved a proposal to build an AI inference and computing center with no more than 12 billion yuan. These large orders and investment plans indicate that industrial capital remains optimistic about the long-term demand for computing power infrastructure. Computing power rental, data center construction, AI inference centers, and other links are all attracting large amounts of capital. At the same time, upstream links such as servers, optical modules, network equipment, power supplies, and temperature control will also benefit from the growth in computing power capital expenditures.

Conclusion: The "System-Level Breakthrough" of Domestic Computing Power

From Zhenwu V900 to Ascend 960, from thousand-card supernodes to million-card clusters, from self-developed chips to open-source ecosystems, domestic computing power is undergoing a systemic breakthrough from point to surface. The core of this breakthrough is not the surpassing of single-point performance, but innovation in system architecture—through supernode interconnection, full-stack synergy, and computing-power-electricity integration, it has forged a differentiated technological path under process-constrained conditions.

As Xinhua Net commented, domestic computing power is shifting from catching up at the application layer to deep breakthroughs in technology, standards, and models. A wall-breaking transformation driven from the "roots" has already begun. As the computing power network is incorporated into the national new infrastructure system, as central enterprises become the main force in computing power investment, and as the open-source ecosystem crosses from "usable" to "easy to use," the ceiling of domestic computing power is being continuously raised. This is not only progress at the technical level, but also a substantial enhancement of China's voice in the global AI computing power competition landscape. In the future, as more domestic chips, supernodes, clusters, and ecosystem achievements land, China's computing power industry is expected to transform from a "follower" to a "parallel runner," and become a "leader" in some areas. And all of this has only just begun.