Amid the explosive growth of AI, we are witnessing an unprecedented siphoning of computing power.
In 2020, we considered the hundred-billion-parameter scale as the hallmark of large models. Just a few years later, top-tier models have advanced to the 10-trillion-parameter scale and are highly likely to evolve toward 100 trillion parameters. AI agents can now handle complex, long-context, cross-application tasks; embodied intelligence and autonomous driving vehicles are bringing AI into the physical world.
All these visible breakthroughs are built on the same premise: abundant AI computing power. The rapid development of AI technology continuously tests the scale, performance, and reliability of AI infrastructure. It is important to note that AI technology does not grow in a stable, linear manner; instead, it often explodes geometrically, pushing the supply relationship of AI computing power to the limit.
How should we respond to this situation? When facing this question, both inside and outside the computing industry today will highly consistently think of one name—"Supernode".
In 2025, Huawei made the Supernode widely known. Next, what Huawei needs to do is to prove that Supernode + cluster is a sustainable and scalable computing power channel. At the Huawei Connect conference, which opened on September 17, the evolutionary path of the Supernode was proven to be true.
During the conference, Wang Tao, Vice Chairman and Rotating Chairman of Huawei, delivered a keynote speech titled "Intelligent New Future, Building the Silicon-Based Fertile Soil of the Intelligent World," and released the world's first Supernode adopting NPO technology—the Ascend 960 Supernode. During the conference, Naojiti conducted an exclusive interview with Wang Tao, conducting in-depth discussions on Huawei's AI strategy and the development goal of building a solid computing power foundation for China.
If in 2025 Huawei opened up a possibility and opportunity for the computing industry with the Supernode, then today, Huawei has led the Supernode onto an evolutionary path of stronger capabilities and more mature products through the technological breakthroughs of UnifiedBus + NPO.
This is the story of the Supernode, and also the story of Huawei turning around to excavate the sea and build ships when faced with a computing power blockade and having no way to retreat.
If computing power is an irrigation canal, the downstream will be at the mercy of the upstream; only when computing power becomes an ocean will fairness and abundance be the background color of the intelligent world.
More and more people are beginning to believe that the Supernode is that ocean of computing power.
No Computing Power Canal Can Quench the Thirst of the AI Era
In the Agentic AI era with the comprehensive rise of AI agents, users' demand for and positioning of AI computing power are undergoing a comprehensive reconstruction. On the one hand, the evolution of model capabilities makes single-chip computing power stretched thin. Prominent features of the new generation of models, such as large-scale MoE, complex context sequences, and long-text memory, all require large-scale parallel computing to ensure model training and inference effects. On the other hand, under the huge computing power costs, the industry has begun to pursue higher computing efficiency, while the inference overhead of single-chip computing power is enormous, which also does not meet the objective needs of AI commercialization development. On this basis, the traditional clustered AI computing power relying on the stacking of computing power scale is also facing the dilemma of diminishing returns. As the computing power unit increases, the wasted computing power in the cluster becomes more massive, and after the computing power scale reaches a certain level, the cost will be unbearable.
On the other hand, China's AI also faces a special computing power challenge, which is the difficulty of obtaining high-end AI computing power from external sources. Meanwhile, domestic semiconductor capabilities are still making continuous breakthroughs. Competing on single-chip computing power means that China will lose the initiative to master advanced AI technologies from the computing power foundation. It is not only Huawei that has no way to retreat on the issue of computing power; the entire Chinese AI has no way to retreat in this game.
Therefore, whether it is the objective requirements of technology and industry trends, or the subjective needs caused by the special environment, integrating multiple chips into a more compact, economical, and easily accessible form of computing power is the inevitable choice for China's AI in the face of the computing power dilemma. If AI computing power remains a small stream, it is destined to be unable to nourish the ambitious AI fertile fields.
Only by turning small streams into rivers and rivers into oceans can we find the true way out for China's AI computing power demand. Since the single-card solution is not viable, we must compete on overall capabilities; we must use multi-path technological innovation to overcome the technological disadvantage of a single path.
When there is no way to retreat, we must seek a victory.
Computing Power into an Ocean: The Supernode is the Inevitable Choice
In Huawei's view, the computing power siege is completely possible to be broken. This breakthrough is not a breakthrough of a single technology, but the use of technological advantages accumulated over 30 years to form alliances and achieve architectural innovation in all aspects, meeting the computing power needs of China's AI. The Supernode is one of the ultimate results of such efforts.
Since Huawei launched the Supernode in 2025, numerous "Supernode" have sprung up like an explosion in the industry. However, the technical cores of these products vary greatly, the standards are very chaotic, and there is a widespread conceptual confusion between Supernode and clusters. In order to clarify the source and standardize industry standards, during the 2026 World Artificial Intelligence Conference (WAIC), Peng Cheng Laboratory (PCL) and the Global Computing Consortium (GCC) jointly took the lead in releasing the "Supernode Definition and Practice White Paper". It clearly defines: A Supernode is a computing system physically composed of multiple computing nodes closely connected through efficient interconnection protocols, possessing the capability of unified memory addressing across physical nodes, and logically possessing the characteristics of "a single computer". The Supernode has three core technical features, including: "memory semantics" and "unified memory addressing", ultra-low latency, and ultra-large bandwidth.
From the perspective of industrial value, the Supernode is also vastly different from the traditional IP-based interconnection cluster solutions. The model floating-point availability rate of the former is greatly improved compared to the latter, which can adapt to more complex model training and inference scenarios, adapt to the computing characteristics of "tight coupling and synchronous parallelism" of large models, and greatly reduce comprehensive cost expenditures, showing a significant characteristic that the larger the scale of the Supernode, the more obvious the comprehensive benefits. Therefore, the industry generally believes that the Supernode has become an inevitable choice for building ultra-large-scale AI infrastructure.
Under the clear standard definition, we can find that the Ascend Supernode is currently the only Supernode in China that can be commercially used at scale. As of now, the Ascend 910C Supernode has been commercially deployed in over 1,000 sets, applied in industries such as the Internet, telecommunications operators, finance, education, healthcare, transportation, and manufacturing, and is the only Supernode in China that has trained a SOTA model.
During this year's WAIC, the Ascend 950 Supernode made its first real-machine debut, and during Huawei Connect 2026, it was announced to enter the stage of large-scale commercial use.
At the same time, Huawei also released the industry's first NPO-based Supernode at this conference. Based on "UnifiedBus + Hi-ONE", the Ascend 960 Supernode built by Huawei can be scaled up to 4,096 cards, which can effectively match the training and inference needs of 10-trillion-parameter large models.
Not only can it achieve the only large-scale commercially used Supernode, but it can also continuously upgrade the Supernode. How exactly did Huawei achieve all this?
The answer lies in the fact that the AI computing power path of the Supernode is not determined by a single capability or component level behind it, but is a comprehensive test of multidisciplinary knowledge, multi-field technology, and systematic engineering capabilities. NPO, UnifiedBus network, and other key technologies supporting the evolution of the Ascend 960 Supernode, conceal Huawei's long-term accumulation in various digital fields and the integration of mastery at critical moments.
Wang Tao mentioned in the interview that Huawei's Supernode cluster solution has a leading advantage. In fields such as communication technology, IP, and optics, especially optical modules, computing power, and systems, Huawei has long-term and large-scale R&D investment in this field. We have the best experts in every field, gathering all these capabilities into systematic engineering capabilities... Huawei's R&D investment is like the "Van Fleet ammunition volume, breaking through a single gap in the city wall".
Under the compulsion of necessity, Huawei urgently solved this problem, and by a stroke of coincidence, turned computing power into an ocean.
Light Chasers: Lighting the Beacon of NPO
The evolution of the Ascend 960 Supernode has a key clue coming from light. Recently, optical modules have received unexpected attention. The reason is that more and more people realize that AI computing power is not only a problem of semiconductors and servers, but a deeper test of processes and technologies.
In large model training, AI computing power units need to frequently perform collective communication operations such as All-Reduce and All-to-All. The data transmission volume shows super-linear growth with the increase of computing power scale. During this process, a large amount of time is consumed in data communication rather than actual computation. The bandwidth, latency, and reliability of optical modules directly determine the speedup ratio and training efficiency of large-scale AI computing power, which is a crucial part of the computing power infrastructure.
Facing the 10-trillion-parameter model training cluster, traditional pluggable optical modules are large in number and low in reliability. Due to too many deployed devices, they are also not suitable for the highly compact Supernode solution. Therefore, relying on advanced optical technology accumulated over many years, Huawei took the lead in launching the world's first NPO optical engine, Hi-ONE, within the Supernode system. This is the industry's first mass-produced NPO product, and also the NPO product with the largest transmission capacity currently, with a total capacity reaching 7.2T. At the same time, it is also the first NPO product to achieve a built-in light source.
By using the NPO solution, the Ascend 960 Supernode has achieved performance improvements. At the scale of a 4,096-card Supernode, power consumption is reduced by over 550 kilowatts, the system's mean time between failures is doubled, and system availability reaches 99.8%. Overall, NPO is currently the most adaptable optical module technology path for the Supernode.
Wang Tao believes that as a single Supernode cluster becomes larger and larger, moving from traditional pluggable modules to NPO is the only way to go. The so-called 7.2T of an NPO now integrates 36 200G lanes into one module. A small Hi-ONE module replaces nine 800G optical modules.
The reason behind being able to take the lead in implementing NPO technology is that Huawei possesses the world's most advanced optical communication technology and the richest optical industry experience. Over a long period, Huawei has always played the role of a "light chaser". This exploration and accumulation have not only been proven in the fixed network communication business but have also been integrated into the ocean of computing power at this critical historical node, becoming an inextinguishable "beacon of light" within it.
There is reason to believe that when Huawei began deploying optical communication and optical module technologies, it could not have foreseen the day when they would be needed to solve AI computing power problems. But that day has arrived.
Only by making all preparations can we cope with all storms.
Breaking the Deadlock through Connectivity: Opening Up the Computing Power Sea Area
In large-scale AI computing power, there has always been a "scissors difference" problem between computing power and the network. That is, the larger the computing power scale, the more computing power is wasted in communication, which then evolves into a vicious cycle that plagues large-scale AI computing power. For example, in the training of large language models, the training failure rate caused by network problems may reach 20%, and a single network interruption may cause computing results of hours or even days to be wasted. Improving the MFU of large-scale computing power has become a world-class problem.
The core reason why Huawei was able to build the only commercially used Supernode in China is that Huawei found a way to rely on connectivity capabilities to break through the "scissors difference" of large-scale AI computing power. The key achievement in this is the UnifiedBus. UnifiedBus is Huawei's self-developed Supernode interconnection protocol and architecture system, and it is also one of the core technical foundations of the Ascend and Kunpeng Supernode. It is not a data center network in the traditional sense, but a unified interconnection bus that connects the three layers of chips, complete machines, and cabinets. Perhaps it can be said that UnifiedBus and Hi-ONE together constitute the neurons of the Ascend 960 Supernode.
In March 2025, the Supernode equipped with UnifiedBus was officially commercially delivered. Subsequently, at Huawei Connect 2025 in September, UnifiedBus 2.0 was officially released, and all technical specifications were opened to the outside world, becoming the first open Supernode interconnection standard in China.
In this Huawei Connect conference, UnifiedBus ushered in a deep integration with NPO technology, opening a new page for the global exploration of Supernode. At the same time, Huawei also released layered and graded UnifiedBus interconnection devices, covering three levels of interconnection: intra-cabinet, cross-cabinet, and cluster, supporting elastic expansion from a single cabinet to a million-card cluster. Relying on UnifiedBus, multiple Supernode can build larger-scale Supernode clusters, which consist of the Ascend 960 Supernode, Kunpeng 950 Supernode, OceanStor M900 memory storage, and Xinghe UBG switches. It can achieve diverse computing power collaboration and graded resource pooling, thereby realizing the construction of computing power scales from tens of thousands of cards to hundreds of thousands of cards, or even millions of cards.
The successful commercialization and opening up of UnifiedBus prove the feasibility of Huawei's integrated innovation of network communication capabilities and computing capabilities. Huawei, which started with connectivity capabilities, has once again demonstrated the importance of this fundamental capability. In Huawei's view, the Supernode is not the accumulation of servers and cabinets, but a complex software and hardware system. And interconnection capability is the prerequisite for hardware to release performance, software to exert effects, and ultimately achieve systematic availability. From the Tau Scaling Law to the Supernode, interconnection has always been the key to opening the door to the intelligent world.
Wang Tao proposed that Huawei hopes to connect all computing power by building a new generation of communication networks.
When the small stream of computing power is connected into an ocean, the inherent rules and limitations will naturally become invisible.
Ecological Inflection Point and Scenario Traction: The Ship Sailing to the Intelligent World
Relying on Supernode + cluster to build a new generation of computing power infrastructure not only requires strong technology, engineering, and capabilities but also needs the support of the ecosystem and the traction of scenarios. The ecosystem is the people who share the same boat, and the scenarios are the people who use its value. The Supernode is not only the close connection of computing power units but also the closer connection between people.
In terms of the ecosystem, Huawei continues to deeply cultivate CANN, the foundation of the Ascend ecosystem. On December 30, 2025, CANN completed the launch of open-source code as scheduled. Currently, monthly active developers have exceeded 5,200, with an average daily increase of 30,000 lines of open-source code, supporting the pre-training of 40 native models. It has become the most dynamic AI open-source community in China, supporting the native training of mainstream large models.
Ascend has reached deep cooperation with over 90 open-source communities such as PyTorch, vLLM, SGLang, Triton, and Tilelang. Ascend has become a computing power platform that can be directly selected and installed on the official PyTorch website, and Huawei has also become the first Chinese hardware manufacturer to achieve this goal. Wang Tao mentioned that the Ascend ecosystem has crossed the inflection point.
In terms of application scenarios, Huawei hopes to comprehensively accelerate AI entering the edge, achieving self-sufficiency and autonomy of edge computing power on mobile phones through Kirin + Ascend. In the automotive sector, it is expected that by 2030, the cloud training computing power of Huawei Qiankun autonomous driving will exceed 100 EFLOPS, the vehicle-side computing power will exceed 3,000 TFLOPS, and the parameter scale of vehicle-side models will exceed 20B. At the same time, Huawei will also create edge-side lightweight AI cores for massive lightweight intelligent terminals, building a computing power foundation for the intelligence of all things.
UnifiedBus + NPO realizes the technological upgrade of the Supernode, Ascend crosses the inflection point, and various intelligent scenarios continue to provide traction. These combined factors have promoted Huawei's effective construction of the computing power foundation.
The seemingly unsolvable computing power cliff of that year now seems to have become a good port that can be navigated.
The Darkness Before Dawn, the Victory When There is No Way to Retreat
Behind the Ascend 960 Supernode is an unprecedented technological breakthrough. It integrates many technologies such as communication, optics, computing, and AI, and the innovative synergy of various systems is intertwined, ultimately completing a symphony-like breakthrough. This is not a head-on technical deadlock, but a strategic expedition where everyone sets out separately, acts independently, and meets at a certain node.
I'm afraid there is no company in the world that can aggregate such diversified comprehensive capabilities like Huawei. But at the same time, no company has ever faced such strict and desperate technological blockades. Even now, Ascend still has many problems, such as insufficient supply and a low global share. But it must be said that relying on Supernode + cluster, Huawei has completed a victory when there was no way to retreat: even though the semiconductor process has not been improved, Huawei has found an alternative path for computing power upgrades and found diverse innovative solutions.
If there is no path in the mountains, go find a path in the sea; even this sea can be excavated by your own hands.
Wang Tao said in an interview with the media that the darkness before dawn must be broken through, and the products must be made well so that users feel that choosing Huawei is worthwhile and correct.
In the end, time will prove which path is the path to victory.