Author | Li Shuiqing Editor | Yun Peng
Reporting from Shanghai on September 17, Zhidx noted that yesterday, an internal interview with Guo Ping, Chairman of the Supervisory Board of Huawei, regarding "Huawei's ICT and computing goal is to become 'NVIDIA'" just went viral. Today, Huawei unveiled its long-awaited computing "blockbuster". At the newly opened Huawei Connect 2026, Huawei made a spectacular debut of its UnifiedBus interconnection architecture and SuperNode cluster solutions, launching a series of new AI infrastructure products including the Ascend 960DT chip, Ascend 960 SuperNode, Kunpeng 950 SuperNode upgrade, OceanStor M900 memory storage, and UnifiedBus all-optical switching equipment.
Wang Tao, Huawei's Vice Chairman and Rotating Chairman, announced on-site that the Ascend 960DT chip will be ready ahead of schedule. Supporting 2 PFLOPS (FP8) and 4 PFLOPS (FP4) computing power, it achieves doubled performance compared to the previous generation. Wang noted that the R&D of this chip has exceeded expectations and is expected to be ready in Q1 2027.
The early readiness of Huawei's Ascend series chips is inseparable from the Tau Scaling Law. Relying on this innovative direction of "time miniaturization" substitution, Huawei's Ascend series chips will continue to adhere to the evolution cadence of one generation per year. In 2028 and 2029, Huawei will successively launch the Ascend 970 and 980 chips to achieve continued doubling of computing performance.
Wang also revealed that Huawei's Ascend 910C SuperNodes have been deployed in over 1,000 sets, and the Ascend 950 SuperNode has begun large-scale commercial deployment.
The deeper focus of this release—UnifiedBus UnifiedBus—is the unified interconnection bus protocol for Huawei's entire computing foundation. In his speech, Yang Chaobin, Managing Director and CEO of Huawei's ICT BG, mentioned that a single cluster based on UnifiedBus can support 1 million AI processors, and the Model Flops Utilization (MFU) of a 100,000-card cluster has increased from 20% to 35%.
Wang stated, "The speed of AI transformation exceeds any technological revolution in history, and the intelligent world is accelerating its arrival." Powerful AI infrastructure is the solid foundation of the intelligent world. To this end, Huawei insists on focusing on building the silicon-based black soil for the intelligent world, including seven major strategies such as "the core of the AI strategy is computing power, adhering to hardware monetization" and "SuperNode + cluster, building a solid computing foundation for China".
Regarding the ecosystem, Wang announced that Huawei Ascend has crossed the ecosystem inflection point. The CANN community's monthly active developers have exceeded 5,000, with external developers surpassing internal developers for the first time, accounting for 61% of the total developer base. The global developer base for Huawei's Kunpeng ecosystem has exceeded 4.16 million, supporting over 560 global open-source projects. The installation volume of openEuler has also surpassed 20 million sets, ranking first in the market share of server operating systems in China.
01.Breaking the Communication Wall of 100,000-Card Clusters Huawei's Major Move Kept Under Wraps for a Year Arrives
As large models advance towards the 10T parameter scale, SuperNodes are an inevitable choice for AI infrastructure construction.
Wang believes that 100,000-card computing clusters will soon become the standard configuration for training SOTA models, but they also face huge challenges: massive cross-node communication overhead makes it difficult to improve MFU (Model Flops Utilization), and the overall effective computing power of the cluster is far lower than the theoretical value of simply adding up the number of cards.
Therefore, optimizing the computing system architecture around reducing the proportion of communication is a reasonable technical choice for building a highly efficient AI infrastructure. The design philosophy of Huawei's SuperNode is to connect all components within the SuperNode equally with a single set of protocols, support unified memory addressing across physical servers, and adopt an interconnection method of "near copper, far optical", thereby enabling nearly linear bandwidth and latency for communication among tens of thousands of chips. It works like a single computer, effectively improving MFU.
To build this complex system, the first requirement is an interconnection protocol capable of equally connecting all components within the cluster. This is the UnifiedBus launched by Huawei at last year's HC conference, and its relevant technical specifications have been fully opened. Today, a brand-new SuperNode "full suite" centered around UnifiedBus makes a spectacular debut.
02.UnifiedBus Unification, 11 Chips Support, Ascend 960 Debuts: Huawei SuperNode Rushes Towards 1 Million Cards
Wang continued, based on UnifiedBus, Huawei has currently developed 11 key chips for SuperNodes and SuperNode clusters.
1. 11 Chips Support the SuperNode, Ascend 960DT is Here
Among all the chips, the Ascend chip is the core component of the entire system.
Wang announced the launch of the Ascend 960DT chip on-site. This chip supports 2 PFLOPS (FP8) and 4 PFLOPS (FP4) computing power, with on-chip HBM supporting up to 288GB and bandwidth up to 9.6TB per second, achieving doubled performance. The R&D of this chip has currently exceeded expectations, three quarters ahead of the original planned target, and is expected to be ready in Q1 2027.
The Ascend 960PR will further support 8 PFLOPS of FP4 computing power, one quarter ahead of the original plan, and is expected to be ready in Q3 2027.
In the future, Huawei's Ascend series chips will continue to adhere to the evolution cadence of one generation per year. In 2028 and 2029, Huawei will successively launch the Ascend 970 and 980 chips, relying on the innovative direction of the Tau Scaling Law to achieve continued doubling of computing specifications. Simulation bandwidth, simulation capacity, and interconnection bandwidth will also be significantly improved.
Wang also showcased the 11 key chips for the SuperNode cluster based on UnifiedBus on-site, which mainly include the following four categories:
(1) Computing chips, including Kunpeng CPU and Ascend NPU. They are key computing chips, undertaking the most important computing functions.
(2) Interconnection chips. It not only includes Scale-out plane high-capacity switching chips, but also requires low-latency switching chips for the Scale-up plane. At the same time, facing future high-speed interconnection across cabinets, it must also have chips capable of achieving high-speed optical interconnection.
(3) Storage chips, in addition to the well-known media controller chips. Huawei has specifically built a Smart Storage Unit for AI's KV Cache to achieve an equal, one-hop direct-access cache system.
(4) There are also various chipsets for management functions to achieve coordination and management of all aspects in complex systems.
Wang stated that it is precisely with these 11 key chips that Huawei can build more competitive SuperNodes and clusters for AI.
2. Launching the Industry's First Mass-Produced NPO, Ascend 960 SuperNode Reaches 4,096 Cards
SuperNodes, and larger-scale SuperNodes, require continuous innovation in optical interconnection technology. To this end, Huawei announced today the launch of the industry's first mass-produced NPO (Near-Package Optics) optical engine—Hi-ONE, building a scalable SuperNode interconnection system with "UnifiedBus + Hi-ONE".
Wang announced that the world's first 7.2Tbps NPO optical engine, Hi-ONE, has achieved mass production. This is the industry's first mass-produced NPO product, the NPO product with the largest transmission capacity in the industry currently, and the only NPO product with a built-in light source. It perfectly integrates high bandwidth, high reliability, low latency, and low power consumption.
Today, Huawei also officially launched the industry's first SuperNode adopting NPO—the Ascend 960 SuperNode. A single Ascend 960 SuperNode can achieve a scale of 4,096 cards, providing up to 8 EFLOPS FP8 computing power, realizing up to 1PB of HBM capacity, doubling the system's mean time between failures, and achieving a system availability of 99.8%. The Ascend 960 air-cooled SuperNode and liquid-cooled SuperNode will be launched in Q2 and Q3 2027, respectively.
3. Kunpeng SuperNode: 4,096 Nodes, 256TB Unified Memory Pool, Agent Startup 30 Times Faster
Massive Agent concurrency and interaction require general computing SuperNodes; to this end, Huawei's Kunpeng SuperNode has been completely upgraded.
As Agent applications deepen, the computing load of CPUs and NPUs in intelligent computing systems has undergone significant changes. On the one hand, multi-Agent interaction requires second-level startup and hundreds of thousands of sandboxes, necessitating millisecond-level sandbox pull-up speeds. At the same time, the massive information retrieval by multiple Agents also requires matching vector retrieval performance, with a single server requiring hundreds of thousands of TPS (Transactions Per Second).
To this end, Huawei has introduced the SuperNode architecture into the general computing system. Through the unified addressing of SuperNode memory, the memory of multiple general computing servers is formed into a unified shared memory pool, which can significantly improve the startup speed of Agent sandboxes, enhance Agent vector retrieval performance, and improve the resource utilization efficiency of the entire AI inference. Last year, Huawei released the world's first general computing SuperNode—the Taishan 950 SuperPoD, and many customers have deployed it at scale. Today, the Kunpeng SuperNode will be comprehensively upgraded. Based on UnifiedBus, it supports up to 4,096 nodes, forming a unified memory pool of up to 256TB.
The Kunpeng SuperNode can significantly accelerate Agent performance. The startup of 100,000-level sandboxes is 30 times faster compared to traditional server solutions, and the sandbox density is further increased by 25%. In complex vector retrieval scenarios, the retrieval efficiency is doubled compared to traditional server performance, reaching 300,000 TPS.
4. Launching AI Memory Storage Solution, KV Cache Latency Reduced by Over 50%
Agents and long sequences require a multi-level KV Cache architecture. Huawei has built the AI memory storage OceanStor M900 based on UnifiedBus.
Huawei's OceanStor M900 is equipped with the UnifiedBus bus, supporting direct connection to Huawei's L3.5 PB-level KV Cache solution. The industry typically adopts an L3.5 storage system interconnected by PCIe plus RoCE, where GPU or NPU access to the SSD pool requires 5 data transfers. Huawei's OceanStor M900 requires only 1 data transfer, reducing I/O latency and first-token latency by over 50%.
At the same time, Huawei's OceanStor M900 can increase the read/write lifespan of SSDs by 16 times, guaranteeing a high hit rate and room-temperature reliability for KV Cache from the underlying level.
5. UnifiedBus + 2-Layer CLOS Networking, Huawei Builds 1-Million-Card Ultra-Large-Scale Cluster
Based on the flexible networking of UnifiedBus + 2-layer CLOS, Huawei has built a 1-million-card ultra-large-scale cluster. A SuperNode cluster is a large cluster formed by interconnecting multiple SuperNodes through Scale-out network switches. The maximum specification of the Ascend 960 SuperNode is 4,096 NPU cards. Multiple NPU SuperNodes are connected together through Huawei's UnifiedBus network to build a larger-scale SuperNode cluster. The SuperNode cluster supports multiple networking modes: first, adopting a 2-layer CLOS single-plane networking, Huawei can achieve efficient interconnection of 32,000 Ascend 960 cards. Second, adopting a 2-layer CLOS multi-plane networking, Huawei's maximum cluster scale can reach 512,000 Ascend 960 cards.
Huawei aims to leverage its comprehensive advantages in "computing + communication" to build SuperNode clusters in the Agentic era, accelerating the training and inference of 10-trillion-parameter large models.
Such SuperNode clusters possess three major characteristics: (1) Unified interconnection via UnifiedBus, unifying multiple interconnection protocols into the UnifiedBus protocol to significantly reduce protocol conversion overhead. (2) It interconnects subsystems such as Ascend SuperNodes, Kunpeng SuperNodes, and KV Cache as peers. (3) Huawei can provide multi-level, high-bandwidth, large-capacity storage systems, and all types of KV Cache can be accessed in a single hop to meet the hot, warm, and cold KV Cache needs in the system, maximizing the overall resource utilization efficiency.
03.Four Major Measures to Promote AI into Terminals, Vehicles, and Homes Building a New Generation of Communication Networks
Lightweight terminals will also become intelligent entry points. In the past, traditional terminals provided users with standalone functions. In the future, AI will reconstruct the relationship between terminals and human-machine interaction, building an integrated intelligent space centered around humans. Humans and intelligent agents need to co-create and co-evolve, and AI entering terminals is comprehensively accelerating. Taking mobile phones as an example, the parameters of on-device models will achieve a 10-fold leap from 2026 to 2030, quickly increasing from 17B-30B to 70-100B. On-device computing power is also expected to increase from 20-80 TOPS to 180-200 TOPS.
Huawei will build capabilities from four aspects:
1. Achieving self-sufficiency in on-device computing power through the "Kirin + Ascend" combination.
2. Achieving on-device intelligence that is useful and easy to use through the "Pangu + third-party large models" approach. Recently, on the newly released Huawei Mate X2 Ultimate Design, Huawei first commercially deployed the native Pangu MoE fully domestic large model, with 30B parameters and 2B activated parameters, focusing on intelligence, security, and speed.
3. HarmonyOS is not only an operating system for the Internet of Everything but will also be an Agent OS with ubiquitous intelligence. Huawei will comprehensively reconstruct the underlying system architecture, operating mechanisms, and interaction logic to support the symbiosis and shared use of Agents and humans.
4. Huawei's rich AI ecosystem is the foundation for the flourishing of the Xiaoyi system's intelligent agents. At the Huawei Developer Conference in June this year, Huawei released the HarmonyOS Agent Framework, which will fully open up the capabilities for northbound application and southbound device AI access.
Huawei will focus on smartphones, AI PCs, vehicles, and home scenarios to build four major on-device computing platforms. Through the synergy of device-to-device computing power and device-to-cloud computing power, it will build distributed swarm intelligence, providing integrated and continuous intelligent services around personal mobile space, office space, vehicle space, and home space.
Regarding the IoT ecosystem, the scale of the open-source HarmonyOS ecosystem has exceeded 1.3 billion. Wang stated that AI Coding has significantly improved the development efficiency of Huawei's terminals, and the era of lightweight intelligent terminals has arrived.
Focusing on "computing usage", Wang finally proposed building a new generation of communication networks to connect computing power together, ultimately forming a silicon-based black soil.
He mentioned that the communication network in the AI era will revolve around computing, as well as emerging communication technologies such as 5G or 6G and 10G optical networks. From the center to the edge, and then to the terminal, it will provide network connection capabilities on demand for different users, ensuring that intelligence reaches every individual, family, and enterprise.
04.Huawei Kunpeng Developers Exceed 4.16 Million Ascend Crosses Ecosystem Inflection Point Balancing Private and Public Clouds
Building an open-source and open computing ecosystem is Huawei's core strategy. Wang announced that, as of now, the global developer base of Huawei's Kunpeng ecosystem has exceeded 4.16 million, with over 7,200 global ecosystem partners, supporting more than 560 global open-source projects. The installation volume of openEuler has also surpassed 20 million sets, ranking first in the market share of server operating systems in China. Meanwhile, Huawei Ascend has crossed the ecosystem inflection point. CANN has become China's most active AI open-source community, with monthly active developers exceeding 5,000. External developers have surpassed internal developers for the first time, accounting for 61% of the total developer base; the number of native training models based on Ascend has exceeded 40.
Its ecosystem has covered over 90 mainstream third-party communities, such as PyTorch, Triton, vLLM, and VRR. With the strong support of the Linux Foundation, Ascend has become a computing platform that can be directly installed on the PyTorch official website, following NVIDIA, AMD, and Intel. Currently, AI computing power is still in a state of supply falling short of demand. How to improve computing power utilization is a common challenge we face in the development of the AI industry. Different application scenarios, different innovation models, and different stages of intelligence put forward diverse demands for computing power supply models. Wang stated that Huawei is building a silicon-based black soil for thousands of industries, using two methods—computing infrastructure and cloud services, which are two sets of core capabilities—to meet customers' intelligent needs with flexible computing power supply.
05.Conclusion: UnifiedBus Breaks the Wall Huawei to Rebuild the Computing Foundation in the Agent Era
From UnifiedBus breaking the communication wall, to 11 chips supporting the SuperNode, to the million-card cluster and the open-source ecosystem crossing the inflection point, Huawei is building the computing foundation for the 10-trillion-parameter large model era with its vertical integration capabilities in "computing + communication". As computing power supply moves from "selling cards" to "selling clusters", and from "hardware monetization" to "ecosystem win-win", the competition in AI infrastructure has upgraded to a contest of systematic capabilities. Huawei's silicon-based black soil is attempting to provide another solid choice for the intelligent world.