The 2026 World Artificial Intelligence Conference and High-Level Meeting on Global AI Governance was held in Shanghai from July 17 to 20 und...
The 2026 World Artificial Intelligence Conference and High-Level Meeting on Global AI Governance was held in Shanghai from July 17 to 20 under the theme “Intelligent Partners, Creating the Future Together.” Representatives from more than 100 countries and international organizations joined participants from government, academia, research institutions, and industry.
During the event, China’s National Development and Reform Commission released the Action Plan for AI Cooperation and Development. The plan sets out eight areas of cooperation, including high-quality data supply, inclusive intelligent-computing services, open-source ecosystem sharing, industry enablement, talent development, standards, safety governance, and responsible AI.
Computing infrastructure remained one of the conference’s central themes. Compared with the previous year, when many product announcements focused on individual accelerator chips, WAIC 2026 put much greater emphasis on supernodes, high-speed interconnects, shared memory, rack-level systems, and large-cluster integration.
Huawei, Alibaba Cloud, Kunlunxin, Biren Technology, MetaX, and Sugon were among the companies presenting systems designed to combine dozens, hundreds, or even thousands of accelerators into larger computing units.
The performance of a single accelerator still matters, but large-model training and high-concurrency inference increasingly depend on the performance of the complete system.
A supernode connects multiple AI accelerators through a high-bandwidth, low-latency scale-up network. The aim is to make a large group of processors behave more like one tightly coordinated computer rather than a collection of separate servers.
This approach is especially relevant to:
The products presented at WAIC show that China’s domestic computing industry is moving from competition at the chip level toward competition in interconnect protocols, memory architecture, server engineering, cooling, networking, system software, and cluster operations.
The following table summarizes the main systems and specifications reported during the conference.
| Vendor | Product or System | Reported Scale | Key Features |
|---|---|---|---|
| Huawei | Atlas 950 SuperPoD | 1,024 accelerators in the displayed configuration | 1 EFLOPS FP8, 2 EFLOPS FP4, 256 TB globally addressed memory, TB-class NPU interconnect bandwidth, 3 μs round-trip latency |
| Alibaba Cloud | Lingjun Zhenwu M890 Supernode Instance | 64 cards | FP8/FP4 support, ICN Switch 1.0, 800 GB/s card-to-card interconnect, support for inference on 10-trillion-parameter MoE models |
| Kunlunxin | 32- and 64-card supernodes | Expandable to 512 cards | High-density integration for domestic AI acceleration; described in media reports as already entering volume delivery |
| Biren Technology | BR2xx and BLink 2.0 supernode matrix | Up to 1,024 GPUs | NPO optical interconnect, distributed disaggregated architecture, shared memory space across the supernode |
| MetaX | Xijing S600 | 64 GPUs per rack | High-density deployment, full interconnection inside the rack, expansion toward clusters with tens of thousands of cards |
| Sugon | Sugon 8000 “Dengfeng” | 100,000-card AI supercluster | Domestic hardware and software stack, connection to China’s National Supercomputing Internet |
The figures above reflect public announcements and media reports available during WAIC 2026. They describe different system configurations and should not be treated as direct, like-for-like benchmark comparisons.
Huawei publicly displayed the physical Atlas 950 SuperPoD at WAIC for the first time.
The configuration highlighted at the conference connects 1,024 Ascend accelerators through Huawei’s UnifiedBus interconnect and supernode architecture. Huawei reported:
The product is intended for large data centers, trillion-parameter model training, and high-concurrency inference.
The significance of the design lies in more than the combined peak compute. Unified memory addressing and lower communication latency can reduce the need to move data through conventional server boundaries, which is important for communication-heavy workloads such as mixture-of-experts routing, distributed key-value cache access, and large-scale parallel training.
Huawei has also described larger Atlas 950 deployments that can scale beyond the 1,024-card configuration displayed at WAIC. The specific figures in this article refer to the conference configuration reported by the original source and Huawei’s WAIC announcement.
Alibaba Cloud officially introduced the Lingjun Zhenwu M890 Supernode Instance, its first supernode-form AI computing service offered through the public cloud.
The service provides a ready-to-use 64-card computing unit with high-speed internal connectivity. It supports FP8 and FP4 low-precision computation and uses Alibaba’s ICN Switch 1.0 interconnect chip.
According to the launch information:
Offering the system as a cloud instance lowers the initial deployment barrier. Customers can evaluate a tightly connected supernode without first purchasing, installing, and operating a complete rack-level system.
The cloud model also places responsibility for hardware maintenance, networking, monitoring, and much of the underlying operations on the service provider.
Kunlunxin also displayed supernode products during the conference.
Media reports described its 32- and 64-card systems as among the earlier domestic supernode products to reach volume delivery. The design uses a high-density integration architecture, supports 32 or 64 accelerator cards in one cabinet, and can expand to a 512-card configuration.
Kunlunxin’s existing product portfolio includes AI accelerators and multi-card accelerator groups designed for large-scale training and inference in cloud data centers.
Because detailed WAIC product specifications were reported mainly through media coverage rather than a dedicated public technical page, deployment status and performance should be confirmed directly with the vendor before making procurement or engineering decisions.
Biren Technology released a next-generation supernode strategy built around:
The system is designed to scale a single supernode to as many as 1,024 GPUs.
Biren states that BLink 2.0 allows up to 1,024 GPUs to share a common memory space. Its broader product matrix includes:
Optical interconnects become more important as the number of accelerators and the physical distance between them increase. Electrical links are effective inside servers and racks, but power consumption, signal integrity, and distance can become more difficult to manage at larger scale.
Biren’s distributed design reflects a wider industry trend: separating compute, memory, switching, and optical components while maintaining a tightly coordinated logical system.
MetaX officially introduced the Xijing S600, a new AI supernode product designed for model training, inference, and intelligent-computing centers.
The system supports:
MetaX positions the S600 as a system-level product rather than a standalone GPU server. Its design emphasizes coordinated optimization across hardware, networking, and software.
The company also maintains the MXMACA heterogeneous computing platform and a developer ecosystem for model adaptation and application deployment.
The Sugon 8000, also called Dengfeng, made its global physical debut at WAIC.
Sugon announced the system earlier in July as China’s first fully domestic 100,000-card AI supercluster. It has been connected to the National Supercomputing Internet, marking a move from clusters with tens of thousands of accelerators toward 100,000-card deployment.
The system combines:
Describe your idea once, and We0 AI can generate a showcase site, pages, and CMS, then help you attract customers and traffic after launch.
One complete project generation for free registration
Best for trying one complete generation flow and seeing a first project draft quickly.
The cluster is designed to support scientific computing, large-model training, AI inference, industrial simulation, biomedicine, materials research, and other large-scale workloads.
A 100,000-card installation is not one conventional supernode. It is a supercluster built from many tightly connected computing and infrastructure components. Its practical value depends on application scaling efficiency, utilization, storage throughput, networking reliability, scheduling, and operating cost.
The announcements at WAIC reflect a clear shift in the competitive landscape.
An individual chip can provide high theoretical compute, but a large model runs across a system. Once hundreds or thousands of accelerators are connected, overall performance depends increasingly on:
Bank of China Securities, cited by the original report, argued that competition in domestic computing is extending toward chip interconnection and system-integration capabilities.
As vendors expand their supernode portfolios, large-scale deployment could also increase demand for supporting technologies such as:
This view is an industry assessment rather than a guarantee of market performance. Actual deployment will depend on product availability, stability, customer demand, pricing, software maturity, and the economics of operating the systems.
WAIC 2026 also featured several AI and robot phones.
The original report suggests that smartphones may evolve from containers for independent apps into platforms that run persistent AI agents. In this model, the operating system and device-level entry point become increasingly important.
Instead of opening a separate app for every task, an agent could coordinate services across communication, search, travel, payment, office work, cameras, sensors, and connected devices.
The appearance of StepFun’s Agent OS at the conference was presented as another sign that agents are moving below the application layer and toward the operating-system layer.
An agent-oriented operating layer could help coordinate:
The concept is still developing, and product capabilities differ widely. A true device-level agent platform must handle permissions, privacy, identity, cross-application execution, reliability, security, and clear user approval for consequential actions.
Embodied intelligence and vertical applications were another major focus of the conference.
Humanoid robots, robot dogs, dexterous robotic hands, and other physical AI systems appeared throughout the exhibition. These products show how the industry is attempting to connect foundation models with sensors, control systems, mechanical hardware, and real-world workflows.
Guotai Haitong Securities, cited in the original article, argued that one of the most important questions at WAIC was whether China’s domestic AI industry can build a complete delivery chain across:
The emphasis is shifting from isolated technical breakthroughs to system-level commercialization and delivery.
For customers, a complete solution must do more than demonstrate a model or robot on an exhibition floor. It must be reliable, maintainable, secure, economically viable, and capable of fitting into an existing business process.
The supernode announcements show that domestic computing vendors are no longer presenting only accelerator chips.
They are building broader platforms that combine:
This does not mean that every announced system has reached the same level of maturity or availability. Some products are in commercial delivery, some are in invitation testing, and others are newly announced architectures or expansion plans.
The common direction is more important than any single specification: AI infrastructure competition is moving from the performance of one chip to the effective performance of the complete system.
An AI supernode is a tightly connected group of accelerators designed to operate as one larger computing unit. It typically uses a high-bandwidth, low-latency scale-up network and may provide shared or globally addressed memory.
A conventional cluster often connects separate servers through a scale-out network. A supernode usually provides tighter internal coupling, lower communication latency, and faster accelerator-to-accelerator data transfer before multiple supernodes are combined into a larger cluster.
The Atlas 950 SuperPoD is Huawei’s Ascend-based AI computing system. The configuration displayed at WAIC 2026 connected 1,024 accelerators and was rated at 1 EFLOPS FP8, 2 EFLOPS FP4, and 256 TB of globally addressed memory.
It is a public-cloud supernode instance built as a ready-to-use 64-card computing unit. Alibaba Cloud says it supports FP8 and FP4, provides 800 GB/s card-to-card connectivity, and can serve very large mixture-of-experts models.
Distributed AI workloads frequently exchange model states, gradients, activations, expert-routing data, and key-value caches. Slow or congested communication can leave expensive accelerators waiting for data and significantly reduce effective system performance.
Sugon 8000 is a fully domestic 100,000-card AI supercluster connected to China’s National Supercomputing Internet. It is designed for AI workloads, scientific computing, industrial simulation, and other large-scale applications.
No. Application performance also depends on communication efficiency, memory bandwidth, software optimization, scheduling, storage, reliability, and utilization. Two systems with the same card count can produce very different results.
The industry is moving from standalone models toward agents that operate across devices and robots. Agent operating systems, intelligent terminals, and embodied systems are attempts to turn model capability into persistent, real-world execution.
WAIC 2026 showed that China’s AI infrastructure industry is moving beyond isolated accelerator products. Huawei, Alibaba Cloud, Kunlunxin, Biren Technology, MetaX, and Sugon presented systems built around supernodes, high-speed interconnects, shared memory, rack-level integration, and very large clusters.
The announcements also showed several deployment models: on-premises supernodes, public-cloud instances, optical-interconnect architectures, rack-scale GPU systems, and a national 100,000-card supercluster.
At the device and application layers, Agent OS, AI phones, robots, and vertical-industry systems point toward a wider shift from model demonstrations to integrated delivery.
The key development at WAIC 2026 was not one faster chip, but the emergence of complete AI computing systems designed to connect, scale, and operate domestic accelerators in production.
Start from one sentence and have a complete website in minutes.
