At the HUAWEI CONNECT event in Shanghai, critical announcements regarding artificial intelligence infrastructures took center stage. Huawei officially introduced UnifiedBus, a new generation computing architecture aiming to eliminate information transfer ends between processors, memory, and storage units. Focusing on the processing power burden brought about by increasing AI spies and massive models, the new architecture enables infrastructures from data centers to SME level to scale much more flexibly and with higher efficiency.

The New Bottleneck of AI: Inter-Component Communication

The geometric growth of AI models and the proliferation of autonomous spies (Agentic AI) with independent decision-making have changed the stability in the hardware world. Today, the biggest challenge is not just having powerful processors; It’s managing the massive data traffic between processors, memory, and storage units without delay.

Huawei directly addresses this bottleneck with its UnifiedBus architecture. Connecting CPUs, NPUs, system memory, and storage resources through a common and direct boundary, this structure significantly increases data transfer speeds while drastically reducing latency. This allows hardware resources to be dynamically allocated to different AI workloads, eliminating wasted capacity across the system.

Flexible Scaling Reaching Millions of NPU Capacity

The most striking aspect of the UnifiedBus architecture is its modular structure, which significantly increases scalability. The system can expand from a single in-cabinet connection to inter-cabinet connections and massive clusters reaching millions of NPUs in enterprise data centers.

Agentic AI SuperCluster and Enterprise Solutions

Huawei also announced its new Agentic AI SuperCluster solution, built on UnifiedBus, as part of the event. The core components of this system are:

• Computation and Acceleration: TaiShan 950 servers and Atlas 960 SuperPoD infrastructure.
• Data and Network Management: OceanStor M900 storage solutions and Xinghe UBG network architecture.

The technology is not only for massive data centers; it also offers accessible solutions for more modest budgets and businesses. For example, two Atlas 650E servers can be directly integrated with each other via UnifiedBus to work together as a single powerful system. Thanks to this, small and medium-sized enterprises (SMEs) can implement large-scale licensing models within their own organizations with cost advantages.

Open Source and Ecosystem Support Expanding

Huawei, reinforcing hardware innovations with software and community support, continues to strengthen its Ascend platform with open-source projects. The company is developing add-ons aimed at accelerating extraction processes, managing AI spyware, and ensuring information security with projects such as DeepSeek Harness, OpenCode, and openJiuwen.

To date, Huawei has prepared operator libraries specific to 26 fields in collaboration with more than 50 business partners and universities, and continues its active cooperation with more than 90 open-source communities. It is emphasized that developers can integrate industry-specific AI spyware into existing corporate IT systems much more easily thanks to this open ecosystem.

Speaking at the event, Huawei Executive Director and CEO of ICT Business Group, Yang Chaobin, emphasized system-level innovation and open architecture, stating that they will continue to offer a strong alternative to the global information processing market.