Huawei has upgraded its Stellar AI Fabric networking platform to support high-density AI clusters, targeting the throughput, reliability and utilization requirements of production-scale agentic AI workloads. The company unveiled the updated architecture at HUAWEI CONNECT 2026 alongside new data-center switches designed for high-performance computing environments.
AI infrastructure is increasingly becoming a networking problem as much as a compute problem. As AI clusters grow larger and workloads shift from model training toward continuous inference and agentic applications, the network connecting accelerators can become a significant constraint on overall system performance.
Huawei is responding with an upgraded Stellar AI Fabric Solution, unveiled at its AI DC Innovation Summit during HUAWEI CONNECT 2026. The company says the architecture is designed to maintain high availability and improve compute utilization in production AI clusters.
The announcement comes as data-center operators expand infrastructure for large language models (LLMs), generative AI and emerging AI agents. Huawei argues that network performance directly affects token throughput and the amount of compute capacity that can be productively used.
The upgraded Stellar AI Fabric combines four components: Rock-Solid Architecture 2.0, StarryWing Digital Map 2.0, Hyper-Converged Fabric (HCF), and Network Packet Load Balancing (NPLB). Huawei is also adding an upgraded NetMaster network AI agent, which it says can automatically identify the root causes of 95% of network faults within minutes.
Those capabilities target a practical challenge in AI infrastructure: keeping large clusters operating consistently while minimizing performance losses caused by network congestion, failures or uneven traffic distribution.
Huawei also introduced two switch families aimed at high-density computing environments.
The CloudEngine SF9300, positioned as a UBG switch, uses a two-layer multi-plane architecture. Huawei says the design can reduce network buildout costs by 30% and lower end-to-end latency by 40% through simplified universal-buffer forwarding. Its Link-Layer Retransmission technology is designed to recover from link instability at microsecond-level intervals and prevent packet loss during link flapping.
The second platform, the CloudEngine XH9300, is an NPO switch built around Huawei’s proprietary 100T and 51.2T configurations. The company says its centralized light source and 3.2T optical engine reduce interconnect power consumption by 40%. Huawei also claims that removing the optical digital signal processor stage reduces forwarding latency by 26%, while pluggable optical engines can accelerate maintenance.
These specifications highlight the growing importance of high-speed networking in AI infrastructure. IDC reported that the global Ethernet switch market reached $18.9 billion in the second quarter of 2026, up 43.4% year over year. Data-center switching grew even faster, rising 64.5% to $12.3 billion as hyperscalers and enterprises expanded infrastructure for AI training and inference.
The competitive landscape includes networking and infrastructure suppliers ranging from Huawei and NVIDIA to established Ethernet and data-center networking vendors. NVIDIA, for example, became the largest vendor by revenue in data-center Ethernet switching in the first quarter of 2026, according to IDC, underscoring how rapidly AI workloads are reshaping the networking market.
Huawei is also extending its networking strategy into regulated and research environments. The Beijing FinTech Industry Alliance, Huawei and 17 partners released a technical specification for high availability in financial data-center networks. The specification addresses redundancy, fault recovery and proactive prevention across intra-data-center and cross-data-center environments.
In scientific computing, Qi Fazhi, computing center director at the Institute of High Energy Physics of the Chinese Academy of Sciences, said the Stellar AI Fabric increased computing efficiency by more than 10% in the institute’s deployment. Huawei says network-wide load balancing helped address training traffic imbalances and improve data throughput for deep-space signal analysis.
The broader direction is clear: AI infrastructure is moving beyond simply adding GPUs. Networking, optical interconnects, storage, software-defined management and automated operations are becoming tightly coupled with the economics of AI clusters.
Gartner estimates worldwide AI spending will reach $2.7 trillion in 2026, with AI infrastructure representing the largest spending category. The research firm specifically includes AI network fabric among the infrastructure required to support anticipated AI workloads.
For enterprises building AI platforms, the question is increasingly not only how much compute they can install, but how consistently that compute can be supplied with data and kept operational. Huawei’s Stellar AI Fabric upgrade is aimed squarely at that layer of the AI infrastructure stack.
Market Landscape
AI infrastructure investment is accelerating across compute and networking. Gartner forecasts AI spending of $2.7 trillion globally in 2026, while AI-optimized IaaS spending is projected to reach $42 billion, driven by LLM training and the operationalization of AI across enterprise workflows.
Networking is becoming particularly important as AI clusters demand higher bandwidth and lower latency. IDC’s second-quarter data shows data-center Ethernet switching growing 64.5% year over year, with 800GbE already accounting for 41.2% of data-center switching revenue.
Top Insights
- Huawei’s upgraded Stellar AI Fabric targets network reliability, traffic balancing and compute utilization in production-scale AI clusters.
- New SF9300 and XH9300 switches are designed for high-density AI environments requiring greater bandwidth, lower latency and improved power efficiency.
- Huawei is applying AI-based network operations to automate fault diagnosis and reduce the operational burden of large AI data centers.
- Rapid growth in data-center Ethernet spending reflects the networking demands created by AI training and inference workloads.
- AI infrastructure competition is expanding beyond accelerators into switching, optical interconnects, network management and cluster-level optimization.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
