Huawei is scaling the AI stack. On September 17, 2026, Huawei Technologies announced an upgraded Stellar AI Network and new chip technologies at HUAWEI CONNECT 2026 in Shanghai, combining enterprise networking, computing infrastructure, and security systems in one broad push to reduce reliance on foreign technology.
The Stellar AI Network now covers three domains: AI Fabric, AI WAN, and AI Campus. Huawei designed the architecture for enterprise demand after daily AI token consumption increased 260-fold over the past year—a useful reminder that the industry’s appetite for infrastructure is no longer pretending to be modest.
Networking Built for Larger AI Clusters
Within AI Fabric 2.0, Huawei introduced the SF9300 series unified bus gateway switch. Its two-layer multi-plane design removes the core network layer, cutting infrastructure deployment costs by 30% and reducing end-to-end latency from 20 microseconds to 11 microseconds.
The SF9300 also uses link-layer retransmission technology to prevent data loss during link flaps. Huawei estimates the design can cut annual cluster service downtime by 100 hours, turning a familiar source of operational misery into a measurable infrastructure target.
Huawei also unveiled the CloudEngine XH9300-EN series of near-packaged optics switches. A centralised light source lowers interconnect power usage from 1,000 watts to 600 watts, while the near-packaged layout cuts optical-electrical conversion latency by 180 nanoseconds and improves single-node latency by 26%.
The switches use a modular snap-fit design for on-site replacement, reducing fault recovery times from days to hours. That detail matters because AI clusters do not become easier to repair merely because their processors become more expensive.
Stellar AI WAN extends computing delivery across spans exceeding 1,000km. Huawei’s proprietary Starnet lossless algorithm keeps remote processing efficiency above 95%, while the architecture protects raw data locally, cuts remote branch processor expenses fourfold, and reduces required network bandwidth fivefold.
Security, Campuses, and the Next SuperPod
The campus networking platform samples telemetry across more than 25 data categories every 100 milliseconds and lowers mean time to repair network faults to under five minutes. Its three-layer security guardrail monitors cryptojacking, prompt injection, and host vulnerabilities.
During an initial week-long customer trial, the system detected 415 high-risk flaws and 322 threat events. Campus switches also embed AI micro-models that reduce endpoint threat detection from five minutes to under ten seconds, which is closer to the response time security teams actually want.
The WAN architecture integrates quantum key distribution over distances of up to 80km. Huawei has placed that feature alongside its network controls rather than presenting security as a separate accessory—an increasingly outdated way to build systems carrying sensitive AI workloads.
Huawei also announced the Atlas 960 SuperPoD, powered by Ascend 960 chips, together with an upgraded version of its UnifiedBus interconnect technology. The Atlas 960 is a faster successor to the Atlas 950 SuperPoD launched late last year, and Huawei says it delivers a 2.3-fold improvement in training and a 2.5-fold improvement in inference for 10-trillion-parameter models.
The Atlas 960 SuperPoD and upgraded UnifiedBus technology can connect up to 4,000 processors, pushing high-speed optical connections deeper into the computing system. Huawei introduced near-packaged optics to its SuperPoD architecture for the first time, extending the same effort to reduce conversion delays and power use into the cluster itself.
Huawei plans to launch the Ascend 970 chip series in 2028 and the Ascend 980 series in 2029. The company presented these technologies as part of its strategy to build advanced AI systems despite US semiconductor-related restrictions, making the announcements as much about technological independence as raw performance.
The timing adds political weight. Huawei’s announcement came less than two weeks before President Xi Jinping’s visit to Washington, and George Chen, partner and chair of digital practice at The Asia Group, said, “The timing — less than two weeks before President Xi Jinping’s visit to Washington — underscores Beijing’s confidence and ambition in technology and innovation.”
Chen also offered a less comfortable forecast for every company treating a current lead as permanent: “AI developments move so quickly that no one can be certain of holding the lead forever.” Donald Trump put the contest more bluntly: “whoever wins in AI, wins.” Huawei’s Stellar network, Atlas systems, and future chip roadmap show how seriously that contest is being built.
Based on
- Huawei debuts upgraded Stellar AI Network for enterprise computing — techmonitor.ai
- Huawei unveils latest tech to boost AI power in push to break China’s Nvidia reliance | South China Morning Post — scmp.com
- Huawei unveils new chip technologies as Chinese firm steps up the AI race with Nvidia — thehill.com
