During HUAWEI CONNECT 2026, Huawei presented its vision for developing infrastructure capable of keeping pace with the rapid expansion of AI models and inference and training workloads. In a keynote delivered by David Wang, Huawei’s Deputy Chairman and Rotating Chairman, the company focused on strengthening computing and connectivity capabilities as the foundation for expanding AI applications.
The company’s vision is based on SuperPoD and SuperCluster technologies, which aim to connect large numbers of computing nodes within a unified, high-speed environment. Huawei says this approach enables large-scale workloads to be handled more efficiently, while providing computing options on customer premises and through the cloud.
Why Is the Need for Specialized Infrastructure Growing?
Huawei links this trend to the growing sizes of models and the capabilities of AI agents. According to figures presented by the company, the parameter counts of large models are approaching 10 trillion, with expectations that they will exceed 100 trillion by 2030. The size of smartphone models has also grown from 3 billion parameters in 2024 to 30 billion currently, with models containing 100 billion parameters drawing closer.
At the inference level, daily volume in China reached approximately 500 trillion tokens, and Huawei expects it to reach the quadrillion level by 2030. This increase is placing greater pressure on system performance and reliability, particularly in environments that combine tens of thousands of neural processing units.
Simulation Results and Their Practical Meaning
Huawei explained that communication operations between components in traditional infrastructure can consume more than 40% of total model-training time, limiting the efficiency of computing-resource utilization. In a simulation conducted by the company’s Markov Lab, a group comprising 100,000 NPUs, built using SuperPoDs each containing 4,000 NPUs, achieved up to 2.75 times higher computing-resource utilization efficiency than a comparable group based on servers containing 8 NPUs each.
This result illustrates the importance of reducing communication time between units when training large-scale models, but it remains a simulation result rather than an independently published measurement of an actual operating environment. Therefore, the material alone does not explain the details of the workloads, software, or testing conditions that produced the reported difference.
Atlas Deployment and Ecosystem Support
Huawei said it has deployed more than 1,000 Atlas 900 A3 SuperPoD systems to date, while the Atlas 950 SuperPoD is seeing broad commercial deployment. It also noted that it supports the training of foundation models on its systems, in addition to running a wide range of models and applications within open computing ecosystems.
The company’s strategy includes limited, medium, and massive computing capabilities, while expanding the use of AI on devices and in vehicles, in parallel with the development of next-generation communications networks. This indicates that Huawei is not presenting SuperPoD merely as a standalone product, but as part of a broader vision that connects processing, networks, the cloud, and devices.