Huawei presented a four-layer 3D data-center architecture on September 15, 2026, at an AI data-center conference in Wuhu, Anhui, China. The reference design stacks cooling, IT equipment, power supply and backup power instead of arranging those infrastructure zones primarily side by side.
The proposal arrives as large AI clusters push data centers toward much higher cabinet power and cooling requirements. Huawei says its architecture is intended to support clusters with more than 100,000 Ascend cards and that one building can support up to 168 MW.
Huawei’s four-layer architecture
The design divides the building into four dedicated vertical zones:
- Cooling: the lowest layer, dedicated to cooling infrastructure.
- IT equipment: the second layer, housing the high-density computing systems.
- Power supply: the third layer, dedicated to power-delivery equipment.
- Backup power: the upper layer, including backup-power systems and batteries on the roof.
Huawei says this arrangement creates dedicated vertical routes for cooling and power, separates infrastructure zones and supports modular construction. The company presents those choices as a way to shorten infrastructure paths and make the design easier to standardize and replicate.
Why AI infrastructure is moving upward
AI systems are concentrating more computing power in each cabinet. Huawei executives have described high-density cabinet power in the range of approximately 100 kW to 200 kW, with demand expected to rise as large models and token workloads expand.
That changes the building problem. Cooling systems, power delivery and backup infrastructure have to be planned as part of the compute architecture rather than added around a conventional server hall. Huawei’s vertical approach places those functions in dedicated layers, with the IT equipment between the cooling and power systems.
The company also describes the architecture as modular and standardized rather than a one-off engineering project. In practical terms, the proposal is aimed at repeating a common infrastructure pattern across large AI deployments.
More than 100,000 Ascend cards and up to 168 MW
Huawei says the reference architecture is designed for Ascend supernode clusters exceeding 100,000 cards. That figure describes the intended scale of the cluster configuration presented with the architecture.
Huawei also says a single building can support up to 168 MW. Here, MW refers to electrical power capacity, not a price or construction budget. The announcement presents the figure as a building-level capability associated with the architecture.
The Wuhu presentation therefore combines two parts of Huawei’s AI infrastructure strategy: a physical design for concentrating cooling, computing and power systems, and a target scale built around very large Ascend clusters.
Huawei’s design library
Alongside the architecture, Huawei announced the 3D Data Center Overview Design Library. It covers architectural and structural planning as well as electrical, mechanical, water-supply, HVAC, and high- and low-voltage infrastructure.
That library gives the concept a broader planning framework. Rather than describing only a server-building layout, Huawei is packaging the design around the systems required to plan and replicate an AI data center from the building structure through its cooling, water and electrical infrastructure.
