Huawei Unveils 168MW 3D AI Data Center

Huawei says more than 90% of the facility’s mechanical and electrical systems can be prefabricated, potentially reducing electromechanical delivery time from about six months to three months. The Wuhu project reportedly took roughly nine months to build, compared with 18 to 24 months for conventional facilities.
Huawei described the data center as a standardized product rather than a bespoke engineering project. Senior Vice President Ren Shulu said, “We treat the data center as a standard product that can be replicated at scale,” with the design intended to allow computing hardware to be replaced without reconstructing the facility’s underlying structure.
Huawei also announced a 3D Data Center Overview Design Library to provide reference designs for the development of future AI infrastructure.
The design is intended to address compatibility challenges that can arise when data centers use different types of AI chips, in addition to handling the power, cooling and space demands of large Ascend-based clusters.
Huawei says its layered architecture supports intelligent operations and zone-based isolation, while Huawei Cloud is building additional 3D data centers in Guizhou, Inner Mongolia and Anhui.
Huawei unveiled a four-layer '3D data center' in Wuhu, Anhui that stacks cooling, servers, power and backup systems vertically to support AI clusters with over 100,000 Ascend accelerator cards. Huawei Connect 2026 reports the 168-megawatt facility shortens power and cooling routes compared to traditional horizontal layouts, a critical advantage as AI workloads demand server-rack power levels of 100 to 200 kilowatts.
The company says more than 90% of mechanical and electrical systems can be prefabricated, cutting delivery time from six months to three months. HackerNoon noted the Wuhu project took nine months to build versus 18 to 24 months for conventional data centers, positioning the design as a scalable, standardized product rather than a custom engineering project.
Traditional data centers arrange IT equipment, cooling systems and power supplies horizontally across floors. Huawei's 3D model stacks all four layers vertically, shortening the distance electricity and coolant must travel. Media OutReach explained the architecture also isolates failures within zones, preventing one problem from cascading across the entire facility while supporting denser hardware arrangements.
This design matters because AI accelerator cards consume far more power than older server equipment. As power demands rise toward 100 to 200 kilowatts per rack, traditional layouts waste energy moving power and coolant over longer distances. The vertical approach reduces waste and keeps the facility cooler with less energy consumption.
Huawei treats the 3D data center as a standardized product, not a custom build. Senior Vice President Ren Shulu said, 'We treat the data center as a standard product that can be replicated at scale.' Over 90% of mechanical and electrical systems can be built off-site and shipped assembled, cutting assembly time roughly in half.
The Wuhu facility took nine months total to become operational. Conventional data centers typically require 18 to 24 months from planning to full power-on. Faster deployment matters because demand for AI computing capacity is growing rapidly, and delays cost cloud companies revenue and customers.
The 3D data center design supports Huawei's larger strategy to reduce dependence on U.S. AI chips and build a complete, homegrown AI ecosystem. Wall Street Journal reported Huawei is developing alternatives to Nvidia's AI systems. The Wuhu facility showcases the company's Ascend accelerators at scale, demonstrating they can power massive AI clusters.
Huawei Cloud is building similar 3D data centers in Guizhou, Inner Mongolia and Anhui. The company also released a 3D Data Center Overview Design Library for reference designs of future AI infrastructure. These moves position Huawei as a competitor to cloud providers like Amazon and Microsoft while reducing dependency on foreign technology.
Publishers
22
Articles
9
Reach
31