Author | Yang Jingli Editor | Li Shuiqing
Zhidx reported on September 22 that recently, Huawei unveiled the world's first 3D data center, reshaping the AIDC infrastructure architecture through engineering innovation. By replacing horizontal layout with vertical stacking and decoupling the cooling, power supply, backup power, and IT equipment into separate layers, it drives the transformation of AIDC construction from customized engineering to productized delivery. This provides a new pathway to overcome the three major bottlenecks of space, power supply, and heat dissipation in the AI computing era, leading a new paradigm for next-generation AIDC construction.
Drawing on the concept of strict separation between the power layer and the IT layer in chip fabs, the 3D data center vertically stacks various functionally zoned spaces that were traditionally deployed horizontally through layered decoupling and vertical stacking. This forms a vertical architecture of cooling on the first floor, IT equipment on the second floor, power supply on the third floor, and backup power on the roof, achieving vertical power supply, vertical cooling, shortest link, and minimum loss. It supports high-density deployment of AI computing clusters while reserving space for computing power evolution. In terms of the construction model, Huawei has turned core electromechanical systems such as power supply and cooling into standardized PODs (the smallest independently operable data center units), completing prefabrication, assembly, and testing in the factory, with on-site work mainly focusing on assembly and connection. Its prefabrication rate for large and small electromechanical systems exceeds 90%, and the electromechanical delivery time is shortened from about 6 months to 3 months.
Beyond project implementation, Huawei also opens up its design and construction experience, publishing the monograph 3D Data Center and releasing the 3D Data Center Conceptual Design Library. This guides the industry to transform AIDC construction into a standard product that can be mass-replicated, enabling the 3D data center to evolve from a single project into a new industry-referenced paradigm.
01.Single Rack Power Consumption Reaches the Hundred-Kilowatt LevelAI Data Centers Evolve from 2D to 3D
Traditional data centers primarily serve general-purpose computing, where functional areas such as IT equipment, power supply, and cooling are usually laid out horizontally on the same plane or in adjacent areas. A 2D layout can basically meet the equipment deployment needs. As we enter the AI era, the situation is changing. To improve computing efficiency, more chips and computing nodes are concentrated into single racks and clusters. Rack power has increased from a few kilowatts and tens of kilowatts to 100 kW and 200 kW, and may continue to rise. As the power per rack keeps increasing, data centers need to provide higher density and more stable power supply, while promptly dissipating the large amount of heat generated by the equipment, which also poses higher requirements for heat dissipation capabilities. Meanwhile, the scale of AI clusters is expanding from 10,000 cards to 100,000 cards.
If servers, switching networks, power supply, and cooling equipment continue to be laid out on a single plane, it will not only occupy more space, but the power lines and liquid cooling pipelines may also be extended accordingly, increasing the difficulty of engineering deployment and transmission losses. In addition, the upgrade cycle for intelligent computing equipment is continuously shortening. Therefore, 3D data centers need to have elastic scalability, requiring more forward-looking planning and construction models. Facing the rapidly emerging computing power demands, infrastructure must be ready in advance, and on-site construction time needs to be compressed through modularization and prefabrication. As the cluster scale expands, the reliability requirements for data centers also increase accordingly.
A 100,000-card cluster carries a massive number of devices and computing tasks. A single power supply or cooling failure could affect large-scale training and inference services. Therefore, it is also necessary to narrow the scope of failure impact and improve abnormal identification and proactive O&M capabilities. Facing these changes, AI data centers must start from spatial layout and system architecture, decoupling the computing power layer from the power layer, and retaining elasticity for future equipment upgrades.
02.Layered Cooling, IT, Power Supply, and Backup Power3D Architecture Supports 100,000-Card Clusters
To meet the needs for high-density deployment, equipment iteration, and rapid delivery, Huawei has launched the world's first 3D data center. Drawing on the concept of separating the power layer and the production layer in chip fabs, the 3D data center vertically stacks the various functionally zoned spaces that were traditionally deployed horizontally, setting up the cooling layer, IT equipment layer, power supply layer, and backup power layer from bottom to top.
The first floor is the cooling layer. The Thermal Management Unit (TMU), serving as the central brain of the liquid cooling system, relies on AI technology to achieve leak detection and a 7-day advance warning for working fluid health, shifting from passive to predictive maintenance, making data center heat dissipation more efficient, reliable, and intelligent.
The second floor is the IT equipment layer, where liquid cooling racks and high-speed switching networks are centrally deployed. The high-speed switching network is positioned as close to the geometric center of the cluster as possible to shorten the connection distance between computing nodes. Power lines and liquid cooling pipelines are vertically accessed from the upper and lower floors respectively, also leaving space for subsequent replacement of computing equipment.
Further up is the power supply layer. The power modules deployed here adopt an integrated hyper-converged design, which is high-density and space-saving, and supports rapid deployment and elastic expansion. Vertical power supply shortens the distance from the power module to the rack from 17 meters to 5 meters; under the same volume, the IT power supply capacity increases from 76MW to 168MW. On the roof's backup power layer, the SmartLi intelligent lithium battery system is deployed in independent compartments.
The compartments are equipped with temperature, smoke, and combustible gas detection devices, and configured with two-level fire extinguishing at the pack level and room level, as well as exhaust, drainage, and explosion relief measures, effectively ensuring the safe operation of the data center. SmartLi also supports intelligent voltage boosting and remote deployment, eliminating the need to thicken cables to meet long-distance power supply requirements, thereby reducing deployment costs. Through layered decoupling, the power supply, cooling, IT, and backup power systems achieve vertical stacking, jointly supporting the high-density deployment of 100,000-card clusters while reserving elasticity for subsequent computing equipment iteration.
03.Electromechanical Prefabrication Rate Exceeds 90%On-Site Construction Shifts to POD Productized Delivery
The four-layer architecture solves the problem of how to deploy and connect the cooling, IT, power supply, and backup power systems, while POD productization further addresses how to achieve rapid delivery and elastic expansion of these systems. The 3D data center uses the POD as the smallest independently operable unit, turning core electromechanical systems such as power supply and cooling into standardized modules, completing production, assembly, and testing in the factory, with on-site work mainly focusing on assembly and connection. The steel structure main building of Huawei's 3D data center is completed in about 6 months, and the electromechanical PODs inside the building are deployed according to actual needs. The prefabrication rate for large and small electromechanical systems exceeds 90%, and the electromechanical delivery time is shortened from about 6 months in traditional solutions to 3 months.
Besides delivery speed, the layered architecture and POD design can better control the scope of failure impact. Power supply, cooling, backup power, and IT equipment are isolated by layers, making different functional units relatively independent. When a single POD fails, it will not spread to other units, achieving minimization of the failure domain. In terms of O&M, after layered decoupling, equipment can be handed over to corresponding teams for management separately, which can reduce mutual interference between different teams and lower the risk of misoperations.
Huawei also applies AI to equipment status and O&M behavior identification. For example, AI can be used to identify risks such as dirt blockage on liquid cooling cold plates and working fluid degradation in advance, and monitor abnormal behaviors during on-site operations, enabling O&M to gradually shift from passive response to advance identification and proactive prevention. Thus, the 3D data center forms two synergistic engineering paths: the four-layer architecture supports high-density deployment and elastic evolution of computing power, while POD productization achieves high-quality and rapid delivery.
04.Conclusion: Expansion of AI Cluster ScaleA New Paradigm for Next-Generation Data Center Construction
Through layered decoupling, vertical stacking, and POD productized delivery, Huawei's 3D data center breaks through the bottlenecks of space, power supply, and heat dissipation, providing a brand-new paradigm for the construction, delivery, and O&M of high-density computing infrastructure. On this basis, Huawei has also published the monograph 3D Data Center, released the 3D Data Center Conceptual Design Library, and participated in the work of AIDC evaluation indicators, reference architecture, and benchmark verification. As these design and engineering experiences are opened to the industry, the construction practice of 3D data centers will provide the industry with new, referable, and learnable ideas, driving the industry to achieve a generational leap.
As the scale of AI clusters expands and the iteration of computing equipment accelerates, data center construction also needs to shift from customized engineering to standard products that can be mass-replicated. In the future, AIDC infrastructure must also reserve elasticity for equipment upgrades, ensuring that the construction pace keeps up with the evolution of computing power.