AI data centers are getting so power-hungry that Huawei wants to start building them upward.
The Chinese technology giant unveiled a four-layer “3D data center” in Wuhu, Anhui province, on Tuesday, with cooling, computing equipment, power systems and backup power stacked vertically. Huawei says the architecture is designed for increasingly dense AI infrastructure, including supernode clusters containing more than 100,000 accelerator cards.
The approach comes as AI hardware pushes conventional data centers toward much higher power densities. Huawei Senior Vice President Ren Shulu said server rack power has risen from several kilowatts to 100 kW and 200 kW, and is expected to continue increasing.
Huawei Deputy Chairman and Rotating Chairman Wang Tao said, “The massive model with trillions of parameters and the rapidly growing demand for tokens place higher demands on AI infrastructure. Ascend supernode clusters with over 100,000 cards will become the basic configuration,” according to Huawei Central.
Huawei wants to build data centers more like products
The physical layout is only part of Huawei’s pitch. The company is also treating the data center more like a manufactured product than a traditional construction project.
Huawei says more than 90% of the mechanical and electrical systems can be prefabricated, reducing electromechanical delivery from about six months to three months. The Wuhu facility took about nine months through its standardized construction process, compared with the 18- to 24-month timelines cited for traditional builds, according to Pandaily.
The design also separates different infrastructure domains, which Huawei says can help isolate failures and allow computing hardware to be changed without rebuilding the facility’s underlying structure.
“We treat the data center as a standard product that can be replicated at scale, rather than a bespoke engineering project,” said Shulu.
More must-read AI coverage
The bigger bet is standardization
Huawei is not presenting the Wuhu facility simply as a showcase. It released a conceptual design library covering architecture, structures, HVAC, water systems, and electrical infrastructure, along with a book titled 3D Data Center.
The company says its cloud business is also deploying the architecture at its core hubs in Gui’an, Inner Mongolia and Wuhu.
The more interesting part of Huawei’s strategy is that it is trying to make data center construction more like a repeatable hardware platform. If facilities can be prefabricated and computing equipment can be swapped as new chip generations arrive, operators could spend less time rebuilding infrastructure for each new AI generation.
That promise still has to survive real-world deployments. Operators will need to validate cooling reliability, power efficiency, maintenance procedures and compatibility with hardware from different vendors before treating Huawei’s architecture as a broadly applicable industry model.
What infrastructure leaders should watch
For data center operators and enterprise infrastructure teams, Huawei’s announcement signals a shift in how AI capacity may need to be planned. The useful question is no longer simply how many accelerators a facility can accommodate, but how easily the surrounding infrastructure can adapt when computing requirements change.
That puts greater weight on decisions that can outlive a particular generation of AI hardware: whether power and cooling capacity can expand independently, how much of a facility can be upgraded without major reconstruction, and whether operators can replace computing equipment without becoming locked into one vendor’s architecture.
Huawei’s Wuhu project will provide an early test of that approach. The numbers to watch now are not just accelerator counts, but operating costs, energy use, deployment time, maintenance requirements and how frequently major infrastructure changes are needed as the hardware evolves.
If those measurements hold up across larger deployments, the lasting significance of the 3D data center may have less to do with stacking equipment vertically and more to do with making the building itself less of a constraint on the next generation of AI.
Related reading: For more on the infrastructure challenges behind increasingly powerful AI clusters, see how power and cooling are being redesigned for the AI era.
