How Alibaba Cloud’s CUBE 5.0 Builds AI Data Centers in 100 Days

Alibaba Cloud’s CUBE 5.0 is best understood not as a faster construction method, but as a change in what a data center is: from a bespoke project assembled sequentially on site into a standardized product manufactured and tested in parallel.

Alibaba Cloud says the approach can deliver large AI data centers in 100 days, compared with reported traditional timelines of six to 12 months in China and 12 to 18 months in the United States. The company also says the design reduces construction cost per kilowatt by more than 10% and produces a near-100% first-pass product-test rate. Those figures are reported claims, not independent benchmarks. 14

The 100-day schedule replaces sequential construction

Traditional data-center construction often depends on one installation stage finishing before the next can begin. CUBE 5.0 shifts much of that work into simultaneous factory production:

• 30 days: Factory manufacturing and site preparation
• 50 days: On-site lifting, installation, and connection
• 20 days: Integrated commissioning and testing

The modules are assembled and pre-tested before being transported to the site, where they are connected in a process Alibaba compares with assembling Lego blocks. This reduces the amount of weather-sensitive, labor-intensive, and coordination-heavy work performed on site. 49

The key metric is the degree of standardization. CUBE 5.0 raises the reported modularization rate across five core systems—from roughly 30% in earlier designs to 90%: power, cooling, security, intelligent management, and fire protection. 17

That makes parallel production possible. Instead of completing a building and then installing each major system in sequence, suppliers can manufacture, integrate, and test standardized units at the same time. On-site work becomes primarily a matter of placing, connecting, and validating modules.

Why speed can also reduce cost and rework

The economics follow from repeatability. Factory production can use consistent designs, controlled working conditions, standardized procurement, and less on-site labor. Pre-testing also moves fault discovery earlier in the process, before modules reach the data center.

Alibaba says this combination lowers unit construction cost by more than 10% compared with the previous generation and produces a first-pass product-test rate close to 100%. Reporting also cites higher equipment density, fewer transformers, and support for both air and liquid cooling. 446

The important qualification is that these are vendor-reported outcomes. A faster delivery schedule does not automatically prove lower total operating cost, higher availability, or better lifetime economics. Those results depend on transport, site conditions, grid connections, workload utilization, and how well the standardized modules fit a particular deployment.

High-voltage DC and hybrid cooling target AI hardware changes

AI data centers face two related infrastructure problems: accelerator power density is increasing, while chip designs and thermal profiles change quickly.

CUBE 5.0 uses high-voltage direct-current distribution to reduce conversion stages between incoming power and IT equipment. It also combines air and liquid cooling so that a site can accommodate different accelerator types and heat loads. Alibaba says the resulting high-elasticity design can support at least three generations of chips and computing architectures without requiring a complete civil and mechanical-electrical-plumbing redesign for every refresh. 914

That claim matters because the useful life of a data-center shell can exceed the life of the chips installed inside it. A modular power and cooling architecture gives operators more options to replace or mix hardware while preserving the surrounding infrastructure. It does not eliminate the need for future upgrades, but it aims to make those upgrades less disruptive.

The 17-year foundation behind the “Lego” model

CUBE 5.0 was not presented as a design created from scratch. Alibaba Cloud traces the architecture to 17 years of data-center construction and operations, beginning with the Feitian era in 2009. 850

The difficult step was converting accumulated engineering knowledge into something manufacturers could repeatedly produce. Alibaba and industry suppliers reportedly translated complete data-center designs into standardized bills of materials, specifying component placement and, in some cases, fastener details. That level of definition allows modules to be manufactured as products rather than redesigned as one-off engineering projects. 4954

This is also why the model depends on a supply chain, not only on a clever building design. Standardized modules create value when suppliers can manufacture them at scale, test them consistently, and deliver them to sites without extensive modification.

Why Ulanqab is the canonical example

Ulanqab in Inner Mongolia illustrates the second half of the AI-infrastructure equation: choosing a site where the operating environment supports dense computing.

The city is near 42°N and has an average annual temperature of about 4.3°C. Reports say data centers can use natural cooling for roughly 10 months of the year, with comparable-condition energy savings of more than 20% in some accounts. 144

Ulanqab is also positioned close to Beijing-linked demand. Sources place it roughly 240 kilometers from Beijing and report dedicated fiber routes with latency of about 4.2 milliseconds, supporting Beijing-connected training and inference workloads rather than limiting the region to offline batch computing. 3338

Energy is another part of the calculation. Local reporting puts the city’s green-electricity share at about 67%. The combination of renewable power, a cool climate, and proximity to a major technology market helps explain why Ulanqab has attracted a growing concentration of data-center projects. 1117

Claims that local basalt geology makes the area especially resilient should be treated more cautiously. Some reports describe stable geology and low natural-disaster risk as site advantages, but the specific causal benefit is not independently established across the available evidence. 4144

From faster buildings to an AI-capacity strategy

Alibaba Cloud has said it plans to increase global modular-data-center capacity by more than two times, not necessarily triple it. 454 The distinction matters: the strategy is about manufacturing capacity and deployment throughput, rather than simply announcing a larger construction target.

Ulanqab shows the cluster effect. Reporting in 2026 described 89 signed data-center projects in the region, while later reports described more than 100 landed projects and operating compute of about 172,000 P, with intelligent computing accounting for more than 90% or, in later figures, more than 95%. The figures vary by date and by whether they refer to signed, landed, operating, or under-construction capacity. 17192629

Reported participants include Huawei, Alibaba, Kuaishou, ByteDance, Century Internet, and CICC Data. 61219 The emerging model combines energy, cooling, network connectivity, standardized construction, and a concentrated customer base into an AI-production ecosystem.

That is the logic behind Ulanqab’s “Token city” ambition: turn renewable electricity into computing capacity, and computing capacity into model training, inference, and token-generating services. The ambition is larger than filling server racks, but it still faces practical tests. Signed projects are not the same as sustained utilization or profitable demand, and rapid construction does not by itself solve grid integration, chip supply, network capacity, or workload economics.

The broader lesson for AI data centers

CUBE 5.0’s main innovation is productization. By moving 90% of five critical systems into standardized, factory-built modules, Alibaba Cloud says it can run manufacturing and site work in parallel, pre-test more of the system before delivery, and make future hardware changes easier to absorb.

The 100-day claim is therefore less about building faster with the same process than about replacing part of construction with manufacturing. Whether the model becomes a new industry standard will depend on independent verification of its cost and reliability claims—and on whether factories, power infrastructure, chip supply, and real AI workloads can scale as quickly as the modules themselves.