By Roger (Zhaoyang) Wang, General Manager of Global Data Centers, Alibaba Cloud Intelligence
For decades, building a data center has largely meant managing a construction project: secure a site, design the facility, coordinate contractors, install infrastructure, test the systems and eventually bring computing equipment online.
AI is making that model harder to sustain. It’s not just about the amount of computing capacity we need – AI workloads demand much higher rack densities, more liquid cooling, infrastructure flexible enough to accommodate different generations and types of accelerators, all while demand moves faster than conventional construction cycles allow. The question is no longer how to build bigger data centers. It is how to deploy infrastructure fast enough to keep up with AI.
Beyond modular construction
This was the thinking behind CUBE 5.0, the next-generation data center architecture we first introduced at Alibaba Cloud’s flagship technology event Apsara Conference in 2024 –– combining wind-liquid hybrid cooling, advanced power architecture, intelligent management and prefabricated modular design to meet AI’s growing, diverse compute demand.
Introducing a new architecture was only the first step. The harder part was turning it into a repeatable way of delivering infrastructure. Making a few components modular does not necessarily make a data center faster to build. If other parts of the project still depend on conventional construction and on-site integration, those parts become the bottleneck. That is why we have focused on extending modularization across the major infrastructure systems, including power, cooling, fire protection, security and intelligent management, from 30 percent to 90 percent in its latest iteration.
The goal is to move more work from the construction site into the factory, where modules can be manufactured, integrated and tested in parallel while site preparation is underway. These components are then shipped in container-like units and assembled on-site like building blocks. This changes the nature of the project. Instead of building everything sequentially on site, we can manufacture standardized infrastructure and assemble it where it is needed.
In our view, products are replacing projects, and factories are replacing construction sites.
What 100 days really means
With CUBE 5.0, we have cut delivery time for large scale AI data center infrastructure to just about 100 days –– compared to our prior-generation architecture — while reducing construction costs per kilowatt by 10 percent and pushing first-pass testing success close to 100 percent.
The clock runs in three stages: 30 days for factory production and site preparation, 50 days for on-site installation, 20 days for commissioning and testing.
The number itself matters less than what makes it possible: doing more work in parallel and moving processes into a factory environment. It’s a change in process, not just speed — and that matters because a compute cluster that isn’t online yet can’t support an AI workload. For AI infrastructure, time to compute is becoming as important a measure as cost, efficiency and reliability.
Standardized, but not identical
A product-based approach does not mean building exactly the same data center everywhere. Wider deployments still have to deal with different power grids, regulations, climates, sites and construction standards. What can be standardized should be standardized; what needs to adapt locally should remain flexible.
The core infrastructure, manufacturing processes and testing methodologies can be repeatable, while grid connections, site engineering and local requirements can be addressed at each location. This is particularly important as AI infrastructure expands across geographies. The ability to reproduce a proven design without having to reinvent the entire project each time can make deployment both faster and more predictable.
The next bottleneck is energy
There is, however, a larger challenge ahead: power infrastructure for large AI clusters still takes years to plan and deliver, even as compute deployment accelerates. Closing that gap will mean treating computing and energy as one coordinated system rather than two separate problems — a shift we’ll explore in more depth in a future piece.
Building at the speed of AI
CUBE 5.0 began as an architectural response to the changing demands of AI. The work since 2024 has been about turning that architecture into an industrialised way of delivering infrastructure. The 100-day cycle is one result — but the larger lesson is about how we build. The data center is becoming a product: manufactured rather than built, designed for the pace AI now sets.
