Alibaba Cloud is making a specific claim about the bottleneck behind China’s AI buildout: the constraint is no longer limited to chips. It is also the time required to turn capital, power, cooling, and equipment into working compute capacity. The South China Morning Post reported that Alibaba’s newest modular design, CUBE 5.0, can cut AI data-center delivery to 100 days, a timetable the company says is about half the conventional construction cycle.
That claim arrives as cloud providers face a familiar problem. Frontier-model training and inference require dense clusters of accelerators, but those clusters cannot be productive until they have stable electricity, cooling, networking, racks, and software management. A faster construction model could give Alibaba an advantage in a market where compute capacity is strategic and customers increasingly want infrastructure that can accommodate different kinds of AI chips.
CUBE 5.0 Makes Modular Construction the AI Infrastructure Product
Alibaba presented CUBE 5.0 at the World Artificial Intelligence Conference in Shanghai. The system uses modularized data-center components rather than treating each campus as a fully bespoke construction project. According to China Daily, Alibaba says the design reaches a 90% modularization rate and can deliver an AI data center in as little as 100 days.
The company’s approach is built around what it calls an integrated AI data-center architecture. Instead of designing power distribution, cooling, racks, and operations as largely separate systems, CUBE 5.0 combines high-voltage direct-current power, hybrid air and liquid cooling, and support for heterogeneous accelerators. The architecture is meant to serve workloads that may run on chips from more than one supplier rather than assuming a single dominant hardware stack.
That hardware flexibility is important in China. The country’s AI ecosystem is trying to expand local chip adoption while cloud customers still have software and performance requirements shaped by Nvidia’s CUDA environment. EastFrontier reported this week that China has ordered state-funded data centers to use domestic AI chips, putting further pressure on infrastructure providers to support Chinese accelerators at scale. A modular system that works with heterogeneous chips gives Alibaba a way to sell infrastructure into that transition.
Alibaba also says CUBE 5.0 can accommodate at least three generations of chip upgrades. That is not a small design choice. AI accelerators, interconnects, and memory requirements change faster than conventional data-center construction cycles. If an operator has to rebuild large parts of a facility for each new processor generation, the value of rapid initial construction quickly disappears. Alibaba’s pitch is that its modules can be revised without abandoning the facility’s underlying power and cooling design.
Power and Cooling Define the Economics of AI Compute
The modularity claim matters only if it improves operating efficiency as well as construction speed. China Daily reported that Alibaba expects power usage effectiveness, the common measure of non-compute energy overhead, to reach 1.15 or lower for air-cooled deployments and 1.10 or lower for liquid-cooled sites. Lower PUE means a greater share of electricity is going to computing equipment rather than auxiliary systems.
Those figures explain why liquid cooling is becoming central to AI infrastructure. Dense AI clusters generate more heat than traditional enterprise servers. Air cooling can remain viable in some layouts, but the power density of large accelerator deployments pushes operators toward liquid systems that can move heat more directly. Alibaba’s hybrid design is therefore a response to a practical reality: not every customer will deploy the same rack density, climate profile, or chip configuration.
The company also emphasizes high-voltage DC power distribution. The argument is straightforward. Every conversion between alternating and direct current produces some loss, and large AI campuses contain many power-hungry components. A design that reduces conversion stages can lower wasted energy while improving the room available for computing equipment. Alibaba is framing this as an infrastructure decision, but it is also a commercial one. Energy efficiency affects the long-term cost of training and serving models.
Alibaba’s infrastructure investment sits alongside the wider domestic chip push. The company is not claiming that a modular facility solves China’s accelerator gap. Instead, it is offering a way to deploy mixed hardware at a faster pace while domestic suppliers improve performance and availability. That is consistent with the current reality described in EastFrontier’s examination of why China’s leading AI labs are still training on Nvidia chips: local chip adoption is growing, but software compatibility and retraining costs still shape procurement decisions.
Alibaba’s 100-Day Claim Tests the Market for Faster Capacity
The practical test for CUBE 5.0 will be deployment rather than presentation. Alibaba must show that a 100-day schedule can be repeated across locations with different grid conditions, permitting timelines, cooling needs, and customer chip choices. It must also show that module-based construction does not create new integration problems for cloud clients running complex AI workloads.
Still, the timing of the announcement is revealing. China’s AI infrastructure race is now about more than how many accelerators a company can obtain. It is about who can put them into service first, at the lowest energy overhead, and with the flexibility to accommodate evolving hardware. The capital commitments are large, but the winner will not necessarily be the operator that announces the biggest campus. It may be the operator that converts a planned site into usable AI capacity most quickly.
Alibaba’s claim of 100-day delivery puts that conversion rate at the center of its data-center strategy. If CUBE 5.0 performs as advertised, the company will have turned modular construction into a competitive AI product. If it does not, the announcement will still have clarified the new benchmark China’s cloud providers are being asked to meet: build faster, cool denser clusters, and avoid locking the next generation of AI compute into a single chip architecture.