Will Modularization Become the New Form of AI Data Centers?
In the past, data centers were more like large-scale engineering projects. Companies had to redesign each facility based on land availability, power supply, server configurations, and customer requirements. Every project involved a significant amount of customized construction. But the AI era is beginning to change this logic.
The question is no longer whether enterprises need computing power. Instead, the growth rate of demand is now outpacing traditional construction methods. A new tension is emerging: AI needs more and more data centers, while traditional construction methods are becoming increasingly slow.
This is why modular data centers are starting to attract attention. The logic behind them is actually very similar to manufacturing. The auto industry moved from manual production to assembly lines to improve efficiency; data centers are moving from “on-site construction” toward “modular production” to improve deployment efficiency. By completing standardized designs in advance and integrating components such as power systems, cooling systems, and server racks, facilities can then be rapidly expanded based on demand. This gives AI companies the potential to gain access to usable computing resources much faster, without having to wait through construction cycles that can take years.
There is also a deeper reason behind this shift. Future AI demand may not be concentrated entirely in a handful of hyperscale data centers. Large-scale model training certainly requires massive facilities, but many AI applications—such as enterprise AI assistants, industrial AI, and real-time inference services—are more concerned with latency, cost, deployment location, and data security. This could drive the emergence of more regional and smaller-scale computing nodes. Data centers may gradually evolve from “large centralized facilities” toward a combination of “cloud centers + edge nodes.”
As a result, the market is beginning to pay more attention to companies that can provide infrastructure solutions. For example, $VRT could benefit from growing demand for power management and liquid cooling systems in AI data centers, while $DLR has a long-standing presence in data center infrastructure and colocation services. At the same time, some companies are exploring opportunities in modular computing deployment. $MAAS’s Stars project is one such attempt to develop standardized computing units. Compared with building hyperscale data centers, modular nodes place greater emphasis on rapid deployment and flexible expansion.
Of course, this approach is still in its early stages. The real test is not whether the concept can be presented, but whether modular solutions can actually reduce costs, be replicated quickly, and generate sustained demand.
I believe AI infrastructure could undergo an interesting transformation in the future: in the past, a data center was a “capital-intensive engineering project.” In the future, it may increasingly resemble a “replicable infrastructure product.” In this process, beyond the giants that control core chips and cloud platforms, smaller players focused on deployment efficiency, energy utilization, and regional computing capacity may also find opportunities of their own.
