Modular vs traditional data center: how to compare
A tailored suit and a well-made off-the-peg suit can both look good. The tailored one fits one person exactly but takes weeks and several fittings. The off-the-peg one is ready sooner, comes in known sizes, and can be swapped if it stops fitting.
Data centers face a similar choice. This piece compares modular and traditional builds on the things that matter most once a site is running: how long it takes to build, how it grows, and how it is looked after. It uses no product figures, because the right numbers depend on the site.
Two ways to build
A traditional data center is usually one large building, designed for its location. Power, cooling and network systems are planned together and installed inside the shell. It is often sized for the full capacity the owner expects to need over many years.
A modular data center is assembled from standard units made in a factory. Compute, power, cooling and network can each come as separate modules, or be combined. Capacity is added by adding units. (We explain the basics in What is a modular data center?)
Neither is better in every case. They make different trade-offs.
Side by side
| Question | Traditional build | Modular build |
|---|---|---|
| How is it built? | Mostly on site, in sequence | Mostly in a factory, then connected on site |
| What sets the pace? | Construction, then fit-out | Power, permits and groundwork; units are built in parallel |
| How is it sized? | Often for full future capacity up front | In steps, unit by unit |
| How does it grow? | Fill empty halls, or build a new phase | Add more units beside the first |
| How easy is a design change? | Harder once the building is up | Newer units can follow a newer design |
| How is it maintained? | On-site teams work on fixed systems | Units can be serviced or swapped as a whole |
| Where does it suit? | Large campuses with long-term demand | Sites near available power or near users |
| What are the main risks? | Long build; capacity sitting idle early | Many small sites to run; site prep still needed |
Build time
In a traditional build, much of the work waits on earlier steps. The shell has to exist before cooling and power can go in. Delays early in the job tend to push everything back.
Modular builds move a large share of the work into a factory. That work can happen while the site is being prepared. The catch is that power, permits and fibre still have to be ready, and these are often the slowest parts of any project. (See Spare power is not the same as usable power.)
Expansion
A traditional site is usually built big, then filled over time. That can make sense when demand is steady and well understood. It also means paying for space, power and cooling that may sit unused for years.
A modular site starts smaller and grows by adding units. Each step can follow what demand actually does. It also means the site layout has to leave room, and power, for the units still to come. (We look at this in Adding AI capacity one unit at a time.)
Upkeep
In a traditional building, systems are often shared across a whole hall. Work on one part can affect many racks, so maintenance is planned carefully around them.
In a modular design, each unit tends to be more self-contained. A faulty unit can be repaired or replaced without touching its neighbours. The trade-off is that there are more separate units, and often more sites, to monitor and keep consistent. Good remote monitoring and a standard design across units help a lot here. (See Why standard units matter.)
Build time, expansion, upkeep
One building for its site
- Build time
- Mostly on site, in sequence; early delays push everything back.
- Expansion
- Often built big, then filled over time.
- Upkeep
- Shared systems; work on one part can affect many racks.
Standard factory-made units
- Build time
- Units are made in a factory while the site is prepared.
- Expansion
- Starts smaller and grows by adding units.
- Upkeep
- A faulty unit can be repaired or replaced on its own.
Fit with AI hardware
AI machines draw more power and give off more heat per rack than most older computing. Cooling methods are shifting as a result. (We cover this in Air cooling or liquid cooling for AI?)
A traditional building designed years ago may need major work to take denser racks. A modular site can add units built for the newer hardware. A new traditional build can also be designed for high density from the start, so age matters as much as the method.
Questions that decide it
The right choice depends on the site and the demand. These questions usually settle it.
- How much power is available now, and how much later?
- How confident are we in demand over the next few years?
- How close does capacity need to be to users or data?
- What rack density and cooling method does the hardware need?
- Who will operate the site, and from where?
- What happens to workloads if one site goes offline? (See Data center outage: what happens when one site goes down.)
What we are building
We're building on the modular path because we expect it to let capacity follow power and demand.
We're building standardized compute close to power, connectivity and demand, joined by one programmable network for AI inference. The network is designed to place workloads by policy, on latency, data residency, model and load, and every workload is designed to generate an auditable record of commitment, placement and delivery.
The method matters less than whether the capacity is there when the work needs it.
If you need AI capacity and want to know where it would run, use Request capacity below.
Related
Modular data center: why standard units matter · Edge data center vs hyperscale: many small or a few big? · Data center outage: what happens when one site goes down · Starfield Fabric