What is a modular data center?

Think about the difference between building a kitchen from scratch and fitting one from a catalogue. In the first case, every cupboard is measured, cut and finished in your house. In the second, the units are made in a factory, tested, and arrive ready to fix to the wall.

Data centers can be built either way. This piece explains what a modular data center is, what is actually inside a module, and why this approach changes how quickly AI capacity can start working.

A data center, broken into parts

A data center is more than a room full of computers. It is a set of systems that work together: power coming in, cooling taking heat away, a network linking machines to the outside world, and a structure that keeps it all safe.

In a traditional build, most of those systems are designed for one specific site and assembled there. A modular data center splits them into standard units. Each unit is designed once, built in a factory, and shipped to wherever it is needed.

What a module usually contains

Modules come in different shapes and sizes, but they tend to fall into a few kinds.

• Compute modules hold the racks of servers and GPUs that do the actual AI work.
• Power modules take electricity from the grid or a local source and condition it for the machines.
• Cooling modules remove the heat that the computers produce. (We look at why that heat matters in Why AI chips need so much cooling.)
• Network modules connect the site to fibre and to other sites.

Some designs combine several of these in one enclosure. Others keep them separate so each can be sized and replaced on its own.

Why factory-made changes the timeline

A traditional data center is built in sequence. The ground is prepared, the building goes up, then power, cooling and IT systems are installed inside it. Each step usually waits for the one before.

With modules, much of that work happens in parallel. The site is prepared while the units are being built and tested elsewhere. When they arrive, the remaining work is mostly connecting them up.

Factory work also tends to be more predictable. Weather does not stop it, the same team builds the same unit many times, and problems are found on a test bench rather than on site. That does not remove risk, but it moves a large part of it earlier, where it is easier to manage.

What still takes time

Modules do not make a site appear overnight. Several things sit outside the factory and often set the real pace.

Power. A site needs a grid connection or local generation that can actually deliver. (We cover the gap in Spare power is not the same as usable power.)

Permits. Planning, environmental and building approvals still apply. Rules vary by place, and this article is not legal advice.

Connectivity. Fibre has to reach the site with enough capacity and, ideally, more than one route.

Groundwork. Foundations, access roads and security still need to be built locally.

Modules can shorten the part of the job that happens after power and permits are in hand. That part can be large, which is why the approach matters.

Why this suits AI in particular

AI hardware changes quickly. A data center designed around one generation of machines may struggle with the next, especially as each rack draws more power and gives off more heat.

Standard units make it easier to keep up. A module can be designed for a known rack density, then a newer design can be added beside it later. (We explain why consistency pays off in Why standard units matter.)

Modules can also be placed where the power already is, rather than waiting for power to reach a large central campus. (That shift is the subject of Why AI capacity is moving toward the power.) Smaller sites close to demand can also help answers travel shorter distances, which matters for inference.

Questions to ask about a modular site

If you are buying capacity from, or hosting, a modular site, a few questions help separate a real plan from a brochure.

1. What is in each module, and what is built on site?
2. Has the power connection been agreed, and when can it deliver?
3. What rack density is each compute module designed for?
4. How is heat removed, and does that method suit the hardware planned?
5. How are modules replaced or upgraded without taking the whole site offline?
6. How does the site connect to others if one location has a problem?

What we are building

Our approach rests on the same idea: repeatable units, placed where the power and the users are.

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 factory makes the unit; the site, the power and the network make it useful.

If you have power and land and want to put it to work for AI, use Request capacity below.