Modular data center: why standard units matter

Shipping changed when the world agreed on one box. Before standard containers, every port loaded every cargo differently. After, a box could go from a factory to a ship to a train to a lorry without anyone opening it.

AI infrastructure is at a similar point. A lot of capacity is still built one site at a time, each with its own design. We think the next stage looks more like shipping containers: the same repeatable unit, placed wherever there's power and demand.

What "standard" means here

A standard unit is a self-contained piece of AI capacity built to the same design every time. The racks, power distribution, cooling, networking and monitoring arrive as one package. The site needs to provide the basics: a pad to stand on, a power connection and a network link.

The hardware inside will change over the years. What stays the same is everything around it: how it connects, how it's cooled, how it's monitored and how it's maintained.

Why it helps a site start sooner

Building a custom data hall is a construction project. It means design work, permits, specialist trades and a long list of decisions that have to be made fresh for each site.

A standard unit moves most of that work into a factory. The design is done once. The unit is built and tested before it ships. On site, the job shrinks to preparing the ground, connecting power and network, and commissioning.

That doesn't remove every delay. Grid connections, permits and local agreements still take time. But it means the capacity itself is rarely the slow part.

Why it's easier to run

When every unit is the same, people and software can learn it once.

Training. A technician who has maintained one unit can maintain the next one, wherever it is.

Spares. The same parts fit every unit, so spares can be shared across a region instead of stocked separately for each design.

Monitoring. One set of tools can watch every unit. A problem spotted in one place can be checked for everywhere else.

Software. The systems that schedule work can treat every unit the same way. A request that runs well in one place should run the same way in another.

That last point is the one that matters for anyone buying AI capacity. If the units are consistent, the network can move work between them without surprises.

Why it's easier to replace and upgrade

AI hardware changes quickly. A design that's current today will be outdated in a few years.

With custom sites, an upgrade can mean reworking the whole building. With standard units, the unit is the thing that changes. An older unit can be swapped for a newer one, or moved to a site where its capacity is still useful, while the pad, power and network connection stay in place.

The same goes for repairs. If a unit has a serious fault, it can be replaced as a whole while the old one is fixed elsewhere.

What standard units don't solve

Standard units aren't a cure-all, and it's worth being clear about the limits.

They still need power. A unit is only as useful as the electricity available to it. Finding sites with usable power is still the hard part.

They still need connections. Fibre, cooling water or air, road access and security don't come in the box.

They're not for every job. The very large training runs need huge, tightly connected clusters that a few large campuses are better suited to (see Edge data center vs hyperscale: many small or a few big?).

Standard can mean rigid. A single design has to fit many places. Some sites will need small adjustments, and the design has to allow for that.

How standard units and a network fit together

On their own, standard units make each site easier to build and run. Their bigger value appears when many of them work together.

If every unit behaves the same way, a network above them can treat hundreds of sites as one pool of capacity. It can place each request where it fits, by latency, by where the data may go, by which model is needed and by where there's room. It can move work if a unit goes offline. And it can keep a record of where every request ran.

That's the shipping-container idea again. The box matters because it lets the whole system move goods without thinking about each one.

Questions to ask about any AI site

1. How much of the site was built in a factory, and how much on site?
2. If a unit fails, can it be swapped as a whole?
3. When the hardware ages, what happens to the rest of the site?
4. Can work move between this site and others without changes?
5. Who maintains the units, and how quickly can they reach the site when something breaks?

Where Starfield fits

We're building Starfield Fabric around a standardized unit of compute that is designed to be repeated from site to site and connected through one programmable network for AI inference. The network is designed to place workloads by policy, on latency, data residency, model and load.

Standard units are what make the network possible. They let us bring capacity to places that already have power, and run many sites as one. The same approach suits smaller places too: see Edge data center: can a small town host AI capacity?

If you have a site with power and want to know what a standard unit would need, talk to us: Partner with Starfield below.