Data center site selection: spare power is not usable power

A site with unused electricity looks like a gift. The power is there, the meter is barely moving, and AI needs power badly. Plug in some servers and you're done.

Almost. Power is difficult to get, and it is still only the first thing. Between spare power and a site that can serve AI every day, there are four more layers: the connection, the cooling, the network and the people. This is the practical side of data center site selection, and it is why AI capacity is moving toward the power only where those layers can be met.

Layer 1: the connection you actually have

"We have spare power" can mean several different things. It might mean the site's grid connection is rated for more than it uses today. It might mean there is generation on site, such as solar or a turbine, that produces more than the site consumes. It might mean the local network has headroom at certain hours.

Each of these is useful, but they are not the same. A rated connection that is used half the time, for example, might only be free at night. Generation that rises and falls with the weather needs a plan for the hours it doesn't. And the electrical equipment inside the site, from switchgear to distribution boards, has to be able to carry the new load safely.

So the first question is not "how much power is there?" It is "how much power can we draw, reliably, at any hour, through equipment that can take it?"

Layer 2: getting the heat out

Every watt that goes into a computer comes out as heat. AI hardware is dense, so it produces a lot of heat in a small space. A room that was comfortable for office servers can overheat quickly with AI equipment in it.

That means a usable site needs a cooling plan matched to the hardware. Depending on the equipment and the climate, that might be air cooling, liquid cooling or a mix. Cooling also uses power itself, which eats into the spare power you started with. And it needs space, water or both, depending on the design.

A site with spare power and nowhere to put the heat is a site that can run for a few minutes.

Layer 3: a network that can keep up

AI inference is a conversation with the outside world. Requests come in, answers go out, models and updates arrive from elsewhere. For that, a site needs good network links, ideally more than one, so a single cut cable doesn't take it offline.

Some places with spare power are well connected. Industrial parks and towns often have fibre nearby. Others, like a remote generation site, may need new links laid before they can do anything useful. That can take longer than people expect, because it involves permits, digging and a carrier willing to do the work.

Layer 4: people and process

Hardware breaks. Fans fail, drives wear out, cables come loose. A usable site needs someone who can get there, swap a part and close the door behind them. It also needs security, monitoring, spare parts and a clear routine for who does what when something goes wrong.

For a big campus, that means a team on site around the clock. For a small site, it more likely means remote monitoring plus a local technician or partner who can be there when needed. Either way, someone has to own it.

A simple example

Picture a food factory that has cut a shift. Its connection was sized for three shifts, so in the evening it has power to spare. The building has a loading yard with room for a container of equipment. Fibre runs along the road outside. A maintenance team is on site every weekday. That site is not ready yet, but every layer has a starting point, and that is what makes it a candidate.

This is an invented example for illustration, not a real site.

Why repeatable designs help

Every layer above takes planning. If every site is designed from scratch, every site takes a long time. That is one reason standard units matter: when the same design for power, cooling and network is used again and again, each new site needs less custom work. The questions are the same each time, and so are many of the answers.

A checklist for a site with spare power

If you own a site with power you're not using, these questions help sort out what you have:

• Power: How much can we draw, at what hours, and through what equipment? Is it from the grid, on-site generation or both?
• Space: Is there room for equipment, either indoors or in a secure outdoor area?
• Cooling: What can we do with the heat? What does the climate allow?
• Network: Is fibre nearby? Can we have two independent links?
• Access: Can a technician reach the site quickly? Who holds the keys?
• Permits: What local approvals would new equipment need?

None of these is a reason to stop. They're the list of things to get right so that spare power becomes something useful.

What we're building

We're building Starfield Fabric around this idea. Starfield is designing standardized compute to be deployed close to power, connectivity and demand, then connected through one programmable network for AI inference. Workloads are designed to be placed by policy, on latency, data residency, model and load, and every workload can generate an auditable record of commitment, placement and delivery.

The work starts with a site that has power, and then turns it into a site that can serve AI every day: connected, cooled and looked after. Many such sites, run as one network, are designed to add up to real capacity.

If you have a site with spare power and want to know what it would take, talk to us.