On-site power for data centers, explained

Think of a farmhouse with its own well. It does not wait for the town to lay a water main. But the owner now looks after the pump, tests the water and plans for dry summers.

On-site power works in a similar way for data centers. Instead of drawing all its electricity from the grid, a site makes some or all of it nearby. This piece explains what that means, the common forms it takes, and the trade-offs a buyer of AI capacity should understand.

What "on-site" means

On-site power, sometimes called behind-the-meter power, is electricity generated at or next to the data center. It reaches the computers without travelling across the public grid first.

The term covers several set-ups. Some sites run entirely on their own generation. Others use it alongside a grid connection, to add capacity or to cover gaps. Many sites also keep backup generators, which is a different thing: those run only in an emergency, not every day.

It helps to ask which one a provider means. "We have on-site power" can describe any of these.

Why builders choose it

The main reason is time. A grid connection can involve a long wait. (We explain why in "Why grid connections hold up AI data centers".) Generating power locally can let a site open sooner, or grow without waiting for a larger connection.

Other reasons include:

• Control. The operator can plan its own supply rather than depend on upgrades it does not control.
• Location. It can make sites possible where the grid is thin but fuel or other resources are close.
• Using what is there. Some sites sit beside existing generation whose output is not fully used.

The common forms

Gas-fired generation. Engines or turbines run on natural gas. They can provide steady output, but they burn fuel, produce emissions and depend on a gas supply.

Solar and wind with storage. Panels or turbines generate when the sun shines or the wind blows. Batteries can smooth short gaps. Steady round-the-clock supply from these alone needs a lot of land and storage.

Fuel cells. These turn fuel into electricity through a chemical process rather than burning it. They are quiet and modular, and their environmental profile depends on the fuel used.

Other options. Small nuclear designs and other technologies are discussed for data centers. Most are at early stages, so treat timelines in announcements with care.

The trade-offs a buyer should know

On-site power solves some problems and creates others. None of these is a reason to avoid it. They are reasons to ask good questions.

Reliability moves to the operator. The grid has many generators sharing the load. A single site has fewer. The operator has to plan for maintenance, faults and fuel interruptions itself. (We discuss what happens when a site goes down in What happens when one site goes down?.)

Cost depends on fuel and equipment. On-site generation may cost more or less than grid power. It depends on fuel prices, equipment and how fully the plant runs. That cost tends to show up in what you pay, even if it is not itemised. (See The costs that aren't on the AI quote.)

Emissions and permits. Burning fuel on site brings air permits and emissions reporting. Rules differ by place. A buyer with its own climate targets should know what the power source is. (We cover how to ask in "Renewable power and AI: what a buyer can ask".)

Neighbours. Generators can mean noise, traffic and visual impact. Projects that plan for local concerns early tend to face fewer delays later.

Lock-in to the site. On-site plants are long-lived. A site designed around one fuel may be hard to change later.

Questions worth asking

If a provider says its capacity runs on on-site power, these questions help:

1. Is on-site generation the main supply, a top-up, or backup only?
2. What is the fuel or energy source, and how is it supplied?
3. What happens if the on-site plant needs maintenance or fails?
4. Is there also a grid connection, and how large is it relative to the load?
5. Which permits apply, and are they in place?
6. How is the cost of this power reflected in my price?

What we are building

Our view is that the power source should be known and stated, whatever it is.

Starfield deploys standardized compute close to power, connectivity and demand, then connects it through one programmable network for AI inference. Workloads are placed by policy, on latency, data residency, model and load, and every workload can generate an auditable record of commitment, placement and delivery.

We're building the network so that work can be placed across more than one site. It is designed so that a single plant's outage does not have to mean a stopped workload.

If you own a site with power available, on-site or from the grid, use Request capacity below.