AI training campuses are being built faster than the utilities serving them can add power capacity. The traditional model — wait for the utility to build a substation on their property — does not match the speed at which hyperscale operators are promising AI compute to investors.
The way developers bridge this gap is with what is called a "customer-side" substation. It is not a new idea — large industrial customers have used these for decades — but the scale and speed at which AI campuses are using them has changed how the work gets organized.
What "customer-side" actually means
A customer-side substation sits on the developer’s land, downstream of where the utility’s power lines reach. The utility delivers high-voltage power — usually 138 kV, 230 kV, 345 kV, or 500 kV — to the property line. From there, the developer’s substation steps the voltage down to what feeds the data halls.
For AI campuses, the step-down is usually to 34,500 volts (the modern standard) or 13,800 volts (older buildings). Inside the campus, that medium voltage feeds smaller transformers at each data hall, which step it down again to the 480 volts that IT equipment uses.
The "customer-side" label matters because the utility’s regulatory and capital-planning rules stop at the property line. The customer pays for the substation, owns it, and lives with it — but builds it on their own schedule, not the utility’s.
Why customer-side is faster than the utility side
Utilities are regulated companies. Their capital projects go through internal planning, get filed in rate cases, and have to fit into reliability priorities for the whole system. A utility substation built for one large customer typically takes 24 to 36 months from agreement signature to going live.
Customer-side substations live outside that timeline. A developer can start construction the day after the agreement is signed, and the substation can be under construction within 60 days. The pacing constraint becomes how long it takes to get the equipment — large transformers and switchgear run 50 to 70 weeks of lead time as of 2026 — not how long the utility takes to approve a project.
For a 600 MW AI campus, that timeline difference is the gap between first servers coming online in month 14 versus month 28. Investor commitments rarely tolerate the longer timeline.
Backup power paths to each building
The standard setup for AI campus power is to give each data hall two or more independent paths for power. Each hall has its own cable runs, its own transformers, and its own switching equipment. A problem on one path cannot take the hall offline — the other path keeps the servers running.
This backup pattern comes from how data centers design their cooling and IT equipment — utilities themselves rarely build power lines to this standard. The doubled cable, transformer, and switchgear cost is substantial. For AI workloads where a single-hall outage represents millions of dollars per hour in lost compute time, the math obviously works.
Some operators are now exploring even more redundancy — two completely separate power paths to the most critical halls. The trade-off is the cost of all that extra cable versus the cost of outages on AI training runs that can last weeks and cannot tolerate interruptions.
Powering up the campus in stages
The other major schedule constraint on AI campuses is how the system gets tested and energized. A campus that finishes construction in month 14 but waits until month 18 to finish testing has not really delivered first power in month 14 — it has delivered first power in month 18.
On modern AI campus builds, testing and energization happen one building at a time, in parallel with construction. The first building gets its power paths tested and energized while the second building is still being built. This requires close coordination between the contractor, the developer, and the operator. The protection system has to be working and tested on the first building before the rest of the campus is even constructed.
Done well, this approach can deliver first server power 6 to 8 months earlier than if everything happened in sequence. Done poorly, it results in a half-powered campus with a protection system nobody trusts.
Temporary power: an emerging pattern
When the utility connection itself is the bottleneck — not the customer-side construction — some developers are now adding temporary on-site power. The pattern is straightforward: a gas turbine power plant, typically 30 to 60 MW, connects to the customer-side substation and delivers power while the utility connection is still being built.
Temporary power lets the campus take first server load on the developer’s schedule rather than the utility’s timeline. When the utility connection is completed, the temporary plants either pack up and leave or stay as backup. The economics work when the alternative is sitting on $500 million of AI servers waiting for power.
Our parent company operates these temporary power plants and pairs them with our customer-side substation builds. The combination — customer-side substation plus temporary power — has become a real schedule path on several recent campus builds where the utility connection was the limiting factor.