An islanded AI power plant: twelve 11.2 MW-class gas gensets, a fast stabiliser and a 100 MW GPU hall. Every training step swings the load by tens of megawatts. Watch the bus frequency.
Concept demo with illustrative parameters, not Bergen or Piller data: gas genset inertia H = 1.5 s, isochronous PI governor with 0.7 s engine lag and 18%/s ramp limit, trip outside 47.5–51.5 Hz for 120 ms. It shows the shape of the problem. It is not a protection study.
Why islanded AI plants need a supervisory layer above the stabiliser, and what an eight-week Phase 1 study with Bergen Engines On Land would deliver.
A fast stabiliser answers the milliseconds and the engine governors answer the seconds. Someone still has to decide how many engines run, how much reserve to hold for the next step, and what the data centre is about to do.
A dynamic plant model calibrated on real load-step tests, run against measured AI training load profiles for two reference plants.
Open-source plant controller: frequency state machine, Modbus genset drivers, flexible-load curtailment, failsafe logic, automated tests. github.com/darkknight4563/mithra
Ali Emadi · Gatekeeper AI · Oslo · emadiali83@gmail.com · +47 920 93 908