Commissioning & Go-Live

Commissioning turns design intent into test evidence. An AI factory needs two interlocked acceptance tracks, facility and cluster, and you must sequence their gates on purpose or the schedule will sequence them for you.

Every commissioning test needs a pre-agreed observable gate and a witnessed result. An AI building is accepted when facility results and cluster results cover the same power, cooling and workload boundary.

Electrical acceptance has to prove, with instrumented results, that the power chain holds a GPU cluster's synchronized megawatt-scale swings. No static load bank reproduces that load, and a chain never tested against it finds out on the first real workload.

Behind the meter you are the grid. Commissioning proves that generation, controls and storage keep an AI campus of up to a gigawatt running when the utility lets go or the largest generator trips, and can restart it from dead in a black start.

Commission cooling in three layers: surrogate load and injection or HIL testing prove the controls and fail-safe actions, liquid-cooled load banks on the manifolds prove the liquid path, and staged workloads confirm operation without making live GPUs the fault target.

IST proves the building survives its design-basis faults, by load-bank trip or scoped injection, at a boundary agreed with the utility, AHJ, insurer and Cx authority. Load banks cannot fully reproduce synchronized GPU swings or cold-plate heat, so staged workloads must prove those before production.

Fabric defects surface quietly, one marginal optic and one mis-cabled rail at a time. Every defect you fail to screen at layer 1 comes back in production as a straggler or a stalled all-reduce, where it costs days instead of minutes.

Burn-in makes a GPU fleet's infant-mortality defects fail on your acceptance clock, under controlled stress, before a bad node can restart a synchronous job. Release a cohort on diagnostics that actually ran plus correctness and thermal results, not on a timer.

A cluster's performance is accepted through the production scheduler and storage. Training needs ML Productivity Goodput measured over a sustained reference training run; inference needs correct responses within its contracted latency. Neither can average away a failed safety, correctness, isolation or recovery gate.

Go-live ramps megawatts and synchronized GPU load in stages behind an operational-readiness gate. Ramp faster than the grid and cooling can absorb, or hand over before operations is ready, and a finished building can become an outage.