How long until a rented GPU runs CUDA? Launch to first matmul on six GPU types, timed

Six single-GPU machines launched through an API, timed from create call to first SSH and first CUDA matmul. 3 to 12 minutes, and why.