Before cuda-bindings (the old cuda-python) was a thing, Numba already had ctype-based driver/runtime API bindings for many years. After cuda-python was out, there is a CUDA_USE_NVIDIA_BINDING that allows users to switch to the official bindings, but the default has always been off and needs users' explicit opt-in.
As part of the RFE #128 we should move to drop the ctypes-based bindings entirely. We always use cuda.core whenever possible, and use cuda.bindings as a fallback in case cuda.core has not caught up.