It's different, more lightweight hardware for inference but can be in the same data center. Training requires beefier GPUs.