What infrastructure (compute, network, and storage) will support OpenLLaMA 7B model training, fine-tuning, and inferencing?

To run Open LLaMA 7B (or similar sized conversation model), I am looking for infrastructure sizing guidelines. I have around 10PB of data collected from various sources along with our own org data. What kind of compute - CPUs/GPUs, network, and storage is recommended. Note that this is for training, fine-tuning and inferencing for enterprise, not for hobby project.