My recipe application is going to be written in a purely functional language with CQRS and a message bus that provides "guaranteed once" determinism. I should be able to spin up no more than a half dozen containers on two VM's to handle this architecture. The database will probably be redundant MSSQL and Oracle (just in case one technology is fundamentally broken). Database writes will be proxied by a Redis instance. I will use 3x 3080 GPU to train a ML model used for scaling the recipe depending on the number of people you wish to serve. Hmmm...I might need a third VM.
You can't do any serious neural stuff on 8GB of VRAM. I'd say save yourself some pains and buy the $3000 professional models with ECC so you can train transformers.
