It doesn't make it a problem because there is no reason other than weakness of current gpus, all of it could have been executed on a single large pc. Or as you can see, smaller models already run on single pc's.
You need to think through the entire request for an llm model response end to end.
What's the system, the thing, that's experiencing consciousness when you talk about llms? Again, if you think through the request path of each token that's generated, this idea falls apart.
every time you ask for the next token you different gpus on different servers are used. each generation of a token is an independent operation from the previous.
a brain is a system of neurons, you can add or remove some, it still some how generates a conscious experience. an llm generating tokens is doing math in a distributed way across what ever gpu happens to be free for that isolated request. there is no consistent system. what do you consider to be the conscious entity? the random selection of gpus? you think something is conscious one token at a time? what is the exact thing that is thinking? the way you can point to a body, you cant do that for an llm string of tokens. thats where i think it falls apart. your doing math that eventually reduces into a final number. theres no one system that continues to run and take input in.