How is it a waste of money if a lot of people, including the author of the article, already have an nVidia GPU in their PC?

Running locally has a lot of advantages - privacy, getting to learn how to run LLMs, not having to deal with quotas, logins, outages.

I use ChatGPT for most practical stuff but I really enjoy running local models. I find it really interesting and I think it's important for people to know how to run these without being beholden to big tech. If you have a used 3090 you can already run some really strong models. There are some really interesting local models as well, like the abliterated ones.

it's nice to be able to have private conversations.

There's also big efficiency increases when batching multiple requests, making clouds inherently more cost effective for normal use cases.

Way better utilization of expensive hardware as well ofc.

Heavily subsidized on heavily price discriminated hardware (nvidia datacenter licensed hardware) kind of cancels out.

Sometimes you don't want your workflow to be subject to the whims of some LLM API provider.