everyone's shoving an llm into their homelab now and i don't get it

i've been reading through a bunch of "here's my 2026 homelab stack" posts lately and there's a pattern that's kind of annoying me. two years ago the answer to "what should i run" was jellyfin, immich, a reverse proxy, maybe home assistant if you're into that. now every single one of these lists has an "ai section," like it's just assumed you're running ollama and some webui and wiring it into your email and your notes and whatever else.

and look, i get why it happened. ollama made running a 7b model stupidly easy, you don't need a beefy gpu anymore, and there's a genuine "look what i can do locally" appeal to it. but somewhere along the way "you can run a local llm" turned into "you should be running a local llm," and those are not the same claim.

what problem is this actually solving

my homelab does backups, media, a couple dashboards, dns blocking. every one of those services exists because it replaces something i'd otherwise pay for or hand over to a company i don't trust. that's the whole point of self-hosting for me, cut out the middleman, keep my data mine.

a local llm doesn't really fit that logic the same way. i'm not trying to replace claude or chatgpt with a worse, slower version running on a machine that also needs to encode video and serve photos to my phone. if i wanted a mediocre assistant that's slower than the cloud version and occasionally forgets what i asked it, i'd just, not do that.

the "n8n + ollama = private ai automations for $0/month" pitch sounds nice on paper but the actual output of most of these setups is triaging email or summarizing notes, stuff that a normal filter or a five minute read already handles fine. you're spending gpu cycles and ram to save yourself a task that wasn't that expensive to begin with.

my setup is too weak for this anyway, and that's not the point

i don't run any of this, partly because my homelab hardware genuinely isn't built for it (nobody's running a 13b model comfortably on what i've got), but even if it were, i don't think i'd bother. it's not a "someday when i upgrade" thing. the actual bottleneck isn't compute, it's that i can't point to a task i'm doing right now that a local model would meaningfully improve.

that's the part that gets lost in these roundup posts. they list ollama and open webui next to jellyfin and nextcloud like they're solving the same category of problem. jellyfin replaces a subscription. immich replaces uploading your photos to google. what does a local llm replace, exactly, that wasn't already solved or wasn't a problem worth solving?

the self-hosted crowd copying ai hype instead of resisting it

the whole appeal of this community used to be doing things deliberately, running stuff because you actually needed it, not because it was trendy. and now half the "what's in my stack" posts read like they're chasing the same ai wave as every saas company shoving a chatbot into their product, except now it's happening in the one corner of tech that was supposed to be about opting out of that.

if you've got an actual use case, wiring a local model into something you already do daily and it saves you real time, that's different, that's the "depends on the use case" answer nobody wants to hear because it's less exciting than a roundup post. but running ollama because it showed up on every list this year isn't a reason. it's just following the same hype everyone else is following, just self-hosted instead of cloud-hosted.