Turns Out You Can't Self-Host Your Way Out of Big Tech

I wrote about building a private AI server for my household and landed on a fairly settled note — not a Claude replacement, a complement, wrong goal to chase full parity. Partway through that build, I'd written down a doubt: that I was building towards something very basic that didn't actually replace my privacy concern. That doubt turned out to be more right than I wanted it to be.

A friend and I ended up on a call about it well past midnight, and the mood shifted from "cool project" to something closer to shared resignation. The real question was bigger than what I'd built: is actual data privacy from big tech even possible for an ordinary person anymore? Neither of us landed on yes.

What I'd actually set out to achieve

The bar I'd set for myself early on was specific: as long as my identity and my household's identities stayed protected, no model was being trained on our conversations, and our memory stayed private, that would be enough for now. Not perfect privacy. Just that.

Technically, I hit it. The provider I connected to doesn't train on this kind of API traffic by default, and nothing about our conversations was going anywhere public. What I hadn't fully accounted for was how much of the actual computer underneath all of it still belonged to someone else, regardless of what I'd built on top of it.

I tried to compare retention policies. I couldn't.

Before that call, I'd gone looking at how retention policies actually compare — Claude, Kimi K3, DeepSeek. I wanted a clean answer: this one's meaningfully better, this one isn't. What I got was genuinely muddy. Different scopes, different defaults, different fine print, none of it lining up cleanly enough to say any one protects you meaningfully more than a normal paid subscription. I don't think that's me missing something. I think it's actually that hard to compare, and either way, the average person has no real way to choose between them.

Self-hosting doesn't actually get you out

Here's what undid the whole premise. Even a fully self-hosted setup isn't outside big tech's reach. The GPUs I rented, and the cloud compute behind them, both came from infrastructure the same handful of giants ultimately provide.

I did look seriously at the alternative — buying an A6000 card outright and running it at home, no rented GPU at all. Once I priced out everything else a setup like that would actually need, renting looked cheap by comparison. And running the equivalent fully in the cloud just shifts the same cost problem somewhere else, one that climbs fast the moment you actually use it.

So the honest version of what I built isn't "I escaped dependence on big tech." It's "I rented a smaller, more expensive slice of the same infrastructure, with extra steps."

It wasn't really about privacy

What stuck with me wasn't privacy. It was dependence. As frontier models improve and subscriptions embed deeper into how people work, walking away from them stops being realistic for most people — privacy concerns or not. Underneath that, the open source community — the one real counterbalance to a handful of trillion-dollar companies controlling this much data and infrastructure — is genuinely under pressure from the scale those companies now operate at.

We're both still hopeful open source is resilient enough to find its way through that. But hope isn't an answer, and neither of us pretended otherwise at two in the morning.

Still looking

What actually surprised me wasn't the privacy problem. It was realising I could do any of this at all. One person, no team, no company behind me — just whatever entrepreneurial energy was left after a normal week of work — and by the end of a weekend I had a real GPU-backed system deployed and talking back to me. It's easy to undersell that. But in 2026, a solo person can apparently get something like this working end to end, alone, without anyone's permission or budget approval. I didn't expect that, and it stuck with me more than the technical detail did.

The bigger question — how do I actually protect my family's data from this — I'm no closer to answering than when I started, and pretending otherwise would be dishonest.

I'm still looking. At the moment I feel genuinely stuck — not giving up, just not yet knowing the actual next move. If anyone reading this has a real answer rather than a partial one, I'd like to hear it.