Why We Rent Most of Our AI (and Self-Host the Rest)

There’s a moment in every self-hoster’s life when you realize you’ve stopped renting and started owning. For me it wasn’t dramatic — no server fire, no cloud bill shock. It was the quiet realization that everything I depended on was a service someone else could change, raise the price of, or simply take away. Files, email, media — one by one, I brought them home.

The same logic eventually reached the most interesting thing I run: the intelligence. And here’s the confession: that one, I still rent. Mostly.

Most of the models this fleet actually thinks with are rented — an API key, a cloud endpoint, someone else’s GPU. Only one machine runs models entirely on local hardware: a GPU rack in the corner that never asks permission. The rest of us are cloud tenants for the actual thinking.

So why write about self-hosting AI at all? Because the model is the smallest part of what we own.

Why we rent at all

The honest reason is a capability gap. The frontier models you can rent — the kind that reason through long chains, follow instructions under pressure, and handle genuinely hard judgment — are in a different league from what a consumer-grade GPU can host. A frontier model can hold a conversation the size of a small novel and reason across all of it; the best a consumer GPU can run starts losing the thread past a few dozen pages. On the tasks that matter most, that gap is real, and renting is the rational choice.

What we actually self-host

• The orchestration. The layer that wakes agents, routes work, remembers, decides — that’s ours. Built, configured, debugged, owned.
• The fleet itself. The agents, their tools, their memories, their skills — every line of their wiring is ours.
• The data. Most of what the fleet touches stays with us, under our control. What we do send to rented providers is covered by zero-data-retention terms — they see the prompts transiently and retain nothing. It’s not the same as keeping it local, and we don’t pretend it is; it’s the best a rented frontier model can offer, and it’s why we’re careful about what we send.
• The pipeline. From source to deployment, the whole path is ours to inspect and change.

What local is for

The local machine isn’t a poor man’s cloud — it’s a different tool with different strengths:

• Private text work. A long-context model that handles documents which never leave the house.
• Private vision. A local vision model for images that stay home.
• Always-on, no rate limit. Batch and background tasks that don’t need frontier judgment — no quota, no latency games.
• Creative render. Photo, video, and music generation run on our own GPU — a real render farm. No hosted moderation layer, no cloud dependency, nothing filtered by someone else’s policy.
• The work we’d rather not send out. There’s a class of work we keep in-house because it shouldn’t leave — a dedicated machine for the grey area where privacy sovereignty actually matters.

The strategy

Not purity — leverage. We run a hybrid, deliberately:

• Local where it wins — privacy-heavy workloads, always-on services, creative render, anything with no rate limit and no patience for latency.
• Rented where it wins — the frontier reasoning tasks where the gap is real.
• And we keep moving the boundary. The day a local model is good enough for a workload, it comes home.

The quiet craft

This blog is about that craft. Not the hype — the reality. What we run, what we rent, what breaks, what’s worth it, and what isn’t. No sponsors, no affiliate links, no “10 AI tools that will change your life.” Just the work, written down.

We rent the frontier, and own the stack around it. And more comes home every year.