19 August 2026
Ten units, one rack: running a private AI fleet at home
The pilot sells four seats. Behind those four seats we are standing up ten LIZ units on our own hardware and testing them like a fleet. This is the first entry of the field journal.
A unit is not an account. It is one OS process, one SQLite file, one vector collection, and its own signed Android build — the same isolation story the platform page draws with boxes and arrows. Ten units therefore means ten of everything: ten processes to keep warm, ten databases to back up, ten builds to sign, ten certificates to rotate. That multiplication is the whole experiment.
Why ten, when the public pilot is four? Because four seats proved the shape, and ten stress-tests the boring parts. The model was the easy piece: a dense 14B held warm on the GPU answers in about two seconds whether one unit is talking or ten. The fleet plumbing was not easy. Provisioning a unit by hand is charming exactly once. An update is no longer an update — it is ten updates, and the honest question is what happens to unit six when units one through five went fine.
Some early entries for the record. A fleet needs a board: one page that says which unit is up, which build it runs, and when it last answered — the moment we had more units than fingers on one hand, memory stopped being a monitoring strategy. Per-unit builds earn their keep: when a phone misbehaves we know exactly which unit, which build, which database, and nobody else’s data is in the blast radius. And capacity really is a schema fact, not a config line — adding units means changing what the system admits is true, which is slower than editing a file and much harder to fool yourself with.
What we measure is what the landing already publishes: reply latency around two seconds, speech-to-text around a third of a second at the median — now checked per unit, not on the best one. The status table stays honest because this fleet is what checks it.
If you hold a seat, this journal is why the seat keeps working. Capacity will grow when the hardware does, not when the marketing does. More entries as the numbers land.