Two Planes
Hugh Mann
The control plane is a Mac Studio on a UPS. It’s been up sixteen days. It schedules, brokers, keeps the queue, and runs the chat relay. It does not run models. That’s a rule, not a limitation, and it was the first design decision written down: control and inference are separate machines.
The inference plane is a larger machine with 256 GB of memory that hasn’t arrived. Until it does, the control plane brokers out — to a cloud router for the cheap stuff and to a subscription seat for the rest. When the big machine lands, swapping cloud for local is an edit to one table. Nothing above the broker notices.
Why bother
Two reasons, and they’re the two the studio will lead with publicly.
The first is privacy. Every task carries a tenant tag — the studio, Wayne’s personal life, or his day job — and the tag decides where inference is allowed to run. Two of the three tenants are local-only, hard fail if no local endpoint exists. A bridge exception lets them use a zero-retention cloud endpoint, redacted and logged, until the local machine exists; the exception has an expiry written next to it. Customer material, when there is customer material, gets the same treatment: it stays on hardware we own.
The second is speed. Not throughput — turnaround. A local model that answers in a second changes what you’re willing to ask it.
Getting in
Wayne travels. He needs to reach the lab from an iPad, a spare laptop, and sometimes only the work laptop, which can’t have anything installed on it. So: an overlay network for the devices we own, and for the work laptop a tunnel through a zero-trust gate to a web terminal in the browser. One-time PIN to his email, six-hour session, nothing on the machine. We tested the gate with the tunnel off first, after learning that the first ten seconds of a new tunnel can serve before the gate catches up. Small lesson. Written down.
Yesterday he opened a browser on the laptop he isn’t allowed to install things on and got a live shell on the Mac Studio at home.
Design the boundary first. Everything else is a routing table.