thought leadership
Air Gap, Open Weights, and the Plant That Cannot Call the API
Many plants cannot send a drawing to Virginia. ITAR, air gaps, a customer who will not let geometry leave the site, a network that is supposed to be boring on purpose. Boards still buy Cloud Copilot for those sites. After Fable you would think the lesson was obvious: permissioned intelligence has a political half-life, and some work never had permission in the first place. The office can live on a nerfed API. The line often cannot call the API at all.
This is Market A and Market B with a fence. Europe will keep buying filtered frontier tiers for headquarters. The cell still needs weights on a rack, a runtime that loads folders, and a harness you own. The Mac Studio in a closet was insurance for a software team. The plant version is architecture.
SCQA (at a glance)
Situation: OT, ITAR, and air-gapped sites cannot depend on a frontier API. Complication: Procurement still equates "approved cloud Copilot" with capability on the line. Question: What do you run on Monday in the cell? Answer: Portable skills, on-prem or local inference, open weights where they pass the eval, a classified route with logs, Market B with a safety case, not a hobbyist narrative.
The API is not the plant
Software teams felt a Friday blackout as a degraded chat. A plant feels a blocked API as: the skill cannot see the drawing, the twin cannot score, the FDE cannot leave. If your golden path assumes egress to a US lab, you do not have a golden path for OT. You have an office path you pretended was universal.
Dual-home it in writing. Which loops may use a frontier API. Which loops must run on-prem. Which loops may use open weights if evals pass. Waiting for the next letter, or the next auditor, is how you donate the decision to panic. I already made this inventory argument after the kill switch. The plant is where it stops being optional.
Skills travel when the SKU cannot
A skill is a directory. Directories walk through a fence. The runtime on the far side, open-weight loop, local Claude-compatible harness, air-gapped Cursor-class tool, whatever you can actually support, loads SKILL.md the same way. That is the hedge: workflow ownership on disk. Vendor-only skills that only exist in a cloud workspace are folklore at the gate.
Set compatibility like you mean it: no network, these packages, this site. Ban secrets in files. Ban consumer chat for drawings, not as a poster, as a path that is faster than the street. If the sanctioned local path is slower than ChatGPT, you will lose the drawing. Every time.
Open weights are not a vibe. They are an eval.
I am not telling you to run an untested 8B on a safety PLC. I am telling you to score. The same ladder that decides whether an agent touches a weld decides whether an open-weight route is allowed for pinout checks, summarising a work order, or drafting a tangent fixture. Frontier where you need it and can send it. Local where you cannot. Mix without evals is how OT security gets to shut the whole lab.
Cost, sovereignty, discontinuity insurance, three words architecture decks avoid because they imply the vendor is not permanent. On a site, they are just true.
The rack is part of the harness
Identity, logging, model route, kill switch, eval harness, plus a box that can load weights without leaving the fence. Platform's job in the triad includes this rack. If platform only offers the cloud SKU, every residency on an air-gapped site becomes a snowflake. Snowflakes in OT are incidents.
Europe's "sovereign wrapping" of nerfed US tiers may be fine for the office. Do not confuse it with a cell architecture. Wrapping a killable API in a local audit log does not keep the drawing on site.
Failure modes
Copilot for the air gap. Procurement theatre. Operators on personal accounts.
USB folklore. Weights and drawings on sticks because there is no rack. You did not avoid the cloud. You invented a worse cloud.
One model religion. Only frontier, or only local. The work does not care about your identity. Routes do.
Skills that assume network. A folder full of fetch is not an air-gap skill. It is a future incident.
Freeze. "No AI on OT." Shadow rises. You will meet it in the incident review.
This month
Classify the loops. Stand up one local route that can load three skills and fail an eval without egress. Write the compatibility line. Make that path faster than the consumer chat. Then let an FDE residency use it. If you cannot get here, stop talking about Physical AI. You cannot run it.
The plant cannot always call the API. Own the folders, the evals, and a route that runs behind the fence. Rent the horse where you are allowed. Keep the harness where the drawing lives. Anything else is a cloud subscription aimed at a door that does not open.
Related: Nerfed Frontier, Open Weights, After the Kill Switch, and Skills Became Infrastructure.