RFD 1055: BEAM workers, local first
Problem
Replicate runs each model as a serverless Cog. That model costs more than it should, and it fights the stack in three ways.
Cog is Python, and this is an Elixir shop. Replicate runs a predict.py, thus no Elixir queue, no supervision tree, and no Nx work can live inside the worker.
The price is a markup, and the retention terms are Replicate’s, not ours. Replicate bills per second above the cost of the card, volume and long jobs make that worse, and a job’s data lives as long as Replicate’s own terms say, not RFD 1058’s zero-trust design.
The boundary leaked into the repository. It produced a duplicate application in cms/, and a passthrough whose model map was empty, thus every job answered 400.
Decision
Blocklist Replicate. Run the worker on this box’s own 4090 first, in plain Docker images, on RFD 1058’s existing Quadlets. vast.ai stays priced and ready, the next tier once this box stops being enough, per RFD 1062’s Gall’s law, not the immediate plan.
The BEAM owns the queue, the retries, and the state, no Python involved, whichever host runs it.
See DETAILS.md for the providers priced for later, the host tiers, the two phases, and what stays unresolved.