In production.
115 Trading runs a futures trading desk against Compute on a DGX Spark it owns. Lumbridge does not host, meter or operate any of it — 115 installed the same Apache-2.0 binary anyone can install.
Compute's resident agent serves a protocol-transparent gateway on 8011 and keeps the model runtime on 8001. The desk's configuration holds one URL and one model id from a Scene — never a weight path, a quantisation or a version. Activating a different Scene changes what is serving without changing what a caller asks for.
Because a caller that does not name a version cannot pin one either, the desk reads the identity back from the gateway and maps the weights behind the alias onto the published Scene name, so an operator sees which Scene the node is in without opening a shell on it.
Limits: one node, one consumer. Compute manages one node at a time and has no fleet API, Scene changes here are run by hand, and there is no failover.