What stays on each side of the boundary?
By Anthropic's announcement, repository checkouts, build artifacts and files a session creates stay on machines the organisation provisions, while prompts, responses, tool results and session transcripts go to Anthropic. The source code stays put; the code Claude reads still leaves. The option is open on Team and Enterprise plans, off by default, and cannot be enabled at all by organisations on the zero-data-retention setting.[1]
The substance of the announcement is a transfer of operating burden. Anthropic itself writes that organisations must staff engineers to build runner images, update the infrastructure and operate an orchestrator for on-demand scaling. Those three items are a cost that never shows up on an invoice and that somebody still pays.[1]
In which geography does inference run?
The August 7 platform notes tie where an agent's model inference runs to the `inference_geo` field inside the model object; the field can be overridden for a single session, and the notes point to pricing alongside the available geographies. The session process and the model inference are therefore placed by two separate controls.[2]
What is missing is a figure. The self-hosting announcement gives no hardware sizing for a runner; the platform note publishes neither a session's cost nor a per-geography price, only a pointer to pricing. What exists today is a deployment mode and a placement switch; measured capacity does not exist yet. Whether adoption is driven by cost or by data-control requirements is also unsettled, because the announcement's own emphasis falls on security and operational controls.[1], [2]