a lot of people are still architecting around a compute dependency that shipped its way out of existence

qwen3 32b, for instance is live on distributed community gpus right now, pipeline parallelism running across nodes no single operator controls, permissionless inference over the open internet
three layers, all open: open weights (alibaba released them), permissionless compute (community hardware, no custody), open transport (no private cluster, no api permission gate). ownership doesn't apply to any of them
the mental model that frontier-capable inference requires hyperscaler rails was a default, and defaults expire when the infrastructure underneath them ships. builders running inference today don't need to ask anyone
the default expired. the stack is already public.