I think we can time share layers if we stop caring as much about chat-bot speed. I think there could be a point in the future where response times aren't the primary motive because the realized cost becomes too much.

Scaling horizontally in that case means distributing tasks within multi-agent flows between data centers.