token-based pricing is much easier to reason about than GPU-time quotas. This feels like the right direction for ollama cloud