Optimizing IP Geocoding for Sub-Millisecond Latency?

We operate a popular web tool, 'What is my City Name', which primarily relies on IP-based geolocation to provide city-level data to users.

Our main technical hurdle now is achieving sub-millisecond latency for our core IP geocoding lookups, especially as traffic scales significantly. We're running into limitations with existing approaches:

• Third-party APIs: Introduce unacceptable network overhead and dependency, pushing response times beyond our target.
• Self-hosted databases (e.g., MaxMind GeoLite2 City): While accurate, optimizing these for high-concurrency, ultra-low latency lookups (especially across different regions for global users) without disproportionate server costs is proving difficult. We've experimented with various indexing strategies and in-memory caches, but hitting a wall on consistency and speed.

We're looking for advanced architectural patterns, specific database configurations, or high-performance caching layers that can reliably deliver this level of IP-to-location mapping performance. Has anyone implemented a custom solution for extreme low-latency IP-to-location mapping at scale? Thanks in advance!