Translate user count into active concurrency, request rates, data volume, and geographic distribution. Keep request-serving instances stateless where practical, use caching and CDNs for suitable reads, partition data only when measured limits require it, and move slow work to durable queues. Add autoscaling with headroom, rate limits, observability, load testing, and cost controls. Millions of registered users alone do not determine architecture; usage patterns and SLOs do.