OpenAI runs ChatGPT for 800 million users on a single PostgreSQL primary with 50 read replicas. Key scaling techniques include: vertical scaling, workload isolation across dedicated read replicas, lazy writes and batch writes to reduce primary load, fixing redundant writes, replacing expensive 12-table joins with multiple targeted selects, auditing ORM-generated queries, connection pooling via PgBouncer, cache locking to prevent thundering herd on cache misses, and hot standby for high availability. For write-heavy workloads, they are migrating shardable workloads to Azure Cosmos DB. Cascading replication (multi-tier read replicas) is being tested to scale reads beyond the current 50-replica limit. PostgreSQL's MVCC is highlighted as a write amplification concern at scale.