Vector database comparison: Pinecone vs Weaviate vs Qdrant

We're evaluating vector databases for our AI search product. Need to handle ~50M embeddings with real-time updates. Cost, latency, and filtering capabilities are key. What are you using and why?