turbopuffer, a company that builds vector databases, published a blog post this week titled "RIP, vector database." We believe this deserves serious attention: when even the shovel-sellers say the shovel is being retired, there's usually real signal behind the headline.
Vector databases — systems purpose-built to store and retrieve "semantic similarity" — have been the default pairing with RAG (retrieval-augmented generation, the pattern of letting LLMs query external material to fill in knowledge) for the past two years. turbopuffer's core argument: as the industry moves toward Agent (an autonomous AI that decides its own next move) + hybrid retrieval, standalone vector databases are being absorbed into general-purpose object storage and are no longer a distinct category.
What this is
The standard pattern for letting enterprise AI query internal documents: chunk the material → convert to vectors (mathematical representations) → store in a vector database → at query time, retrieve the most similar content and feed it to the LLM.
This architecture spawned a wave of companies — Pinecone, Weaviate, Chroma, Milvus. But starting in 2024, three things happened at once: LLM context windows jumped from 4K to 200K, even into the millions; Agents can decide on their own what to query and which tools to use; keyword search and vector search merged into "hybrid retrieval." The result: the standalone value of vector databases is being diluted — general-purpose storage like AWS S3, with vector capability bolted on, is now enough.
Industry view
Plenty of voices back the "dead" verdict. turbopuffer's argument: the future is "storage with vector capability," not "a database that stores vectors." The thesis is already validated among the hyperscalers — AWS, Azure, and Google Cloud are all pushing vector capabilities into general-purpose databases and storage services.
But the pushback is equally clear. Incumbents like Pinecone and Qdrant argue that the engineering complexity of vector retrieval — performance, scale, latency — is nowhere near ready to be replaced by general-purpose storage. One AI infrastructure founder told us privately: "Saying vector databases are dead is like saying databases are dead — the abstraction layer changes name, but someone still has to do the work underneath." This view deserves weight: the death of a category and the disappearance of a capability are often not the same thing.
Impact on regular people
For enterprise IT: From 2025 onward, when procuring AI infrastructure, you no longer need to evaluate vector databases as a separate line — look first at the integrated offerings from cloud providers. IT budget structures will shift.
For individual careers: "Knows vector databases" was once hard currency for non-technical talent breaking into AI. That scarcity premium is fading; what replaces it is "knows Agent workflow orchestration" and "knows business data governance."
For consumer markets: No visible change in the short term. It typically takes 12-18 months for an infrastructure-layer reshuffle to show up in consumer AI products (customer support, search, recommendations).