Cheaper Retrieval Enables Bundled AI
Turbopuffer
This shows that retrieval cost can set the ceiling on product adoption just as much as model quality. When Notion made AI a standard part of its Business and Enterprise plans in May 2025, it stopped gating usage with a separate per member charge and started treating search across a company’s workspace and connected apps as a core workflow. That only works if looking up far more notes, docs, and external records stays cheap enough at scale, which is exactly the workload Turbopuffer is built for.
-
Turbopuffer makes money when customers store more data, write more vectors, and run more queries, so a customer product change like included AI can increase spend without any sales motion. The backend is metered on usage, and the front end removes friction, which creates automatic land and expand.
-
The retrieval tradeoff is concrete. Pinecone, Vespa, and Elasticsearch keep more data in memory or SSD for speed, while Turbopuffer pushes colder data to object storage and pulls hot data up when needed. That is why it is strongest for huge corpora, spiky traffic, and many lightly used namespaces, which looks a lot like workspace search.
-
Comparable productivity apps increasingly bundle AI into the suite instead of charging separately, because charging per seat can punish the vendor when AI reduces work. Notion moving AI into higher tier plans fits that shift, and it turns retrieval infrastructure from a cost center into a direct driver of feature breadth and gross profit.
The next step is more products pricing AI as part of the workspace, then buying retrieval that can support broader indexing without blowing up unit economics. As AI features move from optional assistant to always on work surface, the winning retrieval vendors will be the ones that let customers widen scope first, then monetize the resulting query growth.