Simon Eskildsen of Turbopuffer discusses building a next-generation search engine for AI workloads, leveraging a novel S3-first, NVMe-optimized architecture that eliminates traditional database complexities. He shares how prohibitive costs for vector search at Readwise inspired Turbopuffer's creation, leading to significant cost reductions for customers like Cursor and Notion, and adapting to the evolving demands of agentic AI.
Summarized by Podsumo
Turbopuffer's Innovative Architecture: Built on NVMe SSDs and object storage (S3/GCS), it uses S3's strong consistency and compare-and-swap for metadata, bypassing complex consensus layers and enabling a fundamentally cheaper and simpler database design.
Origin Story & Cost Reduction: The idea for Turbopuffer stemmed from the high cost of vector search (30k/month) for a single feature at Readwise, driving the mission to create a highly cost-efficient solution, exemplified by a 95% cost reduction for Cursor.
Adapting to AI Workloads: The platform evolved from supporting single RAG queries to handling highly concurrent, agentic AI searches, leading to a 5x reduction in query pricing to accommodate the increased demand for parallel data access.
"P99 Engineer" Culture: Turbopuffer focuses on hiring "P99 engineers" who are obsessive, demonstrate high agency in bending software to their will, and excel at articulating trade-offs, fostering a talent-dense team.
Strategic Fundraising & Focus: Simon's candid approach with investor Lockhe, offering to return funds if product-market fit wasn't found, highlights a strong commitment to focused execution, prioritizing search capabilities before expanding to other database functionalities.
"I don't think you're going to find a company over the next few years that doesn't directly or indirectly have all their data available for search and connected to AI."
"I don't think I've said this publicly before, but I just call Lockhe and was like, local Lockhe, like, if this doesn't have PMF by the end of the year, like, we'll just like return all the money to you, but it's like, I don't really, just need that I don't want to work on this unless it's really working."
"I think AI has also changed the Bivers' build equation in terms of, it's not really about can we build it? It's about do we have time to build it?"