Turbopuffer CEO Simon Eskildsen: companies are building their own Google-scale search indexes on top of serverless vector DB

Summary

Simon Eskildsen, CEO of Turbopuffer, describes how AI applications are shifting from pushing entire datasets into model context windows to building dedicated search infrastructure that lets agents query data independently.

The pattern emerged first in coding. Tools like Cursor and Cognition use Turbopuffer to let agents search through codebases rather than feeding whole repositories into context. Eskildsen sees the same dynamic spreading across legal, enterprise, and other verticals. Teams typically start by pushing everything into context, then realize agents need to search the data themselves.

Turbopuffer's commercial pitch centers on Google-scale search indexes for private data. The company recently launched tooling to let customers index hundreds of billions of documents with sub-50-millisecond latency at what Eskildsen describes as extremely reasonable cost. He argues this is something only Turbopuffer can deliver off the shelf.

Compute constraints

Eskildsen acknowledges that compute scarcity is a broad problem. Every company at scale has to request hundreds of machines from cloud providers in specific regions. Turbopuffer runs on CPUs rather than contending for GPUs, but Eskildsen watches DRAM prices closely alongside everyone else.

Customer selection and survivorship

Turbopuffer's user base reflects survivorship bias. Companies come to them because they want to connect more data to AI than competitors do. If you're building a thin layer without deep data operationalization, the market becomes really difficult in an era of powerful models.

Trust and customer credibility

Eskildsen is direct about logo placement and customer trust. Early customers hesitated because Turbopuffer was "three or four people in Canada," a risky bet that kept some founders awake at night. Others worried what their own customers would think. The company has since grown 10x in scale, but Eskildsen still emphasizes honest representation. A small team at Vercel using Turbopuffer does not mean Vercel is puffing the product. He categorizes logos as "core" or "auxiliary" based on actual usage and asks permission before featuring them publicly. He views logo crimes—slapping major company names on landing pages to suggest enterprise adoption that doesn't exist—as a trust violation that founders eventually regret.

Replacement risk and infrastructure positioning

Shallow SaaS tools are becoming easier to replace with agentic composition. Turbopuffer positions itself as deep infrastructure that is hard to replicate and tied to data moat, rather than a commodity tool. Eskildsen expects some companies will build their own light versions, discover the maintenance burden, and return to paying for robust infrastructure.