Simon Eskildsen on Turbopuffer: making vector search 10x cheaper and why AI is starving for context

Summary

Simon Eskildsen co-founded Turbopuffer after eight years running infrastructure at Shopify, where he was one of roughly six engineers on the company's last-resort on-call pager. That experience shaped everything about how he builds software: conservatism, simplicity, and a visceral awareness of how complex systems age. It also gave him a front-row seat to how badly search can fail at scale — the Shopify search system, he says, had the longest recovery times of anything he got woken up about.

Turbopuffer's pitch is cost. Vector search — the technique that lets AI systems find semantically relevant content by plotting text as coordinates in a high-dimensional space — was prohibitively expensive before Turbopuffer. The founding insight came from a concrete number: Eskildsen helped Readwise build a recommendation engine and found it would cost $30,000–$40,000 a month using conventional vector search infrastructure. He optimised the same workload down to $3,000 a month on Postgres — a 10x reduction. The gap convinced him that many companies were shipping constrained products simply because the economics of vector search didn't work. Turbopuffer's storage architecture brings costs down to what he describes as tens, even a hundred times cheaper than prior solutions.

The customer list reflects where AI context retrieval matters most: Cursor, Notion, Linear, and Superhuman all use Turbopuffer to index large datasets and connect them to AI. When a Cursor agent needs to pull relevant code from a large codebase, Turbopuffer is the search layer doing it cheaply enough that the unit economics hold.

Why search is hard

Eskildsen's explanation of why search remains so poor — even in products like Gmail or ChatGPT — is worth unpacking. String matching retrieves documents where a word appears, not documents that are actually about the topic you're thinking of. The semantic gap between what you type and what you mean has historically required PhD-level solutions for edge cases. Vector embeddings solve part of that by converting text into coordinates and finding what's geometrically close, but the data expansion is significant: a kilobyte of text becomes 20–30 kilobytes of vector data. That storage cost is precisely what Turbopuffer is attacking.

Email search has an additional structural problem: it's expected to be exhaustive and date-ordered simultaneously. The best semantic match might be a thread from ten years ago, which conflicts with what users actually expect to see first.

Shopify and agentic commerce

On agentic commerce, Eskildsen is sanguine about Shopify's position. Discovery moving into OpenAI or Meta surfaces is less threatening than it looks, because everything that happens after a customer clicks buy — returns, shipping, order tracking, the full back-office stack — still runs through Shopify. He draws an analogy to Meta's failed attempt at in-app Instagram purchases: the platform tried to own the transaction layer and backed off. Shopify, he argues, functions as the commerce CRM regardless of where discovery happens, and the take rate on partnerships may be better than outside observers assume.

His sharper point is qualitative: Shopify's merchant ecosystem produces real products with real entrepreneurs standing behind them, which is an advantage over the "LLM slop" equivalent of generic SKUs proliferating on large marketplaces.

The economics that drove Turbopuffer's founding are the same constraint facing any AI product that needs to search large, user-specific datasets cheaply enough to justify the feature. Readwise's $30,000-a-month problem was one instance. Cursor's power-user codebase is another. The bet is that there are many more.