Simon Hørup Eskildsen

Simon Hørup Eskildsen co-founded turbopuffer in 2023, building search infrastructure that connects large datasets to AI through vector and full-text search. His consequential design choice was to make object storage the durable source of truth, with SSD and memory caches for actively searched data. His March 2026 technical article identifies him as co-founder and CEO and credits co-founder and CTO Justine with the operational experience behind their preference for fewer stateful dependencies.[2][3][1]

In late 2022, Eskildsen helped Readwise prepare its infrastructure for the launch of Readwise Reader. He built an article recommendation prototype using embeddings of recently read articles. In his account, the recommendations worked, but the infrastructure cost prevented the feature from shipping. His technical article compares roughly $5,000 a month for Readwise’s relational database with an estimated $20,000-plus a month for vector search over the same 100 million-plus documents. He began developing turbopuffer in 2023.[3][1]

Eskildsen designed turbopuffer around the different requirements of search and transactional databases: search can tolerate more relaxed write latency while still needing fast reads. Writes commit durably to object storage; SSD and memory caches reduce latency for active datasets. This lets infrequently searched data remain cheap to store, at the expense of slower occasional cold queries. The design particularly suits workloads in which only some tenants’ data is active at a time.[3]

The query planner and storage engine coordinate to fetch more data per object-storage round trip and limit sequential requests. For cold vector searches, Eskildsen describes a target of at most three round trips. In his Latent Space interview, he illustrated an early approach: fetch a file describing vector clusters, identify the nearest clusters, then download their files together. He explicitly qualified claims of originality: “I don’t know if we were the first”.[3][1]

Cursor became turbopuffer’s first large customer. Eskildsen recalls launching an initially simple Rust service on a single eight-core machine, hearing from Cursor the next day, and visiting its team in San Francisco. During that engagement, he asked Justine, a former Shopify colleague, to become his co-founder, and they worked closely with Cursor on migration and reliability. His article dates the migration to November 2023 and reports a 95% cost reduction. Cursor’s many codebase indexes, only some active at any moment, fit the caching architecture.[1][3]

The Notion engagement tested their willingness to preserve that architecture. Eskildsen recounts starting on Google Cloud, where conditional object updates let servers modify shared metadata only if it had not changed in the meantime. Rather than introduce a separate stateful metadata system to serve Notion on AWS, his team bought dark fiber, tuned TCP windows and prewarmed connections to reduce cross-cloud latency, while absorbing networking costs. He connects that decision to his and Justine’s years on call at Shopify: outages involving inconsistent state in multiple systems were precisely what they wanted to avoid.[1][3]

Eskildsen argues that models can learn to reason about knowledge but still need external systems that preserve it in full fidelity. In the March 12, 2026 interview, he described retrieval shifting from one initial context-building call toward a single agent issuing many concurrent searches. Different search methods can help diversify those requests. He said turbopuffer was reducing query pricing fivefold to accommodate the higher volume. He also argued that AI changes the build-versus-buy decision toward whether a company has time to build infrastructure itself.[1]

His longer-term ambition is for the database to support more query plans as customers’ needs develop, but his stated priority in March 2026 remained search: fuller text-search features and cheaper, faster searches over very large datasets. He warned against expanding too quickly, telling his team that “what we’re most likely to regret at the end of the year is having tried to do too much.”[1]

Eskildsen describes accepting investment as a deliberate commitment to give the company his full effort. He chose Lachy Groom because he felt able to speak openly without preparing a pitch. In his retrospective account, he offered to return the money if turbopuffer failed to reach product-market fit. His phrase for this approach was “I just play with open cards”. He valued Groom’s help with candidates and customers alongside the database expertise already present in the team.[1]

His “P 99 Engineer” framework looks for engineers who understand tradeoffs and have made software approach the performance suggested by basic hardware calculations. Hiring defaults to rejection unless someone strongly champions the candidate; unanimous enthusiasm is unnecessary. As an example, he credited chief architect Nathan, with substantial help from the team, for a six-to-eight-week effort enabling a particular large-scale vector-search workload.[1]

Before turbopuffer, Eskildsen’s career centered on scaling infrastructure. His earlier Shopify work supplies the operational context for the database decisions he later described with Justine.[2][3]

Eskildsen grew up in Aarhus, Denmark. He worked with Rails at the phone-support software startup Relatel during high school, from 2010 to 2013, and competed for Denmark at the International Olympiad in Informatics in 2012 and 2013. In 2013, he moved to Ottawa to join Shopify.[2]

His biography dates his Shopify career to 2013–2021 and lists Principal Engineer/Director responsibilities across infrastructure as the team scaled from 1,000 to one million requests per second. He recalls work on database scaling, sharding and the move from company data centers to the cloud. Self-hosted Elasticsearch was especially difficult to operate, although he qualified that criticism as experience with its roughly 2015 vintage. In 2019, he began maintaining sirupsen/napkin-math and writing about estimating systems performance.[2][1]

From 2021 to 2023, he took roughly three-month infrastructure engagements at friends’ companies, including Replicate, Causal, Datafold and Readwise. He called the arrangement “angel engineering”: solving infrastructure problems while vesting some equity. At Readwise, his initial work included tuning PostgreSQL autovacuum before the recommendation experiment exposed the retrieval-cost problem that became turbopuffer.[2][1]