Building Turbopuffer: Gergely Orosz (@pragmaticengineer ) × Simon Eskildsen (CEO)
Simon Eskildsen, CEO of Turbopuffer, discusses his unique path from early computer fascinations to founding an S3-native vector search company. He shares insights from his eight years at Shopify, detailing infrastructure scaling challenges and the development of Toxyproxy for fault injection. A core theme is his "napkin math" philosophy, using fundamental performance metrics to optimize systems and challenge benchmarks. He explains the impetus for Turbopuffer, its initial minimalist architecture, and how it achieved a 95% cost reduction for its first customer, Cursor. Simon also touches on the unexpected scarcity of CPUs due to AI workloads and his unconventional, principled approach to venture capital and building a remote-first company culture.
Simon Eskildsen’s journey began with early computer fascinations, evolving from creating intricate PowerPoint games to exploring Front Page, HTML, and PHP. Confronted by language barriers in programming resources, he sharpened his English skills through World of Warcraft. A pivotal experience was participating in the International Olympiad in Informatics (IOI), which introduced him to algorithmic problem-solving, a stark contrast to his web development background.
Shopify Years and Infrastructure Scaling
Simon joined Shopify in 2013, bypassing university after his article about ditching his iPhone for a Nokia went viral and caught a recruiter’s attention. Despite initial insecurities about lacking a formal computer science education, he adopted a rigorous self-learning approach, delving deep into every technical concept he encountered, from TCP/TLS to database internals. This curiosity led him to the infrastructure team, where he witnessed Shopify’s rapid growth and the challenges of scaling. Key initiatives included:
- Black Friday Preparedness: Annually preparing for escalating traffic, managing physical hardware orders, and scaling software.
- Database Sharding: Addressing the limitations of single shards, with a critical cutover happening just a week before Black Friday.
- Failure Modes: Tackling the resilience of the system, particularly concerning stateful components like session stores. This involved building a matrix to define service behavior when components failed.
- Toxyproxy: Simon created this Layer 4/7 proxy to simulate database failures (e.g., taking databases offline, introducing slowness). This tool, by injecting faults at the connection layer, uncovered numerous issues in Rails and MySQL drivers that were difficult to identify in production environments.
The “Napkin Math” Philosophy
After eight years at Shopify, Simon sought new challenges. He formalized his “napkin math” project, a GitHub repository detailing hardware performance metrics (DRAM bandwidth, S3 latency, NVMe SSD speeds) and their costs. This project was born from his frustration with poor benchmarks driving infrastructure decisions. He used these fundamental numbers to challenge assumptions, for instance, by calculating that a search query should take milliseconds rather than seconds, or investigating why MySQL could achieve 10,000 writes per second despite f-syncs taking 1ms (due to batching).
Founding Turbopuffer: S3-Native Vector Search
Three key experiences converged to inspire Turbopuffer: a frustrating search project at Shopify, his “napkin math” insights into system limits, and observing the high costs of existing vector search solutions during angel engineering projects. The advent of LLMs with small context windows highlighted the critical need for efficient search. A project with Readwise, where a recommendation engine incurred a projected $30,000/month cost (compared to their $5,000/month total infrastructure spend), solidified the problem. Simon conceived of an S3-native vector store to drastically reduce costs. The initial summer of development was spent overcoming S3’s inherent latency (P99 of 200ms for small objects). His initial, pragmatic architecture involved clustering vectors, storing them as files on S3, and using an Nginx reverse proxy for caching to achieve the necessary speed. He launched Turbopuffer on Twitter, advertising “a million vectors for a dollar.”
Partnering with Cursor and CPU Scarcity
Cursor became Turbopuffer’s first customer. Their founders, already exploring S3 for KV caching, found Turbopuffer’s approach perfectly aligned with their need to manage expensive in-memory vectors. Simon personally flew to San Francisco, building trust by helping Cursor debug their PostgreSQL issues (e.g., autovacuum misconfiguration). This relationship, combined with his co-founder Justine’s arrival, enabled Turbopuffer to reduce Cursor’s vector database bill by 95%. This success underscored the conviction both teams held, despite Turbopuffer being a nascent startup. Simon also shared a humorous anecdote about meeting Jensen Huang, where he lightheartedly asserted Turbopuffer’s reliance on CPUs and SIMD, despite the Nvidia CEO’s GPU focus. He notes a growing CPU scarcity, driven by AI workloads like reinforcement learning and agents, which demand significant general-purpose compute. This scarcity forces cloud providers to prioritize regions based on power availability, impacting even large customers and underscoring the importance of architectural flexibility across diverse machine types.
A Unique Approach to Venture Capital and Remote Culture
Simon’s approach to venture capital is grounded in first principles. Initially skeptical about Turbopuffer’s “venture scale” potential, he focused on building a profitable business where customer billing surpassed their GCP costs. His initial seed round of $700,000 was solely for R&D, explicitly stating they would return the money if the product didn’t achieve significant traction. He outlines six reasons to raise capital:
- Fund R&D: Turbopuffer’s first raise.
- Fund growth.
- Founder’s ego: A dangerous, but common, reason.
- Reward employees: Turbopuffer’s second raise provided employee liquidity.
- Strategic partnership.
- M&A.
Turbopuffer operates as a fully remote company, a natural extension of Simon’s experience at Shopify. To foster connection, they hold bi-annual, in-person company gatherings. They’ve also introduced “campfires” (spontaneous local meetups) and “Turbo Credits,” which reward employees for extracurricular contributions (e.g., conference talks, blog posts) with business class flight upgrades, encouraging more in-person interaction and shared experiences.