Turbopuffer: Why vector search doesn't need GPUs or DRAM

Simon Eskildsen, founder of Turbopuffer, explains how he built a vector database that runs on CPUs and S3 rather than DRAM or GPUs. The key insight: napkin math on theoretical hardware limits drove the architecture, achieving 1 million vectors for $1. Cursor was the first customer, reducing their bill by 95% after migrating from Aurora.

CPU scarcity is emerging as a real constraint, driven by RL workloads and agent-based inference. Simon notes that the cloud providers are not infinite, and even medium-sized companies struggle to secure CPU allocations. This forces teams to adapt architectures to run on diverse instance types.

Simon's venture capital philosophy is pragmatic: raise only for R&D or employee liquidity, not ego. He raised $700K early to hire two engineers, then became profitable. The company remains remote with intentional in-person gatherings called campfires, using turbo credits to incentivize travel.

The engineering takeaway: simplicity and cost efficiency can outperform complex, DRAM-heavy solutions. Teams should evaluate whether their AI infrastructure actually needs GPUs or if CPUs with object storage can meet latency and cost requirements.