Napkin Math Exposes the Real Cost of AI Infrastructure

Simon Ericson’s "napkin math" approach—calculating theoretical hardware limits—exposed that many database benchmarks are misleading. At Shopify, he used it to challenge infrastructure decisions, leading to tools like Toxyroxy for failure testing. The lesson: engineering leaders should demand first-principles cost and performance estimates before adopting tools.

After leaving Shopify, Ericson applied napkin math to vector search, realizing S3’s durability could slash costs if latency could be managed. His prototype used simple clustering and Nginx caching, offering $1 per million vectors—undercutting incumbents by 100x. This forced rethinking where AI infrastructure spend goes: from expensive managed services to commodity storage.

Cursor, desperate to fix unit economics, became the first customer after Ericson personally debugged their Postgres issues. The 95% cost reduction proved that reliability and simplicity can win over feature-heavy competitors. The risk: betting on a one-person startup paid off, but only because of deep technical trust built through hands-on help.

Now, Turbopuffer faces new constraints: CPU and NVMe shortages as RL and agent workloads explode. Ericson’s CPU-centric architecture benefits from broad SKU compatibility, but the fight for compute is real. The strategic takeaway: infrastructure bets must account for shifting resource scarcity, not just current pricing.