What Cursor actually runs

The most telling fact about Cursor’s stack isn’t a tool — it’s a retreat. Anysphere adopted Yugabyte, a distributed database that markets itself as infinitely scalable, and then migrated off it to plain PostgreSQL (confirmed). The company behind nearly a billion lines of code written daily decided the boring database was the scalable one. That instinct — boring where possible, exotic only where the workload demands it — runs through the whole stack.

Most of what follows comes from cofounder Sualeh Asif’s June 2025 interview with The Pragmatic Engineer, which makes this one of the better-evidenced stacks we’ll ever document: nearly everything below is confirmed by the company. Full sourced record in the stack database.

Stack at a glance

Context for scale, per the company: 1M+ transactions per second at the data layer, ~50 engineers, 100x load growth in 12 months, $500M+ ARR at the June 2025 snapshot, and a $900M Series C at a $9.9B valuation. Reports in 2026 put the valuation far higher, but we haven’t seen primary confirmation, so we’re labeling that one reported.

The fork that bought them a company

Cursor is a fork of VS Code — TypeScript and Electron, same as the upstream editor. Asif is direct about why: “We decided to fork because starting from scratch would have taken a massive effort just to build a stable editor. Our value proposition was not building a stable editor… Forking let us focus on the experience, not the editor.” The much-copied tab model, he says, “would have been very difficult without forking.”

The backend follows the same philosophy: mostly one large TypeScript monolith, deployed as one unit, with Rust reserved for performance-critical components and a Node.js bridge connecting the two (confirmed). At 25,000 files and 7 million lines of code, the monolith survived hypergrowth that conventional wisdom says demands microservices.

Two databases, two jobs, one emergency

The data layer splits cleanly. PostgreSQL handles transactional work — after the Yugabyte retreat. Turbopuffer holds the interesting stuff: encrypted code chunks and the Merkle trees Cursor uses to sync codebase indexes without ever storing source code server-side (confirmed). Embeddings — built with OpenAI’s models or Cursor’s own — get vector-searched to answer codebase questions, while the actual code is fetched from the client on demand. A Pinecone instance stores documentation embeddings; Warpstream moves streaming data.

The team’s stated reason for preferring Turbopuffer: not having to deal with manual sharding (confirmed). And it earned its place dramatically — during a large indexing outage, the team executed a migration to Turbopuffer in hours, mid-incident. Migrations as incident response is a genre we expect to see more of.

The GPU estate

Asif: “We are very much a ‘cloud shop.’ We mostly rely on AWS and then Azure for inference. We also use several other, newer GPU clouds as well.” CPU infrastructure lives on AWS; tens of thousands of NVIDIA H100s — much of that footprint in Azure — do inference and only inference (confirmed). Training and fine-tuning run elsewhere: Voltage Park, Databricks MosaicML, and Foundry (confirmed). Terraform manages it all, and Cloudflare fronts many services as reverse proxy and DDoS shield (reported).

For agents, a fully-Rust orchestrator called Anyrun — a pun on Anysphere — launches cloud agents on EC2 with AWS Firecracker providing process isolation (confirmed). And on the model-serving side, Cursor’s fine-tuned “Fast Apply” Llama-3-70b runs on Fireworks with speculative decoding at ~1,000 tokens/second — a 13x speedup over the vanilla model (confirmed, vendor co-published).

The unglamorous perimeter

The business stack is deliberately conventional, all confirmed: Datadog for observability (Asif calls its developer experience “vastly superior to the alternatives”), Sentry for errors, PagerDuty wired into Slack for oncall, Amplitude for analytics, Stripe for billing, WorkOS for auth, Linear for work tracking, Vercel hosting cursor.com. And Cursor builds Cursor — with the house rule that every engineer owns their checked-in code “whether they wrote it by hand, or had Cursor generate it.”

What we’d watch

Whether PostgreSQL survives the next 10x (the sharding complexity they escaped once will knock again), how much inference migrates from Azure H100s to newer GPU clouds as that market commoditizes, and whether the TypeScript monolith makes it to 100 engineers. Our snapshot is June 2025; Cursor moves fast, and we’d welcome a verification from the team.

Sources: The Pragmatic Engineer’s deep-dive with cofounder Sualeh Asif (primary), Fireworks, ByteByteGo, Contrary Research. Research assisted by AI agents, verified and edited by a human. Confidence labels: confirmed / inferred / reported. Spot an error? We correct publicly.