From $0 to $40M ARR: Inside the tech that powers Bolt.new (newsletter.posthog.com)

Bolt — the viral one‑prompt app builder born from a seven‑year StackBlitz codebase pivot — has rocketed from $0 to $40M ARR since relaunching in October 2024, generating millions of apps with a 15‑person engineering team. The product feels magical: a single LLM call turns a natural‑language prompt into a full app, and the preview runs instantly because the “server” is your browser. That business and UX win shows a new, cost‑efficient path for AI app builders: client‑side runtimes cut cold starts, per‑user cloud costs and some abuse vectors, while delivering localhost‑like latency and better perceived privacy. The technical secret is WebContainer: a Node‑compatible runtime squeezed into a browser tab via a Rust‑based virtual file system compiled to WASM and mounted in a SharedArrayBuffer, plus Web Workers that act as OS‑like processes with Atomics for IPC, fake signals/stdio, and a TypeScript shell (JSH). Networking is solved with a Service Worker that virtualizes localhost and bridges WebSockets, plus a relay for raw TCP. A Node‑style module resolver and an ESM↔CommonJS bridge let the npm ecosystem run unchanged. Boot performance uses a slim WASM binary, snapshot‑first file system blobs and CDN package layers so installs and cold starts happen in <500 ms. For the AI/ML community this demonstrates how LLM-generated apps can be delivered at scale by coupling model output with sophisticated client runtimes rather than heavy cloud orchestration.