Reverse Engineering ChatGPT Web: How OpenAI Built for a Billion Users
TLDR Dev
OpenAI's ChatGPT web app serves approximately 1 billion users through a carefully optimized architecture that migrated from Next.js to Remix and now uses React Router 7 with server-side rendering. The initial HTML payload is 84 KB compressed with a time to first byte of 50-65ms, delivering a 548-node DOM that prioritizes the chat input as the primary interaction point. The engineering approach uses off-the-shelf libraries rather than custom frameworks, with the server controlling client behavior through prefetch decisions and streaming data to enable rapid interaction without requiring user authentication.
Why it matters
ChatGPT's web app runs on a boring stack: React 19, React Router 7 (streaming SSR), Tailwind, Radix, and TanStack Query. Every default is questioned in service of letting an anonymous visitor start typing instantly with no auth wall and fast time to first paint.
