How Michael Truell built Cursor from failed ideas to $100M ARR

Executive overview

Most developer tools companies aim to make existing workflows slightly better. Cursor was built on a different premise: that all coding as we know it would change within five years, and that no one else was building for that reality.

The team spent most of 2022 failing — at CAD autocomplete, at encrypted messaging, at niche coding tools — before committing to a general-purpose AI code editor. They shipped a from-scratch editor in three months, switched to a VS Code fork when the cost of feature parity became clear, and grew from $1M to $100M ARR in a single year through relentless product improvement and word of mouth.

The core insight: in markets where end-user preference drives adoption, making the best product is the growth strategy.

From robotics to Cursor: the founding team's background

• Truell got into programming at age ~13 trying to build a mobile game; ended up learning Objective-C from a book.
• In high school, he and a friend built robots trained via reinforcement learning — implementing their own neural network library on microcontrollers because standard ML libraries wouldn't fit.
• All four co-founders had deep AI backgrounds before starting: one built a prototype Google competitor using LLMs in 2021, one worked on computer vision, others on recommendation systems.
• The team was inspired early by PG's essays and the Y Combinator ecosystem.

The failed ideas before Cursor (2022)

• First major idea: a Copilot for mechanical engineers — predicting actions in CAD tools like SolidWorks.
• Chosen because it seemed "boring and uncompetitive" — wrong on both counts, and none of the founders were mechanical engineers.
• Work involved scraping CAD models across many file formats, building training infrastructure, and hacking extensions into non-extensible software.
• Two co-founders simultaneously built an end-to-end encrypted messaging system that hid metadata (not just message content) — technically impressive but unscalable and impossible to sell.
• Both projects had essentially zero users after months of work.
• The team cycled through three to five ideas total before landing on coding.

The pivot to coding

• The team had originally avoided coding tools because GitHub Copilot already existed and seemed dominant (~$100M revenue by end of 2022).
• After months of low excitement and low traction elsewhere, they recognised they were genuinely excited about the future of coding — and that no one was building for a world where AI handled all of software development.
• Early niche options considered: a security review tool, a tool just for quantitative researchers. Abandoned in favour of a general-purpose AI editor.
• Conviction came from two sources: seeing early AI products emerge, and evidence that models would improve predictably with scale.

Building and shipping the first version

• First line of code to public launch: ~three months.
• Built their own editor from scratch using open source primitives (CodeMirror, language servers); added their own Remote SSH, Copilot integration, autocomplete.
• Usable as a daily driver internally after ~four weeks; first beta testers four weeks later; general availability four weeks after that.
• Key early learning: a single universal command that let the AI decide what to do (chat response, code suggestion, codebase search) didn't work well given 2022 model capabilities — the form factor needed more explicit control.
• Rapidly hit the limits of building a feature-complete editor: VS Code had 12 years of development behind it. Switched to forking VS Code to focus time on AI features instead.

Growth from zero to $100M ARR

• 2023: slow, wandering growth — roughly $0 to $1M ARR. Team stayed at under 10 people through year-end.
• Key product bets resisted: building for non-coders (loud early segment), building for a single tech stack. Both rejected in favour of horizontal focus.
• One co-founder built a niche AI Twitter following in 2022 by consistently posting paper analyses — helped seed early awareness and produced a "movie magic" launch demo.
• After launch, the team "lived like monks" in 2023, ignoring growth engineering in favour of product work. Growth engineering sprints never matched product improvement in impact.
• 2024: $1M to $100M ARR in a single year, driven by compounding product improvements:
  • Making Cursor codebase-aware
  • Next-edit prediction (tab completion)
  • Faster, more accurate predictions
  • Agentic sequences of changes
• YC batch adoption went from single-digit percent using Cursor in 2023 to ~80% in 2024.
• Growth mechanism: word of mouth among professional engineers in a market where end-user preference drives tool choice.

Model training as a product lever

• Early on, the team used off-the-shelf models and avoided training their own.
• The original Codex model (behind GitHub Copilot) cost ~$100K to train — used as proof of concept when raising their first round for the CAD idea.
• Over 2023, building their own models became important — especially for features like tab (next-edit prediction), where fine-tuned models significantly outperformed API models.
• At scale, product usage data creates a flywheel: more users → better training data → better models → better product.

The future of coding and advice for students

• Near-term view: professional engineers will work alongside AI as a colleague; code will remain important even as AI handles more of it.
• Long-term view: AI transformation of knowledge work will take decades, not years — many valuable companies to build along the way.
• Programming, like mathematics, remains a strong general education regardless of AI progress.
• Advice: work on things you're genuinely interested in, with people you respect. Resist the pull toward box-checking over building.