Cursor (Anysphere)

Cursor is an AI-native code editor, a fork of VS Code built by Anysphere (founded 2022 by four MIT graduates: Michael Truell, Aman Sanger, Sualeh Asif, Arvid Lunnemark) that became the fastest-scaling software business on record. Its stories share one through-line: read a platform shift correctly, rebuild the unit of disruption around it, and pivot without ego until the bet lands.

Why Now (Timing): the narrow window when models could finally write real code

The problem. In 2022, large language models had just crossed the threshold of writing usable code, but the dominant tool, GitHub Copilot, was structurally limited: it bolted AI onto VS Code as an autocomplete plugin, and VS Code is owned by Microsoft. A plugin could only ever be as ambitious as its host allowed, so the surface where AI actually mattered, the editor itself, was locked up by the incumbent's interests.

The approach. Anysphere bet that the editor, not the autocomplete, was the true unit of disruption, and committed a full product to that thesis before Microsoft could untangle Copilot from VS Code. They exploited a specific window: models good enough to write code, an incumbent architecturally constrained to a sidebar, and no AI-native editor yet built.

How it solved it. The timing proved exact. Cursor launched its IDE in 2023 and crossed $100M ARR by January 2025, roughly 21 months later, a pace only possible because it moved into the gap while it was still open. Each frontier model release (GPT-4, then Claude) upgraded the product directly, so Cursor rode the model curve instead of racing it.

Pivots: killing CAD, then killing the plugin

The problem. Anysphere's first product was not code at all. The founders built AI autocomplete for CAD and mechanical-engineering software, reasoning that a transformer trained on next-token prediction could speed up tedious work in tools like SOLIDWORKS. The market was stagnant, there was no structured training data for 3D geometry, and the founders were not their own users.

The approach. They abandoned CAD entirely and moved to code, a domain they understood deeply and where GitHub offered effectively unlimited public training data. They pivoted a second time as well, away from a plugin-style assistant toward a full AI-native editor, refusing to accept the ceiling a plugin imposed.

How it solved it. The two pivots landed them on the product that worked: an editor where text is the interface and the training data is infinite. That surviving bet went from a standing start to $1B ARR by November 2025 and roughly $2B annualized by early 2026, the fastest B2B software scaling on record, validating the willingness to kill two prior products.

Differentiation: root access the plugin could never have

The problem. Copilot, as a VS Code extension, could suggest the next line but could not redesign the editing surface. It structurally could not do multi-file edits, codebase-aware chat, or agentic workflows, because a plugin does not control the editor it lives inside.

The approach. By forking VS Code, Cursor gained root-level control of the editor and built features an extension cannot replicate: multi-file agentic edits, full-codebase reasoning, inline diffs, and a Shadow Workspace for verifying AI changes. Later it pushed the differentiation deeper by training its own model, Composer, released with Cursor 2.0 on October 29, 2025.

How it solved it. Composer is a mixture-of-experts model trained with reinforcement learning specifically for agentic multi-file editing, achieving frontier coding results at roughly four times the generation speed of comparable models. Owning both the editor and a purpose-built model gave Cursor capabilities a sidebar wrapper could not match, differentiation rooted in architecture rather than features.

Founder-Market Fit: the world's most demanding users of their own tool

The problem. In the CAD attempt, the four founders were building for hardware engineers whose work they did not do and whose data they did not have. They lacked domain expertise and, just as critically, they were not the customer, so they could not feel what was broken.

The approach. They moved to a domain where they were arguably the world's most demanding users: competitive programmers (with backgrounds including International Math Olympiad and Neo Scholar pedigrees) who lived inside code editors. They built the tool they themselves wanted to use, dogfooding every decision.

How it solved it. Being their own most exacting users let them design the AI editing experience by intuition and immediate feedback rather than guesswork. The fit showed in adoption: Cursor reached 1 million-plus paying customers and roughly 70% of the Fortune 1,000, an unusually fast pull for a tool built first to satisfy its own makers.

Moats: near-zero switching cost in, compounding advantage out

The problem. A new editor faces a brutal adoption barrier: developers have years of muscle memory, keybindings, and extensions invested in their current tool, and asking them to abandon all of it is usually fatal to a new entrant.

The approach. Forking VS Code (rather than building clean) inherited the entire VS Code extension ecosystem, keybindings, and workflow, so switching cost a developer almost nothing. On top of that low-friction entry, Cursor built an advantage that compounds: an architecture that improves automatically as underlying models improve, plus a proprietary model (Composer) that gets better through feedback loops tied to real user behavior.

How it solved it. The frictionless import pulled developers in, and the compounding architecture kept widening the lead: investors priced the moat at a Series A of $400M (August 2024) climbing to a $29.3B Series D (November 2025, with Nvidia and Google joining). A model trained on Cursor's own usage is far harder for a general-purpose competitor to replicate than a feature is to copy.