Cursor's $20B Peak: Users Are Quietly Migrating, And You Should Know Where
A valuation peak and a user peak are not the same event. Investors bet on what already happened; engineers bet on what hasn't. When those two curves fork, that's the signature signal of a category topping out.
What holds up $20B
The Series D is a bet on three things:
- ARR going from $100M to $250M — Q1 delivered this one, it's real money
- Enterprise penetration into the Fortune 500, pulling seat price from $20 to $200/seat/month — still a pitch deck
- A vertical tooling moat — investors believe the forked VS Code + in-house Cmd-K + Cursor Composer stack into something defensible
Number three is where this valuation lives or dies. I covered it in the earlier MCP post: that moat already got punched through by protocolization in H1 2026.
Users are leaving quietly, in three directions
I put an NPS survey next to GitHub star growth for 3 open-source Cursor alternatives. The migration map:
| Destination | Volume | Type | Who's leaving |
|---|---|---|---|
| Back to VS Code + Cline / Continue | Biggest | Open-source MCP clients | Engineers who want AI Coding but resent paying for it |
| Claude Code (CLI) | Medium | Anthropic's official CLI | Heavy backend / DevOps / multi-machine work |
| Zed + Agent | Small but sharp | New IDE written in Rust | Performance obsessives / Apple silicon devotees |
Nobody is moving to Codex or VS Code's own Copilot. Worth noticing. Microsoft's OpenAI advantage never made it through to the IDE.
What all three waves share
Everyone leaving is solving the same pain: Cursor turned AI editing into a gym membership. You pay to use it, and the day you cancel the whole IDE feels wrong.
The next generation splits it into tool + protocol + your own model:
- IDE layer (VS Code / Zed / Helix): open source, stable, your pick
- Protocol layer (MCP): interop across tools
- Model layer (Claude / Codex / DeepSeek / local): swap at will
This decoupling has happened before. Browsers decoupled from search — there's DuckDuckGo and Kagi next to Google. Terminals decoupled from shells. AI Coding tools are only at the start of it.
How Cursor saves itself
My guess at the next 12 months:
- Pour money into enterprise — certain to happen, and also the most dangerous; enterprise sales raises seat price and shreds engineering culture on the way through
- Train their own models — they've already acqui-hired for it, but it's a deep hole
- Go vertical — own specific scenarios; Background Agents point this way
The smartest move available is to admit the moat is gone and start building the "MCP server store, optimized for Cursor." An App Store inside the IDE. But that requires saying out loud that Composer-class features get commoditized eventually, and they may not be willing to.
If you're deep in Cursor today
No need to migrate now. But shift your AI Coding workflow from "lives inside Cursor" to "runs inside Cursor."
Concretely:
- Write your common workflows as MCP servers / shell scripts so they're portable
- Don't sink time into Cursor-specific Composer templates
- Keep prompts and specs in markdown, not in
.cursorrules
Whether you actually leave is a question for six months out. What you do today is stop putting every egg in a basket whose valuation just peaked.