Investment Memo: Cursor (Anysphere)
Stage considered: Hypothetical Series A Sector: Developer tools / AI-native software creation
---
Thesis
Cursor is a bet that the IDE — the surface where all software gets made — is being re-founded around AI, and that whoever owns that surface captures a disproportionate share of the value AI creates in software development. Developer tools are usually bad venture businesses (low willingness to pay, open-source substitution). Cursor broke the pattern: it is reportedly one of the fastest companies ever to scale from $1M to $100M+ ARR, on a largely self-serve, bottoms-up motion. If AI coding assistants become as universal as version control, the winning editor is a $10B+ revenue company, not a feature. That's fund-returner math even at an aggressive Series A price.
Product & wedge
Cursor is a fork of VS Code rebuilt around AI: inline multi-line autocomplete ("Tab"), codebase-aware chat, and agentic multi-file editing (Composer/Agent mode) that can plan and execute changes across a repository. The wedge is elegant:
- Zero switching cost in: VS Code fork means existing extensions, keybindings, and themes work day one. A developer can try Cursor in five minutes.
- Codebase context as the moat-in-progress: Cursor indexes the whole repo, making suggestions materially better than a generic chat window — and the more your team's context lives there, the stickier it gets.
- Model-agnostic: It routes across frontier models (Anthropic, OpenAI, plus its own fine-tuned models for autocomplete), so it rides model improvements rather than depending on one lab.
The wedge is the individual developer paying $20/month with their own card; expansion is team plans, then enterprise (SSO, privacy mode, admin controls).
Market & competition
TAM is credible: ~30M professional developers globally; if AI tooling becomes a $50–200/seat/month line item, this is a $20–70B annual market before counting the agentic expansion (AI doing work, priced on output, not seats).
Competition is the hardest part of this memo:
- GitHub Copilot (Microsoft): the incumbent, bundled into the GitHub/Azure enterprise motion, distribution Cursor cannot match. Copilot has been slower and more conservative on product, which is exactly the gap Cursor exploited.
- Windsurf (Codeium): the closest direct analog — also an AI-native editor, aggressive on price and agentic features.
- The frontier labs themselves: Anthropic's Claude Code and OpenAI's coding agents attack from the terminal/agent side; Cursor pays these labs enormous inference bills while they build competing products. This is the strategically ugliest dynamic in the deal.
- Others: JetBrains AI, Replit, Cognition (Devin), Zed, Sourcegraph.
Traction & business signal (public only)
- Reported to have reached $100M ARR roughly 12–21 months from launch, among the fastest SaaS scale-ups on record; subsequent reports place ARR at $300M–$500M+ during 2024–2025. Exact current figures: unofficial.
- Raised successive rounds at rapidly escalating valuations (reported $2.5B, then $9B+, with later reports near $10B+ from investors including Thrive, a16z, Benchmark, Accel). Precise terms: unknown.
- Widely adopted at named tech companies per press coverage; net revenue retention, gross margin, churn, and enterprise mix: unknown — and gross margin is the number I most want, given inference costs.
Risks (the three that actually kill it)
- No durable moat at the application layer. Cursor is a VS Code fork calling third-party models. The switching cost out is as low as the switching cost in — that's the flip side of the wedge. If Copilot closes the product gap or Claude Code makes the editor itself less relevant (agents working from the terminal/CI, not a GUI), Cursor's advantage compresses to "moves faster," which is not a moat at a $10B valuation.
- Gross margin structurally hostage to model providers. Cursor's COGS is frontier-model inference sold by companies that compete with it. If labs price API access to advantage their own coding products, Cursor is squeezed on both margin and capability. Its own models mitigate autocomplete costs but not the agentic workloads where the value is migrating.
- The paradigm shifts past the IDE. The bull case assumes humans stay in the loop inside an editor. If autonomous agents write most code within 3–5 years, the valuable layer becomes orchestration, verification, and deployment — and Cursor's asset (the editor surface) depreciates. Cursor must win the transition to being the agent control plane, a different product against different competitors.
Verdict
Invest at Series A — this is exactly the profile you take moat risk for: extraordinary organic pull, best-in-category product velocity, bottom-up motion in the largest knowledge-work vertical. You underwrite the team's speed as the moat until data/context lock-in matures.
The one question I need answered first: What is net revenue retention and inference-adjusted gross margin on cohorts older than 12 months? If developers stay, expand to teams, and unit economics survive real usage, the moat objection is empirically wrong. If churn is high and margins are thin, this is a beautifully wrapped pass-through to Anthropic — and I pass.