Let's talk about Cursor
While the giants competed on intelligence, Cursor competed on integration. A look at how it got there and the bet others missed.
Every great tech story starts with a group of people who have no business winning, doing exactly that.
Cursor was four MIT students in 2021 trying to build AI for mechanical engineers. When that didn't work, they pivoted to a VS Code fork. No moat. No data advantage. No distribution. Just a better experience.
They crossed $100M ARR faster than anyone. 64% of the Fortune 500 runs it. And last week SpaceX acquired it for $60B in stock, four trading days after the largest IPO ever.
But here's the part nobody talks about.
For two years, the entire AI conversation has been Anthropic, OpenAI and a sprinkle of Google when they make the effort. Who has the smarter model. Who's winning benchmarks. Whose capabilities leap ahead this quarter.
Nobody stopped to ask: what are they actually racing toward?
The model war was not a race to the top. It was a race to the bottom.
Like any innovation wave, the R&D and first-mover costs are astronomical. Building a general purpose model on the world's data is ruinously expensive. That burden was carried by OpenAI, Anthropic, Google. And by Meta with LLaMA, which is what really made open source viable.
OpenAI alone has burned through tens of billions. Anthropic's latest round valued it at nearly a trillion dollars on zero viable revenue at that scale. Google can afford it but is still trying to figure out what to do with it.
Then the second wave arrived. DeepSeek and others trained on what the giants shipped and built competitive models at a fraction of the cost. Meta's LLaMA made that possible. The billions the giants spent became a subsidy for everyone else.
Sam Altman said it best: "GPT-4 is by far the dumbest model any of you will ever have to use again." The point holds. Every model we have today, including the ones powering this entire war, is the worst one we'll ever use. And open source is eating that gap fast. GLM-5.2 already beats GPT-5.5 on coding at 1/6th the cost. This trajectory doesn't slow down. In a few months the gap might be practically nothing.
Cursor watched this play out and positioned itself perfectly.
They didn't enter the model war early. They waited. When they finally built Composer, it was on Kimi's open source models. Cheap, fast, good enough. Composer 2.5 is a workhorse at a fraction of the cost of frontier models. Good enough to ship, cheap enough to run at scale.
While the giants burned billions competing on intelligence, Cursor competed on integration.
2023: Launches as an AI-first code editor. Not an extension. Not a plugin. A fork of VS Code rebuilt around AI. That choice matters.
2024: Ships Agent Mode. The industry calls it a novelty. Cursor keeps building.
2025: Cursor 2.0. In-app browser. Voice input. Their own Composer model. Background Agents that work while you sleep. MCP integration. JetBrains expansion. Jupyter. Half a billion in ARR.
Nov 2025 to May 2026: Composer 2 then 2.5. In-editor code review. Instant grep. MCP marketplace. Jira integration. Shareable canvases.Headless CLI. BugBot for autonomous PR review. Auto-review mode that lets agents run for hours without you.
They shipped them all in seven months.
May 2026: Cursor 3. A completely new interface built from scratch around agents, not editors. Multi-repo. Parallel agents. Local-to-cloud handoff.
Then Origin. Their own agent-first git platform.
And Microsoft? They built VS Code. They had every advantage.
They had a chance to buy or invest in Cursor early. That conversation must have crossed Satya's desk. But it was never actioned or taken seriously. Maybe they didn't see the threat. Maybe they thought bolting GitHub Copilot onto VS Code was good enough.
It was the wrong bet.