If you’ve been a developer in the last year, you’ve definitely heard of Cursor—and in this video I break down why they’re in trouble right now. Cursor announced “Composer 2” as their new in-house model, but the drama is that it looks like it’s basically Kimi K2.5 with some reinforcement learning/post-training on top. The bigger context here is how Cursor is built: a VS Code-based IDE layer plus AI model access. And the uncomfortable truth is most people aren’t really paying for an IDE—they’re paying for the intelligence layer (Opus/Claude, OpenAI, Gemini, Kimi, etc.), which makes Cursor simultaneously dependent on, and competing with, the model vendors. I also explain why Cursor is incentivized to build their own models to survive long-term—because pricing pressure (like Claude Max vs API economics) makes it hard to stay revenue-positive if you’re just reselling someone else’s tokens. The real issue is licensing: Kimi K2.5 is open source under a modified MIT license with a clause for hyperscale commercial users (100M MAUs or $20M monthly revenue) requiring prominent UI attribution. Cursor reportedly didn’t mention Kimi anywhere, and Moonshot employees claim the tokenizer matches and that Cursor didn’t come to them for a license. Net: building your own model is smart, but taking license risk at Cursor’s scale is wild.