Elon Musk's New Grok V9 Was Trained on Your Private Cursor Coding Sessions

The newly finished Grok V9-Medium boasts a staggering 1.5 trillion parameters, making it three times larger than the previous version powering Grok 4.20. While earlier models were mostly famous for generating edgy jokes on X, this new behemoth is strictly targeted at dethroning OpenAI and Anthropic in writing actual, functional code.

To achieve this, Elon Musk's AI team didn't just scrape public GitHub repositories. They tapped into a goldmine of real-world developer behavior by injecting massive amounts of Cursor telemetry data into the training pipeline. This includes millions of live coding sessions, showing exactly how developers accept suggestions, reject bad AI completions, and manually patch bugs on the fly.

This data pipeline was unlocked through a massive, quiet deal struck in April. Under the agreement, SpaceXAI secured an option to buy Cursor outright for $60 billion by the end of the year, or pay $10 billion just to keep the partnership alive. In exchange, the Cursor team got access to the monstrous Colossus 2 supercomputer, packed with a million H100 GPU equivalents, to train their own Composer 2.5 model. While many thought Cursor got the better end of the deal with unlimited compute, the return flow of proprietary developer data back to xAI shows the transaction was far from one-sided.

The resulting model is currently undergoing final reinforcement learning, with a public release scheduled in two to three weeks.

Millions of developers essentially paid a subscription fee to act as free, high-quality labeling monkeys for a billionaire's side project. It turns out that the best way to build a world-class coding AI isn't hiring better engineers, but simply buying the IDE where everyone else does the hard work.

Source: X