Elon Musk's AI startup company xAI was recently exposed for directly using the output data of Anthropic's flagship model Claude to train its Grok coding model over the past few months.

According to The Information, although Anthropic revoked xAI's official API access in January 2026, xAI engineers later turned "underground," continuing to extract data through personal accounts and third-party intermediate service Blackbox AI. Previously, Musk had admitted in court that xAI had "partially" used OpenAI's models for training and called this practice of using outputs from leading models for synthetic data transfer an "industry standard."

Grok, Elon Musk, xAI

However, external dependencies could not conceal internal turmoil within xAI. Currently, the pre-training team has been reduced to fewer than five people, and several co-founders, including four Grok code leads, have left in recent months. Recently, there was a major mistake where an employee accidentally deleted key training data, causing two to three weeks of work to be lost, forcing it to currently rely on temporarily renting computing resources from SpaceX to companies like Google.

This incident has placed xAI at the center of controversy and reflects the general anxiety within the generative AI industry under the backdrop of dwindling high-quality training data. At present, as compliant open-source data reaches its ceiling, using "knowledge distillation" and borderline synthetic data extraction to improve model performance has become a gray shortcut for latecomers to catch up with the first-tier players. xAI's current technical dependency and core team attrition indicate that the model of simply stacking computing power and relying on external data is facing a bottleneck. AI unicorns urgently need to find more stable support in fundamental innovation and organizational stability.