xAI nears Grok 5 training breakthrough

xAI's shift to parallel multi-model training is a high-stakes compute play aimed at shattering the "diminishing returns" ceiling of large language models.

• –Training seven variants simultaneously allows for rapid architecture and scaling law testing that competitors cannot match.
• –Colossus 2's 1-gigawatt power draw underscores the extreme infrastructure moats required for the next leap in intelligence.
• –Deep integration of Tesla's real-world video data aims to ground Grok 5 in physical reasoning rather than just token prediction.
• –Recursive self-improvement loops during training suggest the model is actively helping optimize its own underlying code.
• –The SpaceXAI merger creates a unique vertical moat, potentially leveraging Starlink for distributed training or data ingestion.