SpaceX Powers Anthropic, DeepMind Trains in EVE Online, Gemma 4 MTP, Grok 4.3 API

SpaceX has allocated 300 megawatts of power capacity to Anthropic at the Colossus 1 facility, marking a significant infrastructure partnership for Claude model training. This substantial power allocation underscores the growing energy demands of large-scale AI training operations and represents a notable collaboration between Elon Musk's space company and the AI safety-focused organisation. Google DeepMind has begun utilising EVE Online as a testing ground for AI systems focused on memory and long-term planning capabilities. The massively multiplayer online game provides a complex, persistent environment where AI agents can develop and test strategic thinking over extended timeframes, offering researchers valuable insights into AI behaviour in dynamic, player-driven scenarios. Google has released Gemma 4 with Multi-Token Prediction (MTP) drafters, delivering a threefold increase in inference throughput for the open-source models. The new Gemma 4 variants include both a 26B MoE (Mixture of Experts) version and a 31B Dense model, significantly improving performance whilst maintaining the accessibility of open-source deployment. Meanwhile, xAI's Grok 4.3 has achieved top performance on agentic AI benchmarks, demonstrating superior capabilities in autonomous task execution. Alongside this achievement, xAI has reduced API pricing by 40%, making advanced AI capabilities more accessible to developers whilst Grok 4.3 outperforms its predecessor Grok 4.2 across multiple evaluation metrics.