Jensen Huang claims AI has achieved AGI, but new benchmarks tell a different story. The ARC-AGI-3 test places frontier AI models in interactive environments with no rules or predetermined goals. Humans solve 100% of these tasks. The best AI systems score just 0.37%. The gap reveals a fundamental limitation. Current AI excels at pattern-matching within training data but fails to adapt to novel situations. This inability to handle genuine novelty defines what AI can and cannot replace in actual work today.
The infrastructure layer now captures more investment than model development. This week alone saw $25 billion in deals focused on data systems and real-world applications rather than language models. IBM acquired Confluent for $11 billion to strengthen real-time data streaming. Eli Lilly bought Insilico's drug discovery pipelines for $2.75 billion. Physical Intelligence raised $1 billion for robot control systems. Building better large language models has become table stakes. The defensible value now sits with controlling the data flow between models and real-world applications. Companies that own this middle layer control how AI actually gets deployed and used.