The infrastructure bet is hitting reality while the software layer fractures into competing defaults. Meta quietly stripped language about web scraping and AI training from its terms without changing actual practice, the classic move of a company managing optics rather than operations. Meanwhile Replit demoted Anthropic in favor of OpenAI's models, a straightforward market signal that Claude no longer commands default status among builders even as Anthropic itself has shifted from startup velocity to corporate structure, more senior staff, more meetings, less shipping. The divergence matters because it exposes what was always true: the AI arms race isn't unified. Europe's industrial players are splitting between software-first (Schneider Electric, Siemens) and hardware-first (ABB) strategies. DeepSeek released V4.1-Flash-1017 with claimed improvements in speed and cost efficiency. NVIDIA packaged pretrained foundation models for tabular data into a zero-training interface. Kyutai released REKI, a 2.4B vision-language-action model for robots that matches larger competitors on out-of-distribution tasks. The physical world is where leverage actually lives, data center economics are tightening, genomics remains intractable to current models, materials science needs better 3D reasoning. Meanwhile the financial system is waking up to agent risk (Anthropic hiring lawyers for machine negotiation, Bank of England flagging stability concerns), and AI-generated political ads hit $137 million with 70 percent unlabeled. The pattern isn't consolidation around one winner. It's fragmentation into specialized layers where cost, speed, and task-specific performance matter more than brand.
Sloane Duvall