What a shift in last 6m … what’s the broader impact of these trends? Need to think on it more but 5 random thoughts with kids breakfast … - DC : data center was the $$ source of gravity for pre training … a) one shot training not a moat anymore b) spend shifting to RL environments and CPU/memory .. this should stay steady here - POWER / electrification: continues to stay scarce either way (love this factor on current drawdown) - distillation is table stakes: pretraining data gets less differentiating once base models are good enough to distill - where is puck going? environments, expert traces, tool logs, unique data and reward signals. That lifts vendors and internal orgs that manufacture RL tasks - APP layer: benefits twice… models get cheaper as the lagging tier commoditizes AND the frontier keeps moving to long running agentic tasks…so the product that owns the workflow and the feedback data matters more than the product that wraps a base model