The real AI hardware story isn't the money—it's the constraint. Micron's $250B bet and the GlobalWafers-Micron deal matter because they're racing to solve an actual bottleneck: memory and substrate capacity can't keep pace with model scaling, and Taiwan's industrial infrastructure is about to become a geopolitical chokepoint. Meanwhile, llama.cpp's relentless optimization work (5 releases in one news cycle) reveals what actually moves hardware adoption—efficient inference on commodity chips trumps raw performance specs every time. The papers on diffusion stability and training efficiency are the unglamorous work that determines whether the next generation of chips needs to be overprovisioned or can actually deliver on their design targets; ignore the hype about "learning to reason through video" until someone ships it at scale with real constraints.