New analysis maps out who wins as AI infrastructure gets more expensive
4 minA quantitative scenario analysis argues that rising memory chip costs, increasingly capable open-weight models, and new entrants reselling older hardware will reshape the AI industry through 2030 — but that established players keep a durable cost advantage over newcomers thanks to depreciation on infrastructure they already own. The paper splits future spending into two tracks: expensive frontier training, projected at $18–38 billion a year by 2030, and much cheaper mass-market model refinement. Its bottom line is blunt: the cost gap between incumbents and new entrants doesn't close, no matter how hardware pricing shifts.