Source: Dwarkesh Patel
Dwarkesh Patel argues that inference costs—not training—will become the binding constraint as AI models proliferate and users demand real-time responses, potentially making compute 10x more expensive as demand outpaces efficiency gains. This contradicts conventional wisdom about Moore's Law solving AI economics. The problem isn't building bigger models but serving them at scale, which creates immediate tension between AI adoption timelines and infrastructure spending. If correct, this favors companies with captive compute (like hyperscalers running their own services) over those licensing models, and could slow deployment of generalist AI across industries.