Billions are pouring into LLMs-but will returns keep pace with the burn? This post cuts through the hype with a numbers-first look at LLM unit economics, pinpointing where value truly accrues across chips, cloud, models, data, and the application layer. It contrasts training capex vs inference opex; proprietary data moats vs model commoditization; and open-source pressure vs defensible differentiation. Expect scenario analyses, real-world case studies, and an investor-ready diligence checklist (ROI drivers, per-token margin targets, utilization, payback, retention, and eval rigor). Distinctive for its clear frameworks and sober risk map (energy, supply chains, regulation, hallucinations), it delivers a practical playbook to avoid capex traps and back resilient businesses. For AI allocators, it’s a compass to find durable moats-and dodge expensive mirages.