Why AI economics are fundamentally broken (leaddev.com)

AI economics are fundamentally different from traditional software: every AI interaction carries a non‑negligible, ongoing compute cost, so marginal costs don’t vanish with scale. The article argues this breaks the classic software playbook (“grow now, monetize later”) because user growth directly increases the largest expense line—API/inference costs—rather than diluting fixed costs. Big-picture evidence: ChatGPT inference has been estimated at ~$700k/day, training newer models has exploded (GPT‑3 ≈ $4.6M, GPT‑4 > $60M, GPT‑5 projected > $500M), and venture subsidies have been masking true API pricing that will likely rise as those subsidies end. Technically, the dominant cost driver is conversation length: because each turn reprocesses the entire history, input-token costs scale roughly quadratically with turns (cost ∝ P_i × N²) while output tokens scale linearly (cost ∝ P_o × N). Agentic workflows multiply that cost by the number of tool calls (k×O(N²)). Model choice and token pricing therefore matter differently by use case—low P_o models win for short exchanges, low P_i models win for long chats—and price differences between models can be orders of magnitude. Practical implications: teams must design for sustainable unit economics from day one, using strategies like dynamic model routing, context compression, RAG, response limits, and careful product design to allocate expensive models to high‑value tasks while keeping routine interactions cheap.