AI models as standalone P&Ls (philippdubach.com)

Microsoft’s filings hinted that OpenAI may have lost roughly $11.5 billion in a recent quarter, fueling scrutiny about AI economics: reportedly spending about $5 to deliver $1 of revenue. That underscores a core industry tension—companies must keep building ever-larger, costlier models to stay ahead of open-source competitors, or risk commoditization. The conventional accounting picture looks dire because successive model generations require exponentially more upfront training spend (example: $100M → $200M revenue, then $1B → $2B, then $10B planned), creating overlapping years of apparent losses even if each model individually returns roughly 2x its training cost. Anthropic CEO Dario Amodei reframes this by treating each model as a standalone P&L: each model can be profitable on its own, but running them in sequence—where each generation costs ~10× the prior—produces company-level losses during the scale-up. This reasoning rests on two fragile assumptions: (1) each model reliably returns ~2× its training cost after inference and deployment, and (2) customers will pay enough for incremental capability as training costs balloon. Two outcomes follow: either scaling hits physical/practical limits and firms harvest profits from final-generation models, or improvements plateau and firms face a costly “overhang.” Crucially, open-source parity or faster commoditization would collapse the premium window and invalidate the standalone-P&L defense, leaving profitability and business models exposed.