Chinese Open-Weight Models and AI Industry Cost Dynamics
Original: Who's afraid of Chinese models?
Why This Matters
AI inference cost structures may determine long-term competitive advantage as Chinese models challenge Western pricing.
Ben Thompson's Stratechery analyzes the rise of Chinese open-weight AI models like Kimi K3, arguing that marginal costs (COGS) are returning as a central factor in AI economics, reshaping how companies compete and what 'free' models actually mean for the industry.
Ben Thompson's latest Stratechery piece examines the debate sparked by Kimi K3, a Chinese open-weight AI model approaching state-of-the-art capabilities. The article challenges a common misconception: that open-weight models are 'free.' Thompson draws a sharp distinction between R&D (a fixed cost, independent of revenue) and COGS—cost of goods sold—which in AI is directly tied to inference and scales with usage. Kimi K3 is priced at $3 per million input tokens and $15 per million output tokens, which undercuts competitors like 'Sol' at $5/$30 per million tokens, but it is not free to serve. Thompson argues that AI is unusual compared to traditional software because marginal costs are real and significant, unlike the near-zero marginal cost model that defined the internet era and underpinned his Aggregation Theory. He contends that Chinese open-weight models competing on price reintroduces classic industrial economics—where COGS, scale, and supply-side efficiency matter—back into what many assumed was a software-economics game. The broader implication is that both the short-term competitive dynamics and the long-term industry structure of AI will be shaped by these inference cost realities.