Open Weights Are Not Free: The Economics of Chinese AI Models

My notes

Summary

Stratechery argues that panic over Chinese open-weight models (Kimi K3, Qwen3.8 Max) is mostly an economics misread. R&D is a sunk fixed cost, but inference (COGS) scales with revenue, so “free” open weights are not free to serve. Frontier labs like Anthropic and OpenAI likely hold the better cost structure and charge premium prices mainly because compute is scarce, not because Chinese models are genuinely cheaper on a marginal-cost basis. The real risk is not pricing, it is cybersecurity, where US frontier model guardrails have pushed defenders toward Chinese models.

Key Insight

• R&D vs COGS distinction: R&D is a fixed cost independent of revenue, while inference COGS scales directly with revenue (100x more revenue means roughly 100x more COGS at a given token cost). This is why “free” open-weight models are not actually free to run. Kimi K3 costs $3 per million input tokens and $15 per million output tokens, cheaper than a frontier model’s $5/$30, but not zero.
• Tokens are not a commodity, intelligence is. A token from one model is not fungible with a token from another because different models need different amounts of reasoning and agentic tokens to reach the same correct answer. Kimi reportedly needs significantly more tokens than top frontier models, eroding its apparent price advantage. What is fungible is the correct answer, not the raw tokens.
• Five COGS-of-intelligence levers: model footprint (memory and accelerators per serving replica), inference efficiency (e.g. Mixture-of-Experts), memory efficiency (KV cache size, which drives concurrency), serving efficiency (batching, scheduling, prefix caching), and token efficiency (fewer tokens to a correct answer means lower cost).
• Commodity market mechanics: in a true commodity market, price is set by the marginal (highest-cost) supplier needed to meet demand. Low-cost suppliers earn fat margins, the marginal supplier earns roughly nothing, and anyone above marginal cost carrying fixed costs or debt risks bankruptcy. The numeric example in the piece: three suppliers at $10, $15 and $20 cost, price clears at $20, so supplier A earns $10 per unit, B earns $5 per unit, and C earns $0 per unit and eventually goes bankrupt.
• Why frontier labs still command a premium today: demand for frontier AI compute currently exceeds supply, creating a price umbrella. Frontier labs can pay markups to compute resellers and still profit because they charge higher markups downstream. The market is not yet treating intelligence as a true commodity because buyers still prefer specific top labs.
• Frontier lab “paranoia” explained: (1) a legacy mental model where training costs dominated, now shifting as the inference market grows faster than training cost; (2) applied inference generates data that improves the next model iteration, a moat that open-weight competition threatens; (3) product and harness stickiness (coding agents) as a differentiation layer beyond raw model quality; (4) a safety ideology that is undermined by the mere existence of viable open alternatives.
• China’s strategic play: commoditize your complement. Xi Jinping’s July 2026 speech explicitly linked AI openness to bringing AI into the physical world (robotics), where China already has leverage. Alibaba reversed its earlier decision to stop releasing open weights, reportedly in response to that speech.
• Distillation as structural advantage: Chinese labs use frontier US models as teachers for reinforcement learning post-training, compressing the expensive last-mile gap between a strong base model and near-frontier performance. Even Western open-weight builders reportedly rely on Chinese models for RL cold-start, meaning Western open models end up distilling the distillation via China, a genuine structural disadvantage.
• The real cybersecurity risk (the one legitimate fear): Hugging Face’s security team, breached by an autonomous AI agent, found that US frontier model guardrails “cannot distinguish an incident responder from an attacker” and blocked their own defenders. They used China’s open-weight GLM 5.2 model instead, run on their own infrastructure, to analyze 17,000+ attack logs. US policy restrictions on using top US models for cybersecurity work are pushing defenders toward Chinese alternatives.
• Policy recommendation in the piece: loosen restrictions on using frontier US models for cybersecurity, pass a law making AI training data collection explicit fair use, and bar distillation-blocking terms of service for US companies, on the grounds that stopping distillation (which is literally just querying the API) is nearly impossible anyway.