Who’s Afraid of Chinese Models? | Stratechery by Ben Thompson

Episode Timeline

1. 0:01

   Ben Thompson revisits his Kellogg strategy class to argue that AI brings marginal costs and commodity-market dynamics back to tech, contrary to the zero-marginal-cost logic of the internet era.

   • Thompson's strategy professor taught that universal principles apply across industries, but he initially believed tech was different because software had zero marginal costs and zero transaction costs.
   • AI reverses this: running inference costs real money, so COGS is directly correlated to revenue, unlike the fixed expense of R&D that open-weights models let you skip.
   • Open-weights models like Kimi K3 are not free to serve—Kimi K3 costs $3 per million input tokens and $15 per million output tokens, cheaper than Claude's $5/$30 but still a real marginal cost.
2. 2:08

   Tokens are not a fungible commodity because reasoning and agentic workflows require different token counts per model, making intelligence the true commodity.

   • Jensen Huang's 'token factory' framing works for the ChatGPT era but breaks down in the reasoning era, where chain-of-thought tokens explode and Kimi reportedly uses far more tokens than Claude, negating its price advantage.
   • A token from one model is not equivalent to a token from another, but the correct answer constructed from tokens is fungible—so COGS for intelligence depends on model footprint, inference efficiency, memory efficiency, serving efficiency, and token efficiency.
   • Intelligence for many economically beneficial tasks is becoming a commodity, so profitability will come from superior cost structure rather than higher prices.
3. 4:10

   Thompson walks through commodity-market mechanics to show that the highest-cost supplier sets the market-clearing price and risks bankruptcy.

   • In commodity markets everyone charges the same price set by supply and demand, and the supplier with the worst cost structure sells at their marginal cost, earning zero profit.
   • Using suppliers A, B, and C producing at $10, $15, and $20 per unit with demand for 25 units at $20, supplier A earns $10/unit, B earns $5/unit, and C earns nothing and goes bankrupt.
   • Fixed costs like R&D and debt financing come back to the forefront because bankrupt suppliers can't price with those costs in mind, and their exit raises prices until new supply enters.

Key Concepts

• open weights models— Chinese open-weight models like Kimi K3 and Qwen 3.8 Max are the central subject of the episode's analysis.
• marginal cost— The return of marginal costs (COGS) to AI is the core economic argument distinguishing AI from classic software.
• commodity market— Thompson argues intelligence is becoming a commodity where profitability comes from superior cost structure, not higher prices.

Notable Quotes

• A token from one model, however, is not the same as a token from another model. What is funible is what is constructed from tokens, which is to say intelligence.

  💡

  — It overturns the common assumption that tokens are a uniform commodity and relocates fungibility to the intelligence they produce.

  Jump to 4:10

• I highly doubt that Chinese models are cheaper to serve on a marginal cost basis. They just seem cheaper because anthropic and openi are so supply constrained that they are charging far more than they would if there were sufficient supply to meet the demand for intelligence.

  💡

  — It reframes the apparent Chinese price advantage as an artifact of Western supply constraints rather than a genuine cost advantage.

  Jump to 9:16

Actionable Takeaways