Newsletter digest — July 21, 2026: Why open-weight AI is an inference-cost story

AI infrastructure and model economics

Stratechery: "Who's Afraid of Chinese Models?"

• Open weights are cheap to obtain, not free to serve. Thompson separates fixed research and development spending from variable inference COGS, then notes that Kimi K3's listed price is $3 per million input tokens and $15 per million output tokens. A provider still incurs serving costs as usage grows, unlike traditional software with near-zero marginal cost. 1
• Tokens are a poor comparison unit for reasoning and agents. Different models can use different amounts of reasoning or workflow tokens to reach the same answer, so a lower price per token does not automatically mean a lower cost per unit of intelligence. Thompson's cost model also includes model footprint, inference and memory efficiency, serving efficiency, and token efficiency. If several providers can handle a basic task such as building a CRUD app, he expects the cost structure to matter more than a headline model price. 1
• The open-weights strategy has an industrial and security angle. Thompson describes China's approach as "commoditize your complements": widely available models could support innovation in physical-world industries, while reducing the advantage held by U.S. frontier labs. He identifies cybersecurity as the sharper risk, citing a Hugging Face incident report in which defenders turned to a Chinese open model after unnamed U.S. models' guardrails obstructed their analysis. His proposed response is to give defenders access to capable models they can run locally and keep U.S. open-weight builders on an equal footing. 1

Product and growth

One thread to watch