Video for those who hate reading: https://youtu.be/75XNXOSPaJ8?is=CwsyPaqTLk2nQJuE
I thought this was an interesting article on the economics of frontier labs and open source. I've seen a lot of hyping of open source models (which seem justified) but just wanted to share an article skeptical of the hype. Ben Thompson is excellent, but much of the article is speculation and reasoning rather then hard data.
Summary (generated by Opus 5)
Open weights are free to acquire, not free to run. You skip the R&D; you still pay for every token you serve. Inference is real COGS (cost of goods sold) and it scales with revenue. The cost to train is not factored into cost to serve.
Frontier prices are inflated by scarcity, not by cost. Compute shortage plus a legacy of funding training runs off inference revenue means today's prices sit well above what the labs would charge with enough GPUs.
Frontier models are cheaper per unit of intelligence. They're more efficient — better token efficiency and serving scale — so they reach the same answer with less compute. Thompson doubts Chinese models are cheaper to serve on a marginal-cost basis at all.
U.S. open-weight labs are structurally disadvantaged by distillation rules. Chinese labs freely distill frontier models as RL teachers, skipping the expensive last mile to near-frontier. U.S. open labs are bound by frontier terms of service, so they end up distilling Chinese models instead. The US government should take steps to make US open source labs more competitive (e.g. distill frontier models directly)
Lastly the Trump administration is stupid and their approach is stupid and they should feel stupid but they are too stupid to realize that (this part is not from the AI summary).