3 Comments
User's avatar
Patrick Mathieson's avatar

From what I understand the $5.6m figure for DeepSeek is not apples to apples with U.S. training run costs, because this model relied on distillation from existing LLMs and also omitted plenty of costs (e.g. the physical infrastructure itself). Not sure why this wasn't mentioned in the post.

Expand full comment
Alexander Campbell's avatar

Yup, totally fair. I had an earlier version where I tried to break down the total cost and suggest they didn't even have enough cash to finance the hundred million stack. Maybe I'll edit this thanks.

Expand full comment
Mackay miller's avatar

Great collection of data points, though it seems odd to lay it all at the feet of doomers. It’s not like doomerism on its own explains why China is taking the lead. Your piece points to the other factors that are independently crucial, eg power infra, open source, state-led coordination, etc.

Expand full comment