I remember when Wallstreet thought Deepseek would reduce GPU demand due to efficiency gains. Most people here thought Jevons Paradox would apply.

+ Most people here

- Wallstreet

but:

China built hundreds of AI data centers. Now many stand unused

https://news.ycombinator.com/item?id=43492171

But even in that article, it mentions that H100 are not in high demand anymore but the H20 is. Same message as this article.

Deepseek killed a lot of AI labs looking to train their own model. Thus, H100 is not in high demand. But inferencing Deepseek is in high demand and the H20 is better for that task apparently.