Deepseek - A model chinese researchers developed and trained was trained under 6 millions whereas researchers in US has burned more than bllions in training.
Art of training model doesn't only depend in how strong the compute performance is but also fine tuning and model architecture. I believe companies will start focussing on doing more with less. Scarcity is mother of all invention.
This should lead to weakening demand in NVDIA for training.
- Most of the companies wants to leverage AI capabilities which require inferencing. AMD inferencing server is cost effective and performant.
- AMD announced integration with deepseek for their new chips!