I get the stupid basic excuse about big tech purchasing loads of silicon. That reasoning seems deeply flawed and idiotic to me. If Deepseek R1 democratizes training more for less, then that means customers with mid to small size data centers like universities now have access to train and research models in this space. Nvidia does not have a real competitor, so they get the sale. Their potential customer base just grew exponentially right? OpenAI should be devalued massively by this change, but I don’t see why anything impacts Nvidia negatively in this instance. Am I missing something or is the market this level of stupid? (I have no skin in this game)
In a gold rush, be the guy selling shovels not looking for gold.
OpenAI: we made a LLM look how good it is.
Investors: wow, here is money for more, how did you do it?
NVIDIA: we sold them the hardware to do it. Our hardware is the best, expensive to buy, and had large markups.
Investors: awww yea, time to make money, buy buy buy.
Chinese AI: we made an AI without NVIDIA hardware*.
NVIDIA/investors: wait what?
Investors: sell sell sell
They used Nvidia hardware though. Just less of it.
DeepSeek showed you can do the same work with less cards so Nvidia is going to sell less cards.
They didn’t use cuda though and there are reports of AMD chips doing very well on R1.
Yeah, they used what seems like the GPU equivalent of assembly language, which if anything leaves them even more tied to Nvidia.
PTX can be used on AMD GPUs through the Scale toolkit, which allows CUDA programs to run natively on AMD graphics processors. Scale is a CUDA compiler that uses open-source LLVM components to compile CUDA sources for AMD GPUs.
I think they are also using Huawei chips for inference. Article