You can't say that DeepSeek is failing. Rather, you can see that the AI has been trained with Chinese content that has just been filtered. Of course you can also train DeepSeek with other content. BTW there is also a lot of woker content in ChatGPT because they use Wikipedia as a source.
DeepSeek's performance when building LLM models is better by several factors. This is based on the MoE approach Mixture of Experts. Instead of calculating the entire network (as with ChatGPT), hotspots, so-called expert networks, are calculated. In other words, only parts. This results in savings in the calculation. But this is open source and described in numerous scientific articles. So it can also be used by competitors or improved even further.
What we may see is that the dominance of HW is tipping in favor of SW. So maybe out of $NVDA or $AMD (nothing is happening there anyway) and into SW $PLTR
Question: How many NVIDIA GPUs and how many AMD CPUs does DeepSeek use?
DeepSeek: DeepSeek typically uses **NVIDIA GPUs** for training and inference of AI models. As for **CPUs**, it is less specifically known whether DeepSeek uses AMD or Intel processors.
ChatGPT: I'm sorry, but I don't have any specific information about the number of NVIDIA GPUs and AMD CPUs DeepSeek uses.
Grok: DeepSeek uses: - **NVIDIA GPUs**: 12,288 NVIDIA H800 GPUs and 1,024 NVIDIA A100 GPUs. - AMD CPUs**: It is mentioned that DeepSeek runs on a dual-socket AMD EPYC platform, which indicates the use of two AMD EPYC CPUs per server. However, since it is not specified exactly how many servers are used, the exact number of AMD CPUs cannot be determined exactly. determined.