On what are fast and new LLMs created if not on a GPU? The reason Nvidia dominates is that all the AI research is done on their GPUs for the past decade. That'

On what are fast and new LLMs created if not on a GPU? The reason Nvidia dominates is that all the AI research is done on their GPUs for the past decade. That's why any new approach in AI will be on Nvidia's GPU and Nvidia will always fine tune their ecosystem based on research results. That's why the 6 months of A100s used to traing GPT-3.5 become 10 days with Blackwell. That's why Nvidia dominates since nobody even considers alternatives in AI research validation than Nvidia. And one major reason is that CUDA is dominant and AI researchers have known it for a decade. You see, Nvidia is talking about "all AI", so training, inferencing, networking, applications, everything. HW competitors only mean inferencing when they talk about AI and wonder why traction is so slow. Nobody on this planet would buy a Nvidia data center for research and training and then a competitor's data center for inferencing next to it. You simply extend your Nvidia data center seamlessly (thanks to Nvidia tools) and use it dynamically at full utilization for training, inferencing and whatever you need otherwise. To beat Nvidia, you have to beat them at everything, not just inferencing and even there on specific fields lol.