But it all boils down to the hardware. You need h100s or similar high performance gpus to train and deploy models at that scale, which is still a significant barrier for most companies regardless of their methodologies or frameworks. It’s not the lack of knowledge for the most part but the hardware