> This makes no sense. Already trading at 50x EBITDA on forecasted datacenter demand. How does this actually pan out? AWS, Microsoft, other large internet comp

> This makes no sense. Already trading at 50x EBITDA on forecasted datacenter demand. How does this actually pan out? AWS, Microsoft, other large internet companies buy a bunch of GPUs to train models _one time_, or at best at an annual cadence. Others rent once a month to train. OpenAI trained on Nvidia cards, but doesn't OpenAI's ChatGPT run on Nvidia cards as well? So it's not just the training, but the execution as well, right? I train Dreambooth models for Stable Diffusion, which takes up a lot of GPU time, but then I generate images with those Dreambooth models, which takes less time (but still uses my Nvidia card).