Well, their huge GPU clusters have "insane VRAM". Once you can actually load the model without offloading, inference isn't all that computationally expensive for the most part.