From Stable Diffusion to Veo3, why generative media is completely different than LLM inference, and how to scale to $100M ARR while writing custom kernels| www.latent.space