“Google's Gemini model is trained and served on TPUs, without using CUDA.”