A former Google TPU architect and PaLM inference-efficiency lead, he left Google a week before ChatGPT launched to design chips purpose-built for LLMs from first principles.