From Bits to Atoms: The AI Infrastructure Shift From LLMs to VLAs