Arvind Krishna on sovereign AI (Q3 2026): public LLM use leaves traces in logs; enterprises turn to on-prem AI infrastructure and open-weight models they can run on private clouds with controlled logs. Bob now runs on-prem for regulated clients. Prepared remarks cite Apptio for AI spend visibility, llm-d (with Red Hat/Google) for 5–10× lower cost via context reuse, and Spyre cards for mainframe AI acceleration.