The open vs. closed AI model debate misses the bigger issue. Confidential inference secures model weights and data during runtime.