“Cloud providers are adjusting SLAs for GB200 systems, guaranteeing uptime for only 64 out of 72 GPUs in a rack to account for higher failure rates.”