StreamSense runs a weekly hyperparameter sweep that normally costs $12,000 for on-demand ml.p3.2xlarge instances. To reduce cost they enabled SageMaker managed spot training and configured checkpointing and a max_wait_time 2x the training duration. The cost report for the sweep shows the managed spot run billed $1,200. Based on SageMaker managed spot behavior, what maximum savings percentage vs the original on-demand cost is most consistent with SageMaker managed spot training potential savings?

Amazon SageMaker Managed Spot Training can provide significant cost savings, often up to 90% compared to on-demand instances. This is because it leverages Amazon EC2 Spot Instances, which offer unused EC2 capacity at a discounted price. The "maxwaittime" and checkpointing configuration further optimize for cost by allowing the job to wait for spot capacity and resume training from the last checkpoint if interrupted, maximizing the utilization of cheaper spot instances. The reported $1,200 cost for a job that previously cost $12,000 represents a 90% saving, which aligns perfectly with the potential maximum savings of SageMaker Managed Spot Training. The other options are incorrect because they underestimate the potential cost savings of this feature.