What strategies would you use to reduce training and inference cost for large foundation models at Datadog?
