AWS personalize limit on number of items considered when training

Amazon personalize puts a limit of 750,000 on the number of items considered by models during the training process. Considering the situation with 2 million items (videos) with no interaction but I want to somehow show these items to users. The first approach I considered was to randomly mix these cold items with the ones coming from the recommendation model. The problem is these cold items might be irrelevant to the user so the user will not interact with the item and it still remains cold. What can I do to make these random recommendations of cold items a bit more relevant to the user? I wonder why Amazon doesn't handle more than 750000 unique items while handling 500M interactions records. Are 750000 items sufficient for all use cases?