It's not that hard (I've done this recently for clients). If your traffic is 90% of normal for a few months, scaling down with load makes a lot of sense, then you can scale back up when ready. There are also autoscaling solutions like serverless.
This is true for simple services. It's different when you're dealing with something like a Cassandra or Elasticsearch cluster, where the system isn't stateless and scaling up takes some time for re-balancing.
Sure, it won't apply to all services, particularly datastores are best left alone, but if you need a quite few web workers say which sit in front of your search or db to service millions of users, and can cut that in half during low traffic times, you may save significant money, as the user pram above has noted.
