The S3 bucket is hosted/managed by the U.S. Geological Survey. The only costs borne by them are the hosting, S3 requests, and S3 egress. In this case there's no egress since I fetch from the same region, so the marginal cost of this experiment to them is roughly $0.000001 per tile a user loads.

They've chosen to leave this particular bucket fully public for the time being. If usage got too high, they could convert the bucket to a requester pays bucket, where the requester (me) would pay for the S3 requests and egress. They've done this for their Landsat bucket (usgs-landsat), which is set as requester pays.