Setting up your own with any of Ceph, Gluster, Riak, Swift or similar depending on your specific needs. Heck, I've got setups where we've been served well for years with a combination of inotifywait, grep and rsync (combined to trigger rsync instantly on modification events). It really depends on your access patterns, and there are lots of potential savings from making use of domain specific knowledge for your specific system.
In general, you can beat AWS with 3x replication across multiple data centres even with renting managed servers, especially as your bandwidth use grows as AWS bandwidth prices are absolutely ridiculous (as in, anything from a factor of 5 to 20 above what you'll pay if you shop around and depending on your other requirements). Lease to own in a colo drops the price even further.
Anything from 1/3 to 1/2 of AWS costs is reasonable with relatively moderate bandwidth usage, with the cost differential generally increasing substantially the more you access the data.
Most people also don't have uniform storage needs. If you use your own setup, people tend to be able to cut substantially more in cost by reducing redundancy for data where it's not necessary etc.
See also the Backblaze article that mentioned Reed-Solomon - if your app is suitable for doing similar you can totally blow the S3 costs further out of the water that way.
In general, you can beat AWS with 3x replication across multiple data centres even with renting managed servers, especially as your bandwidth use grows as AWS bandwidth prices are absolutely ridiculous (as in, anything from a factor of 5 to 20 above what you'll pay if you shop around and depending on your other requirements). Lease to own in a colo drops the price even further.
Anything from 1/3 to 1/2 of AWS costs is reasonable with relatively moderate bandwidth usage, with the cost differential generally increasing substantially the more you access the data.
Most people also don't have uniform storage needs. If you use your own setup, people tend to be able to cut substantially more in cost by reducing redundancy for data where it's not necessary etc.
See also the Backblaze article that mentioned Reed-Solomon - if your app is suitable for doing similar you can totally blow the S3 costs further out of the water that way.