R2’s writes are much slower than S3/GCS. It’s not a good fit unless your workload does many reads for each write (e.g. assets). Also it is notably not multi-region - it just picks a region automatically based on where you created the bucket from.

KV is super expensive - once you’re operating at non-trivial scale, reading another configuration value per-request in your worker starts to cost thousands per month. KV’s tail latencies are also surprisingly bad (I’ve seen over a minute), even for frequently read keys that should be easily cached.