I agree on the conclusion that fee depends on what users are willing to pay, since if it gets too expensive they just stop using the service, but I don't agree with the pessimism.

Can we increase it without incurring a cost or stress-testing bottlenecks? No.

Yes we can. We've already tested that completely filled 2 MB blocks (currently max allowed for witnesses to set) with over 7k transactions per block (somewhat mainnetish mix of comments, votes, transfers and custom jsons) are totally fine for normal desktop PCs with a lot of margin. That's about consensus (and witness) nodes. I fully expect that after we implement the optimizations that are already on the wish list (and we know how to make them), consensus node should be possible to run on an old smartphone level of hardware (by old I mean something like my yet-to-be-replaced SG9+ 😉). Of course it gets more hardware hungry when you want to run HAF, Hivemind and whatever other apps are going to sprout in the future to push price of Hive to 2k$. While normal servers can handle the data intake from such large blocks (tested), amount of data is going to be challenging for PostgreSQL - the database grows quite fast and we are yet to perform prolonged tests in such environment. But that only means we are going to need to work on the solutions. F.e. we could have slower but more capacious servers to handle relatively rare queries for old operations and fast servers that only keep data from last month and handle most of the load (requires implementation of HAF pruning). The bulk of the cost is likely to be not in the hardware, but in network connections. But with expensive Hive it should not be a problem 😊 and even backup witnesses should be able to afford to have "full nodes". Third party apps are likely have their own servers where they only keep operations that are needed for their specific app.