It will be interesting to see how this will work out in the long-term. Currently the cost of Bitcoin transaction is already at around $50 which makes Bitcoin transactions quite expensive. The only reason why this cost doesn't matter yet is because the inflationary factors (new coins, etc.) are smaller than the deflationary factors (more adoption, etc.).

I see Bitcoin as something like the 2nd-gen P2P file sharing systems (Gnutella, etc.) where a radical decentralization approach was taken to counter weaknesses of the earlier generation (P2P: Napster, Currency: E-Gold). In the P2P world we figured out that a small amount of centralization provides benefits and is acceptable (e.g., Bittorrent). I suspect that in the e-currency world we'll be going a similar route, where successors to Bitcoin will use Ripple-like concensus systems instead of proof-of-work [2]. This should not only allow to significantly reduce the cost per transaction, but might also allow to make the network more secure in practice (as the number of validators in a concensus approach will be much larger than the number of mining pools with signficant hashing power in the Bitcoin world).

[1] that's basically the cost of finding a new block divided by the number of transactions per block - see https://blockchain.info/en/charts/cost-per-transaction.

[2] Proof-of-stake is a nice theoretical approach but due to the complexity involved it will probably take some time until a solution is found that also works well in practice. Proof-of-stake as implemented in Peercoin is basically outdated, something like Slasher (http://blog.ethereum.org/2014/01/15/slasher-a-punitive-proof...) looks more like a viable solution.

EDIT: As there has been some confusion about the $50 transactions costs: I'm not talking about direct costs for the person who initiated the transaction, but about externalized costs that are paid by the whole network. See my responses below.