Officially they spec 1.4 GHash/J but in practice it measures 1.8 GHash/J: see the "0.57 watts per GH/J" quote in the link I gave. Only a few low-performing units bottom out at 1.4.
> But if you have to replace your hardware every year this makes the whole thing even worse
You used to have to replace your hardware every year, but not anymore. There was a big race to get the ASICs to the next fab node as Bitcoin startups progressively got better and better funded and could develop for more and more expensive processes. In 24 months we went: ASICMINER 130nm -> Avalon 110nm -> BFL Single SC 65nm -> Bitfury 55nm -> Jupiter 28nm -> Neptune 20nm. But now that we have reached the top-of-the-line 28-20nm, Bitcoin ASIC developers are just waiting for foundries to make the next node available (16nm), so the lifetime of a Bitcoin miner should grow from less than a year to at east ~2 years.
If a Bitcoin miner lasts ~2 years, its hardware cost is relatively small compared to electricity costs, so replacing them is not much an issue: a $1500 2kW Neptune consumes $3500 of electricity over 2 years at $0.10/kWh. TCO = $5000. The hardware represents only 30% of that. So not only it is affordable but it makes sense to throw that hardware away and replace it for a more efficient one after 2 years.
Okay, that will gives us total transaction costs with hardware, electricity, rental, profit and so on in the range 1.5 to 2.0 of the electricity costs which again makes electricity costs a pretty good proxy for transaction costs.
So the remaining question is how the total costs of running the Bitcoin network will evolve with changing transaction volume. Will competition among miners drive up or lower the total costs? Will total costs evolve proportionally to the transaction volume - which I think is important for security - or will they decouple? It would probably be interesting to model the whole system and see what happens. Is there an equilibrium for transaction costs and Bitcoin adoption? And is it stable or will the system run away?
> But if you have to replace your hardware every year this makes the whole thing even worse
You used to have to replace your hardware every year, but not anymore. There was a big race to get the ASICs to the next fab node as Bitcoin startups progressively got better and better funded and could develop for more and more expensive processes. In 24 months we went: ASICMINER 130nm -> Avalon 110nm -> BFL Single SC 65nm -> Bitfury 55nm -> Jupiter 28nm -> Neptune 20nm. But now that we have reached the top-of-the-line 28-20nm, Bitcoin ASIC developers are just waiting for foundries to make the next node available (16nm), so the lifetime of a Bitcoin miner should grow from less than a year to at east ~2 years.
If a Bitcoin miner lasts ~2 years, its hardware cost is relatively small compared to electricity costs, so replacing them is not much an issue: a $1500 2kW Neptune consumes $3500 of electricity over 2 years at $0.10/kWh. TCO = $5000. The hardware represents only 30% of that. So not only it is affordable but it makes sense to throw that hardware away and replace it for a more efficient one after 2 years.