From memory friedcat indicated that the cost of further chips was under $1 each - to that has to be added the PCBs, cooling, power supplies, rack-mounts etc. But the bulk of expenditure to date has been on the actual development of the ASICs not the production of them - and that's a one-off sunk cost (for this design).
The problem with this is that from memory giga's contract seems to have been perpetual at some point. People who avoid making actual contracts (which means, gpg-signed) usually avoid it for a reason.
That aside, from what I understand the entire AM run was pretty much based on mooching. That is to say, someone knew someone who was able to insert some chips into somebody else's tailrun or such.
This does not scale. It is perfectly true that if you are a normal customer doing a small run, coming back and doing that run 10x will probably cost you 80% less per item (so you get the 10x run for 2x what you paid on the original run). However, if you're a mooch customer you will be asked to actually pay normal rates on your 10x run, which may very well
increase your per-item costs. This, obviously, being all speculation, we've not seen for instance a single bill issued to this company and paid by it. (Which is, incidentally, no way to do business if you have a real world side).
That's why you can't compare ASICMINER to other mining companies - as they just can't expand their ASICs for the same low cost ASICMINER can.
You can't compare Avalon to BFL because Avalon just isn't as cool as BFL, says BFL. Unless you have those other companies on the record saying this, I'd exercise some caution.
We don't have enough information to accurately work out the cost/TH for ASICMINER to produce more hardware - but I'd be shocked if it was anything like as high as 25% of what other miners currently have to pay to get BFL/Avalon hash-power.
Not only would I not be shocked at least, but more importantly: AM has failed so far to report their chip blowout rate. We don't know how many chips that were cut out of the wafer had to be discarded at bumping, we do not know how many PCBs had to have chips replaced. Common sense would indicate the former figure is not zero. The evolution of hashing power seems to indicate that the latter figure is also not zero.
This particular tight-lippedness, especially if coupled with the other naive/noobish mistakes they've made to date would seem to suggest we are confronted with yet another run of the "lala company" let's call it. That is, the company which sweeps any sort of trouble under the rug in order to maintain what they misrepresent as "respectability", only to crash spectacularly once the quantity of shit under the rug becomes larger than the rug itself. If you were to believe Tom, everything was fine and they were set to take over the world up until five hours before going bankrupt. If you are to believe Inaba BFL will actually at some point deliver. On it goes, the history of BTC is replete with examples. Of particular concern in this line of reasoning is that the updates they do release become increasingly longer, increasingly content-free (even on an absolute basis, leave alone percentages) and increasingly vague. This circumstance works with the theory admirably well (even though, as you point out elsewhere, it does not constitute proof).
We already know BFL are happy selling their FIRST batch at a 50% discount (in exchange for return of pretty much useless to them FPGAs) - which is the batch in which they need to be making serious inroads into development costs (as price-cutting won't be too far away once the back-log gets cleared). That indicates that production costs are such that they could sell at half the price and make a good profit (if not, then they're majorly screwed as they'll never recover development costs).
I'd remind you that a. BFL is a horrible company, created by a convicted felon in violation of his parole; b. BFL has a record of lying atrociously about everything, times, prices, specs, you name it; c. BFL burned a wafer just because they could. More generally, this "X is ok because it's better than Y" is a very poor criteria, seeing how pretty much all the ASIC makers are a little cuckoo, so to speak. Certainly more powered by ambition and forum psychosis than actual experience, capital and ability.
I can't see component/PCB/assembly costs being higher for ASICMINER in China than for BFL in the US.
I can however see them being lower for Avalon, for instance. Those guys have a history of underpromise and overdeliver. These guys do not, in spite of the community effort to paint pretty colors on it. Moreover, it is my hunch that the major ASIC player has not yet emerged, and by December 2013 we'll all be buying our ASICs there. It's after all how these things usually play out - but in any event, presuming that the players are all known and proceeding deductively from there is fallacious.
And do remember that the next 50 TH of chips are already done - with assembly starting soon.
How many of them are being used to replace blown first batch chips is a much better question.
) 
