Waiting longer doesn't mean you get wider buses for the same cash.
There are two kinds costs in electronics -- the kind that halves it's price every year and a half, and the kind that stays constant regardless of time. More signal lines and more memory *chips* are static costs, whereas faster chips and larger chips are the kind of costs that go down with time. For a console that will presumably be produced for 5 years+, this is very significant -- if you pick a wide bus with slow chips, your console will never be cheaper to produce than today. Because of this, they will have a narrowish bus that connects to the fastest and largest chips available at launch.
I still disagree with that. Nvidia has a product that see ~three shrinks the old GTX8800. The ship almost still ship nowadays with few changes, it started with a 384 bits bus wide at an high price and chip nowadays at ~100$, there have been really few changes still they moved from a 384 bits bus to a 256 bit one. The move from 90nm to 65nm to 55nm. It's an interesting "case study" but with Nvidia renaming frenzy I really have a tough time gathering proper informations on die sizes as I wonder if they moved from 384bits bus to 256 bits one in one jump or if they pass by by a 320 bits one (if not with this exact product but one of the same family).
Anyway point is that in a corner case you could say that it's the tiny bus that doesn't scale down. For example AMD GPU use 64bits wide memory controllers, so you have two of them on a 128 bits wide bus, so far so good. I would not be surprised if the configuration is the same in Xenos ie two 64 bits wide memory controllers so if MS want to use a 64bit bus they would need GDDR3 clocked @1.4GHz. Either it's not available or to costly but they did not deemed this choice as worthy (they are not concerned by cost? I don't think so).
On the other hand if with their next system they start with a 256 bits bus and nowadays cheapest GDDR5 clocked @900MHz, say two or three years down the road they could move to a 192 bit bus and use GDDR5 clocked @1.2GHZ which will be super cheap at this time and offer overall the same characteristic.
Then there is the matter of the number of RBEs tied to a given memory controller, but if they plan ahead it may not be a problem depending at which granularity AMD can disable RBE on a given design. HD 6870 has 32 RBEs 8 per memory controller, on a 192bits bus that would be 24 RBEs, say one is planning ahead they could disable 2 RBEs per memory controller in the first revision (coarse grain redundancy that may help with yields too).
Anyway point is tiny bus don't scale as to shrink them you need the bandwidth per memory controller to go up by a 2X factor which is unlikely to happen.
Overall if we were to go by this absolute belief that wider bus is not an economical option the 360 would be stuck on 64bits bus as the xbox (or a bit of a sarcasm the GTS 250 would not sell at 100$), moving from a 64bits bus to a 128bits was also a X2 increase with associated costs. Point is manufacturers need that extra bandwidth, actually MS went further they added an EDRAM chip (even more cost) to meet ~720p bandwidth rendering requirements (and to have a single pool of RAM which comes handy). Nowadays is another matter, cheapest GDDR5 will buy you +100GB/s worse of bandwidth on a 256 bits bus, modern GPUs do really well at 720p and more (and with AA) with this bandwidth. In few years cheapest GDDR5 will offer the same bandwidth on a 192bit bus bringing cost down 'it needs only GDDR5 @1200MHz). And one will will never have to deal with EDRAM an extra chip and its associated costs. In the same time given enough bandwidth (and they would have enough) editors will be way happier with a single pool of RAM for real and no limit and no arbitrarily set size for say their G-buffer.
