Kojima and team have delivered software that pushed the envelope time and time again
Well, Metal Gear Solid 2 pushed the envelope for the prologue in the ship only, which is long at most one fourth of the game. The rest is franlky sub par both gameplay and graphically wise, and the difference is so appalling it seems they developed a grand demo which took most of the time and budget in order to impress, followed by another game which was short on both. I felt almost cheated when I played the latter.
Giving a logical guess, during the development of MGS2 they noticed it would require much more time to keep up with the complexity of the first part of the game, and lowered their profile... which isn't very different from what we are reading about MGS4.
I didn't play MGS3 but from the movies it looked very interesting while not revolutionary graphically wise (after all they already did MGS2 using the same hardware and knew their technical and artistic limits).
Sony's choices for the hardware should look amateurish to industry veterans: for the results they were aiming to a multicore processor is far less important than a solid GPU. Games are all about illusion of realism and everything is approximated, so they only needed to push the graphics and a CPU core which excells at floating point calculations isn't simply suited for the job, except maybe for a few specific tasks such as physics (which anyway don't require high precision in gaming) and audio processing (which isn't needed to this extent). It's almost as if they really didn't know about what they were thinking. They also choose a processor which gives about 10% / 20% yield results (that is, at least 80% of the times a cell processor comes out it has less than 7 working SPE cores, thus being unusable for this job), which is not economically viable in the consumer market (the processor was in fact originally intended for research purposes) since you are also paying for the processors which failed testing, as often happens for example with professional audio equipment, with the difference that you are not getting any benefit out of it. Not to mention the need to reinvent the wheel from scratch for the engines available on the market, since are all based on a very different architecture (again, this doesn't matter in research since they would write anyway their software from scratch). When the price will be acceptable, the hardware will be already old (and it looks like it is already happening).