Well, the issue is that the costs involved in creating a complex chip go down over time with refined processing. That is, as you get better and better at making such things, you succeed more, and thus it becomes cheaper. By contrast, the cost of a multi-chip board like the above would be more fixed by nature (cost would still go down over time, just not by as much).
I think this is the primary thing that current GPU manufacturers have been banking on. In the short term it may indeed be cheaper to produce a, say, 4-chip NV44 board (assuming NV44 would be, for example, a 4-pipeline NV4x part) than it would be to produce one NV40 board, but as the cost to develop the NV40 chip itself decreases over time through refined engineering, it becomes cheaper to build the single NV40 board. Besides, you don't have to worry about the inefficiencies of distributed processing (it's nontrivial to allow each chip to access external memory optimally).
That said, before long it will be necessary to go for multi-chip boards in order to continue to continue improving performance. We'll just be hitting hard limits in clock speed, transistor densities, and heat. But more chips will still increase performance, and can serve to ease those burdens.
