It's the lead free solder that's the problem on g80, 92, 94, and g200
Yeah, I think your right. I do recall this being a problem with some of the 2xx cards like the 285.
All lead-free BGA joints are more vulnerable to thermal fatigue and mechanical failure than their leaded counterparts from before 2006. They just have less elasticity.
But, what makes some models more prone to failure than others is the amount of heat dissipation through package contact surface area, and mechanical design of the cooling system. G80 is just a hot chip with huge transistors due to 90nm process. G92 shrunk considerably but resulted in a very compact package, but made amount of heat dissipated over it's small surface area too great. When G92 was shrunk down to 55nm the clock speeds also went up, so despite smaller transistors the chips didn't get any cooler. G94 was a better balance between heat dissipated and package surface area than G92 was, but I doubt you'd want one now. So here's what makes the GT200 chip so much better, it has a very large BGA footprint, in part thanks to enormous 512-bit wide memory bus, lots of BGA pins to dissipate heat through. Another thing that helped is the excellent reference cooler, that supported the board over the entire surface area and had generous thickness thermal pads over memory and sprung screws. Unlike G92, GT200 also has IHS (or "capped" if you like the slang term). People have mixed opinions about IHS, but for reliability of the card in stock form it seems to have helped. There haven't been that many GT200 boards with non-reference coolers, most were with the reference cooler. The reference cooler makes the card into a solid brick which prevents the cooler from rocking on the BGA chip.
The result was very reliable GTX 260, 275, 280 and 285 cards (and the Quadros). Because they are so cheap now there's hardly any reason to not get the 285 from among those, but a 260 is a good card also.
This trend however did not continue into the future, Fermi cards again had very high BGA failure rate. GF100 boards are just 'fugetaboutit, and wasn't really much better with GF110. Too hot for the surface area of the package and Nvidia was less generous with reference cooler design on those, despite fancy cosmetic design. Thermal pads were thin and hard, applying more pressure onto BGA chips, and boards were fastened with unsprung screws to coolers, which increased mechanical point stress. I still have several boxes of DOA Fermi boards.