Well, a graphic card is not just the chip though as it also contains power hungry GDDR5 memory and VRMs etc. An LGA2011 CPU running at stock might be 125W but the motherboard will draw a lot more than that (LGA2011 with a bit of overclocking will easily consume as much as an AMD's FX8350 for example).
There are a few trends pushing these power guzzling GPUs though:
- Silicon process nodes have slowed down a lot (no more 18 months cycles and even when they finally come the node may not offer much for high power draw chips like GPUs: TSMC's 28nm to 20nm will not only have taken a lot longer than 18 months it will also be optimised for low wattage SOCs).
- Some (so-called) hardware enthusiasts are willing to pay far too much for some kind of 'premium' product.
- GPUs are very parallel which and once the basic blocks are designed (GCN, Kepler etc.) it is relatively easy to scale those up.
- While it is possible to get more performance out of running large chips slower, very large dies tend to have yield issues. So large-and-slow is an approach more suited to mobile.