It's not exactly half GP104 like GM206/GM204. We're talking 192-bit and 48 ROPs according to rumours (Ellesmere has 32 ROPs according to the recent GPU-Z leak). Even if it's only 1280 SPs at >1607 MHz (Geforce GTX 1080 clocks) it should be faster than Geforce GTX 970. Probably too early for performance predictions anyway.
Inno3D seems confident it will replace Geforce GTX 980 as well
While it's true that the bus width is 3/4 that of GP104, a 192-bit bus with 7GB/s GDDR5 only gives 52.5% the bandwidth of the 256-bit bus with 10GB/s GDDR5X used in GTX 1080. If the GTX 1060 used 8GB/s GDDR5, that would increase to 60% of the 1080's bandwidth. But Maxwell and Pascal are already pretty good about making the best use of memory bandwidth; I expect GP106 is going to be constrained by shader throughput in most applications.
GTX 960 had the same base and boost clocks as GTX 980. My preliminary prediction for GTX 1060 was based on the same being true here with regards to the 1080's clocks. Existing GP104 chips seem to max out at 2050-2150 MHz on reference PCBs, regardless of whether an AIB cooler is used. If Nvidia decided they needed some extra performance on GTX 1060, they could presumably let the boost clock go up to a full 2000 MHz by default without hurting yields too much. That would likely match the GTX 980 in performance, but perf/watt would suffer a bit. Final clocks will depend on what metrics Nvidia thinks is most important.
Of course GP106 will replace all the GM204-based products. This will happen even if it falls behind GTX 980 by a small margin. Nothing says the products in a new line have to exactly match the performance levels in the old line, especially since GTX 980 was not a very popular card after the 980 Ti effectively made it obsolete.