10^12 is how much more work you have to do (according to the complexity estimate for the GNFS), regardless of how you do that work. If your custom-built hardware is 10^12 times faster than a general-purpose CPU, you can factor a 2048-bit number on your hardware as fast as a 768-bit number on general-purpose CPUs, but it will still take 10^12 times more work than factoring a 768-bit number on the same hardware. And I've never heard of GPUs or FPGAs giving a speedup of more than perhaps 10^4 on any problem, which still leaves a factor 10^8.