This has been discussed several times. 1200W is the minimum rated capacity for the DC output of the power supplies. Since they're about 91% efficient, they use 9.8% more power than they deliver. Also, after extensive testing and discussion with the power supply manufacturers, Spondoolies has concluded that it is safe to run each PSU at an output capped at 1360 W, which translates to input of around 1494 W.
By the way, I don't like the idea of 3000W for 1 machine ... I live in 13A@220V zone, meaning 3000W JUST exceeded the supply of 1 socket ... thus the machine will need 2 sockets, wasting the socket slot.
Also, putting too much into 1 machine means when a small part in it fails, the whole machine fails. Where as if I split into, say, 10 x S3, it's more robust, more deployable, and flexible.
I'm running into a similar issue at the DC I'm building, where the power jump from 2750W to 3000W might push me over the edge of the amount of current each circuit can supply. My approach to dealing with this issue will probably be to boost the voltage from 240V to around 254V by switching the taps around on my transformers. 254V is probably better anyway. I doubt that would be an option for you, though.
You don't have to run it at that power level. If you use the Quiet mode in the firmware, you'll get a max voltage for the ASICs of 642 mV instead of 750 mV. I would guess that would reduce your J/GH by a factor of about (642^2 / 750^2) = 0.73 to about 0.47 J/GH, and reduce performance by a factor of about 642./750 = 0.856 to around 3.8 TH/s at 1900 W. At least, that's what would probably happen if your ASICs were actually using the 750 mV max voltage, which they're likely not. Guy says 0.55 J/GH when downclocked:
https://bitcointalk.org/index.php?topic=521520.msg8091352#msg8091352By the way, you can see the voltages and power limits in the Spondoolies Minepeon code:
https://github.com/Spondoolies-Tech/minepeon/commit/f9d5535c3e525bdede299773c9d0b5f26d85fd5b