If my math his not totally off, a 40nm gen2 will be 10 times as fast as their current 130nm design ((130/40)2=10.5). That means that they can deploy 1PH in the same datacenters they currently use for 100TH. It could be even faster since their first design wasn't optimum.
A single chip might be faster or slower, depending mainly on how physically big it is, but also on many other factors. So the chip speed is not always proportional to chip area divided by a square of feature size. But it doesn't matter anyway.
What does matter is cost efficiency (cost per GH/s) and energy efficiency (J/GH). The former affects how cheap we can potentially sell the HW (and how much do we earn from a given retail price), while the latter determines how much hashpower can fit in any given datacenter. In both cases the math is not so simple, especially when we add NRE costs and volume savings into the equation.
It is important to remember to avoid randomly entering panic mode, thinking that 130nm is totally worthless compared to 28nm, because the latter is exactly 21.556 times better.