The confusion here is that people refer to several distinct exponential growths as Moore's Law, and some of them have already faltered.
The CPU frequency growth has halted primarily due to the fact that going any higher means we can no longer cool the chips. Power usage (and, consequently, the heat you need to dispose of) is proportional to the square of the frequency (higher-frequency gates also usually require higher voltage to get the switching times down, which means it's really the cube).
Another aspect is the size of the transistor. Some features are already getting down to the point where it's more useful to talk about them in terms of monolayers--i.e., the exact number of atoms. Some elements of a 14nm transistor are already two or three monolayers in size, and the distance between two transistors is already about 100 monolayers, which puts a hard cap on the maximum possible minimization, since you can't make transistors smaller than a single atom. In terms of the smallest transistors that can be feasibly made, the general consensus is that there is at most around 3 shrinks remaining.
There's another dimension, too: whether or not it's cost-effective to keep doing these shrinks. The 14nm node itself has given Intel lots of trouble, and the 10nm node doesn't look like it's much better. Intel has already been forced to give up its Tick-Tock cycle, and the semiconductor industry as a whole may explicitly give up maintaining Moore's Law as a collective research goal shortly.
In short, then, Moore's Law either ended a decade ago, is just now ending, or will end in a decade, depending on what you want it to mean.
The CPU frequency growth has halted primarily due to the fact that going any higher means we can no longer cool the chips. Power usage (and, consequently, the heat you need to dispose of) is proportional to the square of the frequency (higher-frequency gates also usually require higher voltage to get the switching times down, which means it's really the cube).
Another aspect is the size of the transistor. Some features are already getting down to the point where it's more useful to talk about them in terms of monolayers--i.e., the exact number of atoms. Some elements of a 14nm transistor are already two or three monolayers in size, and the distance between two transistors is already about 100 monolayers, which puts a hard cap on the maximum possible minimization, since you can't make transistors smaller than a single atom. In terms of the smallest transistors that can be feasibly made, the general consensus is that there is at most around 3 shrinks remaining.
There's another dimension, too: whether or not it's cost-effective to keep doing these shrinks. The 14nm node itself has given Intel lots of trouble, and the 10nm node doesn't look like it's much better. Intel has already been forced to give up its Tick-Tock cycle, and the semiconductor industry as a whole may explicitly give up maintaining Moore's Law as a collective research goal shortly.
In short, then, Moore's Law either ended a decade ago, is just now ending, or will end in a decade, depending on what you want it to mean.