Why do we need 12 atoms to store 1 bit of data? Recent research at IBM has found a way to store 1 bit of data in 12 atoms.
While that is a big accomplishment compared to what we have today, it does seem like a waste to a non-physics eye like me.
From this figure on the same page:

it looks like we can determine 1 and 0 based on the alignment of magnetic properties of 12 atoms.
But why is a smaller unit, like just one atom not good enough?
 A: 1 atom can be used to store information, it's just not quite as convenient.
10 years ago IBM used an Atomic Force Microscope (AFM) to write the word "IBM" in individual Xenon (?) atoms on a silicon wafer. 
Of course you needed an AFM and the sample had to be kept at a few Kelvin in presumably a rather impressive vacuum.

A: I don't think that this is a physics restriction, but one of current engineering capability.
As you link points out, using 12 atoms allowed the information to be retained without effecting the information stored next to it.
You will also need enough data-mass to allow for the reading and writing of the information without affecting the data next to the one of interest.
In theory the binary data could be stored in other characteristic so an individual atom, but we (IBM) currently don't have a way to do this.
A: For any kind of magnetic data storage you need a magnetic state that is stable over time. The magnetic moment of an isolated single atom does not have any preferred direction, therefore the energy states are degenerated. 
The 12 atoms used in this experiment is not a lower limit, in principle it can also work with 2 atoms given the right magnetic interactions between these two atoms. As you lower the number of atoms the probability of a flip of the magnetic moment due to thermal fluctuations of the whole cluster increases dramatically, so your information would be lost after a short time.
One atom can only be used if you use a different property, like position to store data. IBM even developed such a system with an atomic force microscope that could in principle use only a single atom but I am not sure how far they are with a practical application (Millipede memory).
