Spheres in a hypercube
Here is one example.
Take a square of edge length $4$. Place $2^2=4$ circles of radius $1$ into that square, in the obvious arrangement. Then place a circle at the center of all these circles, and make that fifth circle as large as possible. To compute its radius, look at the distance between the center of the square and the center of one of the large circles. That difference vector is $1$ length step for each of your two dimensions, so the distance is $\sqrt2$. As the big circle has radius $1$, your little circle has radius $\sqrt2-1$.
Now go to dimension $n$. You have a hypercube of edge length $4$, into which you place $2^n$ spheres of radius $1$ and one central sphere of radius $\sqrt n-1$. For $n=4$, the central sphere will already be as big as the outer spheres, and for $n>4$ it will be larger. At $n=9$, the radius of the central sphere will become $2$ so the central sphere will touch the containing hypercube. For $n>9$ the central sphere will no longer fit into the hypercube, even though it still touches the insides of all the radius-1-spheres contained in the hypercube.
This seems pretty counter-intuitive to me, even though the algebra is simple enough.
Finite sphere packings
A remotely related phenomenon: take a finite number $m$ of unit spheres, and arrange them in such a way that the volume of the convex hull becomes minimal. You may imagine that for a few spheres, placing them in a straight line will be best, but the more spheres you have the more compact a clustering would appear in comparison. For 3D, more compact packings exist for e.g. $m=56$ spheres.
For $n\geq42$ dimensions (note that number!) it has been proven that the “sausage conjecture” holds: for those dimensions, the straight line sausage arrangement will always be optimal no matter the number of spheres. It has been conjectured but not proven that the sausage is always optimal for $n\geq5$ dimensions. The German Wikipedia article on this subject has some nice illustrations, even if you don't understand the language.