Good question.
As it is, a random number can only be chosen randomly from a given range of numbers and the range is not itself random and it won't contain an infinity of numbers.
Here, we would need instead to have an infinite range of numbers to choose from, and we would therefore need to be able to somehow represent an infinite range of numbers, and not just in the abstract by writing (0, ∞). An interval of Real numbers, say [0, 1], contains an infinite number of Reals, so that's a start, but that's also a theoretical abstraction, something which is not good enough for computers. Computers don't have any representation of this infinity that they could use to pick a number from. Typically, a computer can only represent numbers between 0 and 1 that have a decimal part with a given and finite number of positions, which will only represent a finite number of Reals.
I think you could go beyond this limitation by using a recursive algorithm. You could generate random numbers that would be arbitrarily large. However, I think that in this case, the larger the number, the longer the time to produce it and you end up selecting numbers that are not at all random in an infinite range, but random within the range that your computer can handle within a reasonable amount of time, which is presumably a finite amount of time, producing a finite range of numbers to choose from.
So, I think computers don't have any infinite range of numbers to choose from for creating random numbers.
In theory, if real space is continuous, picking a particular point in space could produce a random number from an infinite range. However, the range from which this number would be so obtained would be limited by the precision of the measure, so that you would also end up with a finite number of measures to pick from.
Brains, however, are similarly limited and for broadly the same reasons. So, while this limitation will apply to all AI systems, it applies to them no more so than it applies to the human brain. So, this is a factor to take account of but not one that would limit AI relative to human intelligence.
EB