In most kinds of logic circuit families, a XOR is slightly slower and more expensive than a NAND or a NOR, but it has about the same cost as an AND or an OR, so it does not seem appropriate to call it slow and expensive.
Even if in theory you can make any logic gate from a combination of NAND or NOR gates, the XOR gate is almost never done so.
In most logic circuit families, including in the most popular families, i.e. TTL and CMOS, there are special circuits for implementing XOR, which are better than a NAND or NOR combination.
Also, in most logic circuit families, including in the most popular families, i.e. TTL and CMOS, even if you can implement NAND with NOR and NOR with NAND, this is not done so, but there are distinct circuits for NAND and NOR.
In conclusion, in most logic circuit families you do not implement everything with a single kind of universal gate, but you have 3 kinds of basic blocks, NAND, NOR and XOR.
In certain logic families you have extra basic blocks, which can be implemented in a better way than the equivalent schematic made of universal gates.
Moreover, even if TTL circuits are usually presented as being composed of NAND gates, it is more useful to view the NAND TTL or DTL gate as an AND gate (a minimum-computing circuit made with diodes) followed by an inverter made with a transistor (which can be expanded to a NOR gate by adding parallel transistors).
So the real basic blocks of TTL circuits are AND, NOR and XOR (the NOR being not the same as the NOR from the NAND+NOR+XOR description, but a sub-circuit of that), but these simpler AND and NOR blocks cannot be used by themselves, because the AND does not have TTL output levels and the NOR does not have TTL input levels, so you always must have an AND (possibly reduced to a repeater) followed by a NOR (possibly reduced to an inverter), to make a complete TTL gate.
My point is that while it is useful to know that any gate can be implemented using only a single kind of universal gate, such implementations are not efficient and you must use a small number of basic blocks, not only one.
On the other hand, when you want to understand the theory of logic gates, basing them on one kind of universal gate is also no useful.
You can better understand logic operations when you see them as based on minimum (or the equivalent maximum) and on inversion (for values between 0 and 1 inversion is 1-x, for values between -1 and 1 inversion is -x). These 2 operations are also valid when applied to non-binary logic values. The physical implementation of these 2 operations can actually work with real-valued inputs and outputs.
An alternative view for the binary logic case is that mentioned by someone above, of having addition & multiplication modulo 2 as the basic operations.
Either the minimum/inversion view or the addition/multiplication view give much more insight than reasoning about NAND or NOR, which just combine the 2 basic operations in such a way that a cascade of the combinations can recover each of the basic operations, so they provide an alternative.