Possible reasons could be:
A - Chip includes significant power-related circuitry, which works better at older/lower resolution nodes. For example, if you're moving around a lot of power, like Thunderbolt most definitely does, it's easier to fab the circuits that control the power at an older or lower resolution (bigger) process node, like 90 nm or 130nm. That's (one reason) why many power drivers and other circuitry use the older nodes. You can mix and match nodes (5nm CPU, 90nm TB controller), but only if you're doing chiplets.
B - They had a 3rd party design the chip for them, and they wanted to reduce the co-dependence. So they make the CPU, 3rd party makes the TB controller. If they're late, no delay on the CPU.
C - They want the option to dual-source the TB controller. Perhaps they think their fab might not be able to make enough. As a separate chip, they can stick an Intel controller there, or one of their own (or 3rd party) and have more supply if one or the other runs out of chips.
D - Cost. The newest nodes (5nm) cost $$$$$$$ per piece. Older nodes like 40nm, 90nm or 130nm are much much cheaper. Yes you get more for your money, but it's not a linear cost. The newest nodes are MUCH more expensive.
But A is probably the answer. TB, USB, PCIe are all power-related, and their circuits work better on older (bigger) nodes. There's a good video on Youtube about why car chips and older nodes are in such demand and why that's not going away anytime soon.