Regardless of whether PD use samples or AES, they still need to record cars.
Recording isn't the difficult part of producing samples, largely because it's such a mature and accomplished industry. In the case of both Forza and GT (there is more evidence for the latter than the former, that I'm aware of), there are cars which have been recorded but those recordings not used in game (modified sounds in Forza's case, examples like the stock R8 V8 in GT's case), and there are still more cars they recorded but that aren't even in the games.
The difficult part is asset production from raw recordings, which is an artistic task, and management of those resulting assets. The size of that management task grows in a non-linear fashion with the number of assets, whether that be a large number of cars with few samples each, or a comparatively few cars with many samples each, as time (and hence quality targets) progresses.
For AES, the asset management is perfectly linear with car count (1 file each), and the quality is dependent on the synthesis engine, not the number of assets. That is, synthesis is code-limited, samples are content-limited in their range of expressiveness (and resulting perceived "quality"). For synthesis, you pay the same few programmers and a fixed number of sound designers for a fixed car count to upgrade the sound scheme. For samples you pay incrementally more sound designers every time you need to increase the quality, even on a similar number of cars.
So in the case of samples, the size of the task will continue to increase at an untenable rate for the sake of competition (especially against synthesis), whereas the synthesis approach will level out fairly quickly (because content will be limited by other tasks, e.g. modeling). Equally, the same content can be re-used in the synthesis case, even for a major overhaul of the quality and features, because the system is based on physical parameters in the first place, which can't have changed. In the case of samples, if there's a change in the way the sounds are to be represented and played back, e.g. longer loops, more samples across the rev range etc., the whole sample bank needs re-doing and the existing loops (not the original recordings, necessarily) are effectively useless (in the same way that old models are).
That's not to say that synthesis won't need artistic attention during such a change. But that work will only ever increase modestly, no matter the detail of the underlying model, because it'll only ever amount to hand-tweaking some filters to match timbre, requiring the same number of sound designers for a given car count and time target. With samples, the work increases with every sample you add, plus the need to balance and match all samples against each other - again, non-linear growth of effort.
In terms of budget, you get more for your money (in the long run) with synthesis. Sampling is effectively throwing your money away.
...wrong end of the train.
