I'd assume the NSA's fabrication capability is more on the scale of the pilot plants fabs build at each new process scale. Some universities certainly have fabrication equipment testbeds as well, so the NSA effort may be more that modest scale.

If I were tasked with the problems the NSA faces, I think I'd at least focus in on:

1. CMOS reverse engineering equipment that can shave down dies, image and analyze the structures, etc.

2. Small scale fabrication for extremely sensitive infrastructure. These roles probably aren't performance critical. Eg if you have some microcontroller that plays a role in say nuclear weapon arming protocols, you need that to be pretty much beyond suspicion.

3. Some way of sampling commodity parts for unexpected behavior non-destructively. If this could be done efficiently enough, you could use it in combination with #1 to get reasonable confidence for off the shelf parts.

One thing I'd suspect is that if the NSA did find highly targeted flaws they probably wouldn't disseminate that fact unless absolutely necessary. Keep an adversary using a strategy you know rather than provoking improvement.

Personally I doubt the NSA forces backdoors into commodity chips. In theory there might be some way of introducing a flaw that would cripple specific large computations like crypto-analysis of a particular code, or biasing a particular random number generator. But that just seems too likely to backfire.

I'd always thought it was interesting that the pentium FDIV bug was most easily found by code calculating twin primes. But there may be a mundane explanation for that rather than cloak and dagger stuff.