If I were a somewhat different person, and life played out quite differently, I can imagine being a senior NSA bureaucrat.
I'd read about Google's cat-recognition engine, and the likelihood of a mouse-level AI within the decade, and I'd be thinking that the NSA needs to get there first.
Not because there's an obvious military application, but because there could be a weapon in there somewhere and because someone in China is thinking the same thing.
So I'd add a few hundred million a year to my off-budget budget; just pocket change really. Then I'd build a data center for my testing and I'd get to a mouse level AI in six years. If I needed to I could pry the secret sauce out of Google's hands; I'm sure there a ways to do that but it's probably not necessary. Google publishes much of its AI research.
I'd build it just to see what it was like, and so I could assess the military potential.
Problem is, modern AI experiments take a lot of power and produce a lot of heat. I wonder how I'd disguise it...