0: Unlike nuclear weapons, AI capabilities can be used without immediately angering the public or damaging the enemy.

1: In fact, the most effective doctrine of AI weaponry is to not to use it at all against the enemy, but only for its own development. Besides the recursive effect, using it against the enemy spoils your current capabilities which weakens your deterrence and potentially makes it easier for the enemy to catch up.

2: The amount of nuclear weapons can be hidden for sufficiently large mainland if the government is well-funded.

3: Compute generates heat, requires electricity, and requires not only for heat capture and electricity but also for the compute itself a large industrial base, which has to, additionally, for security reasons, to be somewhat self-enclosed once waste chips begin to accumulate. 

4: They can be put underground, but this makes sense if you can also hide all input and output traffick from satellite intelligence. They can be distributed approximately at less than 50% ± 25% overall performance decrease, but unless the distribution is extremely effecient camouflage, you have to make much more units of compute ane infastructure, which can be counterproductive to hiding.

5: The cost of hiding, both in time and money, easily becomes prohibtive in an arms race against adversarial nation-states, due to the time cost of not having the best AI capabilities as soon as possible; cost of hiding in addition to above reasons when having to be done as soon as possible increases much faster than when constructing non-time-penalized projects.

6: Therefore it makes no sense – under the presumed unaforelayed premises – to hide the compute.

Edit log from earliest to most recent:

Fixed note indexing and added the three separators and everything below them because user "(hidden)" pointed out that "Format note: your list is missing a number 3.", which is correct. For reference, I separately also note that the indexing of notes in above text was conventional rather than syllogistic.

In the previous edit log entry, user "(hidden)" was actually merely subjectively hidden being actually user "Gyrodiot". For reference, I also note that I prefer avoiding nested edit logs in my LessWrong "posts."