[middleca]

Schneier's estimate assumes rented hardware time, but purchased hardware and multiple attempts would drive down the cost dramatically over time. This estimate also doesn't include optimizations by leveraging previous cycles or spends, or the impact of storing something like a rainbow table.

Given a government budget and time-scale, and ignoring sunk costs, I would think a collision attack could already be in a negligible cost range for any well funded sophisticated attacker that has setup shop.

edit: also that opinion post is 2 years old