I'm no chip designer, but maybe a 'small' speculation cache which allows quick moving to the real caches might be better just in terms of less cache needs. If there's not enough space then you can't speculate farther, and that's that.

Of course, that'd probably need more complex logic to manage this new cache, which makes things more difficult. Then again, not sure how two caches interacts with potential speculation across multiple branches (does Intel even do that?)