[selfhoster11]

If 20B reasoning models are the goal, we can do that a whole lot cheaper than for $1T.