I think the race is actually between locally optimized rigs with extremely strict context management, and the rest.
Those rigs are running one of the open Chinese models though, right?