Once they implement their models directly in silicon, the cost will come down and the speed will go up. See Taalas.