Seems like in a mature implementation your phone/laptop would route the message to a GPU farm and the work requirement could be scaled appropriately. And for whitelisted everyday contacts the computation cost could be zero.