I suspect the hypothetical benefit lies in iterating new models that drastically reduce the cost of a task of this calibre: something that, frankly, has already happened with other tasks. We spend 20 million dollars on NS today so that, by, say, 2030, a problem of this level costs $20.000,00.