Training requires a ridiculous amount of both, but again the raw compute requirements would be much lower with huge memory pools.
