how do you justify the compute investment for something like nemotron ? especially if all the labs are willing to pay for those same GPU clusters for inference or training runs?