What's your opinion of high bandwidth flash? It seems ideal for storing and retrieving model weights. Do you see that dramatically changing inference chip design over the next couple of years?