“The performance slowdown from increasing AI model size has historically been offset by speed gains from more powerful chips.”