By burning the transformer architecture into our chips, we can run AI models an order of magnitude faster and cheaper than GPUs.