I mean I think I read their spec sheet that it was 2.5kw on the asic but closer to 25kw with GPUs, for the same number of tokens (different latencies of course), but there's no way that a dedicated from-scratch matmul+memory chip can be less efficient than a GPU.