An important detail is that this is not 130 petaflops of double precision floating point. Their target is >130 single or even half precision.