“Google's TPU-AT scales up to 9,600 TPUs per pod, a 2.8x improvement over Ironwood, delivering 121 exaflops of FP4 compute.”