What's this about?
Google's AI infrastructure chief covers accelerator failures, TPU specialization, co-design with DeepMind, optical switching, power and the extra CPU and storage demanded by agents.
What exactly is goodput?
Throughput is theoretical work per second; goodput is useful work actually delivered after failures, retries, checkpoint restoration and operational overhead.
Why is this discussion important?
full publisher transcript. It reframes the compute race around reliability and watts: at 100,000 accelerators failures can occur multiple times an hour, while autonomous loops remove the natural delay created by human users. Key sections: 05:30, 15:32, 23:30, 34:16, 37:50 and 42:35.