Google's AI Infrastructure Chief, Amin Vahdat, on the Physics & Economics of Frontier AI

Google treats power provisioning and real-world failure recovery, not raw chip FLOPS, as the true bottlenecks shaping its AI infrastructure strategy.

• Goodput over FLOPS: Vahdat says Google holds itself accountable to "goodput" — useful work delivered through real failures — not theoretical flops, since at 100,000-accelerator scale something fails multiple times an hour.
• TPU split into 8i/8t: Google released separate inference (8i) and training (8t) TPU chips in 2026 after projecting inference could reach 30-60% of workload, justifying specialization versus one general-purpose chip.

Themes, sentiment and reasons are AI-written paraphrases, checked so they never quote an episode. They may contain errors. Not investment advice. Episodes remain the property of their creators.