Power, hardware failures and specialized TPUs in Google's AI buildout

A post about Google's AI infrastructure buildout says Google forecasts $200 billion in 2026 capital spending. It describes “goodput” per watt as a key metric and says something fails multiple times an hour at a scale of 100,000 accelerators. It also discusses separate TPUs for training and inference, rising demand from long-horizon agents and Google's preference for grid-connected power.