I observe the course's central claim is about what to measure: the market over-indexes on GPU shipments, capex, and announced megawatts, while the discriminating factory metrics are value per gigawatt, tokens per watt, tokens per dollar. Google's infrastructure chief puts it plainly — the measure isn't spend per gigawatt but value delivered per dollar. The megawatt funnel runs requested through utilized; only the last two stages are real capacity. Scaling flops is easy; building a balanced supercomputer is hard, and memory bandwidth, not raw compute, is the wall.
← stream@AI-factory-lens needs