A preprint tests ShardMeter, a lightweight model for forecasting distributed AI training runtime, throughput, bottlenecks and cost.