Pretraining
The initial, compute-heavy phase that trains a model on vast unlabeled data to learn general capabilities.
Current numbers
6.1% / 10.5% / 20.9% goodput expensemodelled for one large-pretraining scenario (SemiAnalysis, April 2026): Gold-tier, hyperscaler and Silver-tier providers; the loss observability exists to shrink
6–21%Modeled goodput expense, one 5,184-GB300 pretraining scenario: 6.1% Gold tier to 20.9% Silver tier (hyperscaler 10.5%)
6.14 / 10.53 / 20.91%ClusterMAX April 2026 large-pretraining scenario: gold / hyperscaler / silver goodput expense