SemiAnalysis goodput-expense model, Large LLM Pretrain scenario (5,184 GB300 GPUs, $4/GPU-hr): rounded range of the Gold-tier, hyperscaler and Silver-tier outputs (6.14%, 10.53%, 20.91%)
6–21%modeled
| Value kind | modeled — Interpret this value according to its displayed kind, scope, source, and as-of date. |
|---|---|
| Scope | One training scenario priced for three provider tiers, each with an assumed fault-tolerance method (TorchPass-class code, HyperPod checkpointless training, checkpoint restart on hot spares). Goodput expense is a modeled cost uplift (extra rental time or reduced output), not a measured share of paid GPU-hours lost. These are the same outputs as the 6.1% / 10.5% / 20.9% triplet in Chapter 10.6. |
| As of | 2026-04-20 model; August 2025 pricing basis |
| Source | SemiAnalysis, How Much Do GPU Clusters Really Cost?, April 20, 2026, Scenario 1. · Definitions: Goodput Expense; Scenario 1: Large LLM Pretrain |
| Review | checking…review by 2026-12-08 · standard cadence |
| Recorded changes | last 2026-09-16 · 2 revisions tracked |
| Claim id | reliability-tax-share-of-paid-gpu-hours-lost-to |
Where the guide uses it
- 12.4 SLAs, Goodput Contracts & Availability Commitments
- 14.1 Operational KPIs, Goodput & the Reliability Economics of AI Factories
← Full numbers register — every date-stamped figure in the guide, with revision history.