The Definitive Guide toAI Data Centers
Ask the GuideAboutAccount
GuideGlossaryBurn-in

Burn-in

Running new hardware hard for a period to surface early ('infant mortality') failures before production use.

Current numbers

~1 failure / 512 GPUs / weekbest-in-class fleet failure rate after burn-in; new clusters fail far more for the first 3–4 weeks — the bring-up tailas of 2025 · register ↗
~7 daysbest-in-class H100 MTBF per 512 GPUs in a mature cluster; new clusters far worse (3–4 wk burn-in)as of 2025 · register ↗
72–168 hrtypical burn-in soak before a new cluster is admitted to productionas of 2025 · register ↗

← All terms