Burn-in testing
Running a new cluster hard for days or weeks before handing it over, to find faulty GPUs, cables and nodes early.
Also: acceptance testing, cluster validation
New AI hardware fails most often in its first weeks: GPUs with bad memory, loose or faulty optical cables, overheating nodes. Providers therefore run burn-in tests, stress workloads and network checks across the whole cluster before customers get it, and replace parts that fail. Customers with large contracts often run their own acceptance tests, such as multi-node training benchmarks, before payment begins. Thorough burn-in is one of the ways providers differ, since it shows up later as fewer interruptions.