> (accumulating GPUs is not trivial) Also like accumulating milk rather than gold. A approximate failure rate I have been able to find is: https://www.tomshar

> (accumulating GPUs is not trivial) Also like accumulating milk rather than gold. A approximate failure rate I have been able to find is: https://www.tomshardware.com/tech-industry/artificial-intelligence/faulty-nvidia-h100-gpus-and-hbm3-memory-caused-half-of-the-failures-during-llama-3-training-one-failure-every-three-hours-for-metas-16384-gpu-training-cluster https://www.tomshardware.com/tech-industry/artificial-intell... Which would imply 42% loss over 5 years if linear attrition at 10% failure rate per gpu year. (((1-(148+72+19+17+6)÷16 384)^(1÷54))^(365) Dunno how many faults are repairable in practice. The failure rates of CPUs were alot lower for comparison.