“The high failure rate of GPUs in data centers is a deliberate design choice to maximize performance by running them at higher power and frequency, with software managing the resulting instability.”