Can a datacenter practicably upgrade to the next generation of GPU every year?