The jury is still out on this. Those tax based deprecation schedules are largely a relic of traditional data centers, where workloads are fairly moderate compared to AI use cases. Additionally, power and rack space constraints can complicate things quite a bit. If next gen chips are significantly more efficient and you are currently constrained by power availability, you might pull your old servers and replace them with the newer ones regardless of how much useful life you have left.