“The disaggregation of prefill and decode workloads in AI inference will extend the useful life of GPUs to 10 or 15 years.”