A one-size-fits-all GPU refresh cycle simply does not exist. When we look at frontier training fleets, their lifespans are typically planned for 2-3 years. Conversely, production inference fleets are designed to operate for 4-5 years or even longer. As a result, any disagreement over the useful life of a GPU is fundamentally just a debate about which specific fleet you are evaluating.