AI GPUs probably live longer than three years (www.seangoedecke.com)

Recent discussions in the AI community have challenged the notion that inference GPUs have a lifespan of only three years under high utilization. This claim, primarily stemming from a tweet referencing an unnamed Google architect, suggests that as AI funding declines, the current GPU infrastructure will quickly become outdated, leading to unsustainable inference costs. However, anecdotal evidence and statements from several AI firms indicate that many GPUs, including Google's TPUs, have successfully operated for much longer, often exceeding six years with minimal failures. Historical data from supercomputers like Oak Ridge's Summit supports this, as a significant portion of their Nvidia V100 GPUs demonstrated high survival rates even after several years of consistent use. The implications of these findings are vital for the AI/ML community, particularly regarding economic longevity amidst a potential downturn in AI investments. While economic considerations might push providers to upgrade to more efficient GPUs like the B100, the evidence suggests that current GPUs can remain operational and profitable for an extended period. This challenges the narrative that AI inference will become prohibitively expensive once initial funding wanes. Instead, firms may continue to leverage existing infrastructure for profitability, utilizing older models while managing operational costs effectively. Thus, the long-term viability of GPU-based AI operations appears more secure than previously thought.