The Extended Useful Life of AI GPUs

Carte mentaleMind map

The author challenges the traditionally short (4-5 years) useful life of GPUs, especially for AI. Recent evidence from Coreweave and Jensen Huang (NVIDIA) suggests that chips like the A100, introduced in 2020, can be re-contracted and profitably used until 2029, or even beyond. This extension of useful life is made possible by NVIDIA's CUDA platform, which allows for continuous software upgrades.

This trend is explained by two main factors: persistent computing constraints and the diversification of inference workloads. While the most complex inference tasks will always require the latest chips, a "long tail" of more trivial requests can be efficiently handled by older GPU generations. This opens up significant opportunities for companies capable of developing intelligent software layers to route requests to the most appropriate GPU, thereby maximizing the profitability of existing infrastructures.