Recently, at the 2026 Berlin GTC conference, NVIDIA CEO Jensen Huang presented a new perspective: the effective economic lifespan of data center GPUs for AI could reach eight years or even longer. Traditionally, the industry has assumed an update and depreciation cycle of about four years for AI accelerator cards. Huang stated that continuous software optimization through CUDA and rising demand for AI inference workloads can extend the actual usage value of hardware. Even without changes to the chip hardware itself, older GPUs can continue to generate commercial revenue through software updates and should not be quickly phased out or scrapped.
Elon Musk expressed a similar yet differently angled view during the earnings call. He believes that as large models and software algorithms continue to iterate, older GPUs can deliver higher "intelligence-per-watt." This means that with newer and better AI models, the same old hardware can produce more effective AI output per unit of power consumption. He cited older accelerator cards like the H100, whose current rental prices are even higher than when they first launched, as surging demand for computing power has boosted the commercial value of legacy hardware.
Note that both parties are discussing data center AI computing GPUs, not consumer-grade gaming graphics cards. This does not mean the hardware is immune to physical aging; rather, it refers to the usable lifecycle from a commercial and economic perspective. GPUs still undergo depreciation, but at a slower rate than previously estimated by the market. While hardware specifications remain fixed, ongoing software and model optimizations can continuously unlock the computational potential of older hardware.
Some Wall Street analysts remain cautious, warning that once computing power supply exceeds demand, the rental value of older GPUs will decline rapidly.