Via nvidia.com
Jensen Huang's dual message on GPU longevity is clever marketing, but supply and demand may have their own opinions
Nvidia is running two sales pitches at the same time, and they point in opposite directions. Buy the newest AI accelerators, CEO Jensen Huang says. But also: don’t worry about the chips you already bought, because those will hold their value for years.
The message from Huang #
On August 13, 2026, Huang posted on X that Nvidia’s A100 GPUs remain effective from 2020 through 2029, crediting the CUDA software ecosystem for extending the chips’ useful lives. CUDA, Nvidia’s proprietary parallel-computing platform, gets regular updates that can squeeze more performance from existing hardware, meaning a chip doesn’t become obsolete the moment a faster one ships.
The timing of the statement matters. Nvidia is simultaneously shipping its Blackwell GPU series, has already released the Hopper line (the H100 and H200), and has the Rubin architecture in its near-term pipeline, expected sometime in the 2026 to 2027 window. That’s three generations of product stacked up at roughly the same moment, which creates an obvious question for anyone sitting on older hardware: should I upgrade, or is what I have still good enough?
CoreWeave, the AI cloud infrastructure company, gave Nvidia a useful data point to anchor the argument. The company has a rental contract for A100 GPUs running through 2029 at pricing it considers attractive, which publicly demonstrates that someone with sophisticated procurement capabilities still sees multi-year commercial value in hardware that first shipped in 2020.
The tension underneath the pitch #
Analysts watching the company in late August 2026 flagged exactly this tension: Nvidia wants to maximize sales of next-generation chips while simultaneously telling existing customers not to panic about depreciation.
So far, the scorecard looks decent for Nvidia. Secondary-market pricing for both A100 and H100 chips has remained strong, supported by sustained demand for inference workloads specifically. The AI compute market is shifting from a world dominated by large-scale model training, which requires the absolute newest and most powerful hardware, toward inference, which means running already-trained models to produce outputs. Inference workloads are less demanding at the cutting edge and can often be handled efficiently by older GPUs, which creates genuine demand for chips like the A100.
What it means for buyers and the broader market #
CoreWeave’s A100 contract is the clearest illustration of how this plays out in practice. A hyperscaler-style customer locked in multi-year GPU access at pricing it found favorable, which suggests the market agreed on a value for older hardware that made a long-term rental economically sensible.
Strong used-market pricing for A100 and H100 chips reflects that customers are actively using and monetizing these assets, which supports the broader thesis that AI compute demand is deep enough to absorb multiple hardware generations simultaneously.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our