Nvidia's 2020 A100 Chip Contracts Extended to 2029, Reigniting Debate Over Old GPU Lifespan
Nashnova编辑部
CoreWeave extended an A100 GPU lease to 2029 — nearly a decade after the chip launched — directly challenging the short-seller thesis that old GPUs become worthless in two to three years.
Why does one lease deal shake the entire AI investment narrative?
CoreWeave CFO Nitin Agrawal told analysts the company signed a deal extending A100 rentals to 2029, adding that older-generation Nvidia chips are "essentially sold out."
This means → a chip released in 2020 still commands fresh lease revenue, putting its economic life far beyond the two-to-three-year window short sellers assumed.
In plain terms = bears said old GPUs turn to scrap fast; a real-money contract says otherwise.
What was the short-seller thesis — and where did it break?
The bearish argument: Nvidia iterates too fast → old chips lose economic value in two to three years → companies must accelerate depreciation → profits get squeezed hard.
GPU-price tracker Silicon Data offered counter-evidence: A100 lease prices rebounded strongly in 2026 and have held steady since.
The firm wrote publicly: "The answer is clearly not the two-to-three years some casually assumed." This reflects real lease-market pricing overturning paper depreciation assumptions.
Why can old chips still make money?
Crusoe SVP Erwan Menard identified the key mechanism: GPUs can migrate from one workload type to another as they age.
In plain terms = the newest chips train cutting-edge models; once training is done, inference (making the model answer questions) and fine-tuning (small adjustments) need far less compute — old GPUs handle these jobs just fine.
Lambda executive Matt Rowe added: effective GPU lifespan can reach seven to eight years; five-year warranty contracts mean failed chips get replaced, extending the entire fleet's useful life.
Is a $500 billion bet riding on this very assumption?
Nvidia announced partnerships this week with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, targeting over $500 billion in AI infrastructure investment over time.
This means → the return logic behind that massive capital pool partly rests on the premise that AI hardware holds economic value for years, not quarters.
The A100 attracting new contracts nearly a decade after launch is the most direct stress test of that premise to date — if old chips really died in two to three years, the $500 billion return model would need rewriting.
Content is for reference only, not financial advice.