Nvidia chief executive Jensen Huang is asking financiers to treat the company’s AI accelerators as long-life assets rather than fast-depreciating parts. In an interview with Yahoo Finance’s Deirdre Bosa at Nvidia’s GTC gathering in Berlin, he said the systems stay productive for eight years or more and that lending against them is less risky than lending against anything else. The pitch lands as buyers and their backers search for ways to finance ever-larger compute bills.
- Key Point: Huang says Nvidia accelerators keep earning for at least eight years, well past the four-year replacement cycle commonly applied to GPUs.
- Key Point: He argues the hardware is unusually safe security for a loan because it holds residual value across many owners and many types of work.
- Key Point: Elon Musk has made a related case, saying older accelerators can beat newer silicon on inference economics when measured per watt of electricity.
- Key Point: B200 units are reportedly rented out or traded at 158% of their original price, though condition, supply and workload shape second-hand values.
A depreciation clock that runs past four years
The four-year working life long attached to graphics processors no longer describes what happens to a data-center system, according to Huang. He reached for three words to describe the accelerators — “productive,” “fungible” and “durable” — and pointed to ongoing software work as one reason the hardware keeps improving after installation.
Raw performance, the ability to take on a wide range of jobs and repeated software gains all stretch the period in which the chips earn money, he said. On that view, investors and lenders are using the wrong yardstick, valuing a capital asset as though it were consumer electronics on a short replacement cycle.
The case for compute as collateral
Huang’s central claim is about resale value. Because one buyer’s finished workload is another buyer’s starting point, Nvidia hardware keeps a large share of its worth when it changes hands, which in his view makes it dependable security against a loan. Financiers, he suggested, have not yet caught up with how adaptable the equipment is.
On lending against Nvidia hardware, Huang put the risk at “the lowest of any asset.”
Elon Musk has arrived at a similar place from a different direction. His argument is that older accelerators can out-earn new ones on inference work, because what counts is how much model output a machine produces per watt of electricity rather than its peak throughput. Musk has cited the H100, which has reportedly been worth more in the market than it was at launch, as models and inference techniques coax better results from unchanged silicon.
Resale data, and what it leaves out
Early signals from the secondary market support the thesis. The A100 and H100 have held more of their value than much of the industry expected, and the newer B200 has been even more striking: units are said to be rented or traded at 158% of what they originally cost.
Prices alone remain an incomplete guide. People who trade the equipment say a used accelerator’s worth hinges on its condition, on how many comparable units are available, on how the surrounding server is configured and, above all, on the workloads the next owner intends to run.
What follows depends on whether lenders adopt Huang’s framing. If they do, accelerators could shift from an operating cost that buyers write off quickly to an asset that supports borrowing, changing how AI build-outs are funded. If they do not, the distance between Nvidia’s view of its hardware’s durability and the depreciation schedules kept by its customers will stay a quiet dispute over who absorbs the risk as the equipment ages.