In an interview timed to GTC Berlin on October 21, 2026, NVIDIA CEO Jensen Huang said the usable life of GPUs has stretched from the four years traditionally assumed to "probably eight years, or even longer." Yahoo Finance reported the remarks, which feed into a broader argument that older-generation GPUs retain ongoing demand and can serve as collateral for loans. The day before, on SpaceX's earnings call, Elon Musk explained that advances in AI models increase the value that can be extracted from existing GPUs. However, the period over which a GPU keeps generating revenue is not the same as the period over which it can support the latest software. It is also worth examining how much older GPUs actually benefit from software improvements.
What "8 years or more" means for GPU lifespan
In an interview with Deirdre Bosa, Huang explained that NVIDIA's computing systems are used for a wide range of purposes and can become more productive during their operating life through software improvements. He then said the lifecycle, long thought to be around four years, has now stretched to probably eight years or even longer. This is not the result of a test running a particular GPU for eight years, nor a lifespan guaranteed for every GPU. What Huang is describing is economic life: the period during which a GPU can keep doing work and generating revenue.
The broad pool of GPU renters also underpins his view. Even if one customer stops using a GPU, the asset does not lose its value when the original contract ends as long as another company needs it. Citing productivity, breadth of use, and the potential for long-term operation, Huang also raised the possibility of borrowing against the GPUs themselves. Whether economic life extends matters not only to companies that own GPUs but also to the lenders that finance their purchase.
NVIDIA's official blog likewise explains that not every workload requires the newest system, and that the right hardware depends on the nature and complexity of the task. The company points to the A100, which launched in 2020 and remains in commercial use six years later, with CoreWeave extending its reservations through 2029. Still, the period from a product's launch to the end of a contract cannot simply be treated as the operating life of the same GPU. Such cases show continuing demand for older-generation GPUs; they are not measured results for the average lifespan of GPUs overall.
The same GPU yields more "intelligence"
On SpaceX's third-quarter 2026 earnings call on October 20, Musk said that as AI software and models advance, the amount of "useful intelligence" obtained from the same power consumption increases. According to a Motley Foolは transcript published on Yahoo Finance, he also claimed that older GPUs such as the H100 are being rented out at higher prices than when they were new. The implication is that even as semiconductor generations turn over, the value of already-installed equipment does not necessarily decline at the same pace.
Even on the same GPU, advances in software optimization allow computing resources to be used more efficiently. And if model improvements raise the quality and usefulness of responses, value rises in ways that raw processing speed does not capture. NVIDIA's emphasis on continuously optimizing software and kernels for existing hardware is consistent with this thinking. Replacing GPUs with the latest ones is not the only way to raise the productivity of equipment.
That said, the "intelligence" Musk referred to is not a performance metric with a clearly defined measurement method. Comparing the efficiency of different AI models requires aligning response quality, latency, and input conditions, then testing how much power it takes to accomplish a given task. As for H100 rental prices, no specific contract rates or comparison conditions were given, so it cannot be concluded that prices are rising across the GPU rental market as a whole. Performance gains from software improvements and the rates renters actually pay should be evaluated separately.
Musk has also reflected this expectation of rising GPU value in his leasing policy. He indicated a preference for short-term contracts priced close to the spot market over fixing rates in long-term contracts, and said that rental rates per unit of power had risen when contracts came up for renewal. Behind this approach is his forecast that shortages of computing resources will continue and demand for older GPUs will hold. But future demand and rising rental prices are not guaranteed.
GPUs may last long, but software support has a deadline
In a developer explainer dated August 4, 2025, NVIDIA said that CUDA 13.0 ends support for offline compilation targeting GPUs with compute capability below 7.5. The affected architectures include Maxwell, Pascal, and Volta. This restriction concerns the step in which the CUDA compiler, NVCC, generates code for target GPUs before a program runs. It is not an announcement that support for the A100 or H100 is ending.
Nor do GPUs that fall out of support suddenly become unusable. According to NVIDIA, applications built with an older CUDA Toolkit can continue to run as long as the driver supports the hardware. For those who want to keep developing for the affected older GPUs, NVIDIA advises continuing to use CUDA Toolkit 12.9 and the 580-series drivers. The 580-series driver is supported until mid-2028, so continued use of older GPUs comes with constraints on development environments and drivers.
This distinction matters when evaluating the argument that software improvements extend GPU life. Being able to run existing applications is not the same as being able to build new ones with the latest development tools. Even if enough computing demand remains for an older GPU, whether the software you want to use runs in that environment must be checked case by case. The claim that GPUs can be used for a long time does not mean every generation will receive the benefits of the latest software indefinitely.
Lifespan is not the only thing that drives collateral value
Huang's suggestion of GPU-backed loans and Musk's short-term leasing policy share an expectation that existing GPUs will keep generating revenue. But the long-term residual value lenders rely on in lending decisions is a separate matter from the rental-rate increases short-term contracts aim to capture. Both depend on whether companies will still want to rent GPUs in the future and whether those GPUs can run the workloads required at that time. In the first place, "GPU lifespan" has several meanings: physical life, economic life, accounting useful life, and the period of development environment support do not necessarily coincide. Huang's "eight years or more" should also be understood as a view about economic life.
| Category | What period it judges | What can be read from the remarks and documents |
|---|---|---|
| Physical life | How long the equipment keeps operating normally | Huang's estimate is not a measurement of failure rates or the operating life of individual GPUs |
| Economic life | How long it keeps generating value that justifies operating costs | The 8-years-or-more estimate rests on software improvements and continuing demand |
| Accounting useful life | The period set for allocating acquisition cost across accounting periods | Longer GPU life or higher rental prices do not make depreciation unnecessary |
| Development environment support | How long development tools and drivers for the target GPU remain available | The CUDA 13.0 compilation restriction must be distinguished from continued execution of older applications |
This table organizes Huang's remarks and NVIDIA's developer documentation by the purpose for which GPU lifespan is being judged. These are not lifespan categories officially defined by NVIDIA, and the periods cannot be compared on a common basis. Moreover, even if economic life may extend, actual revenue depends not only on rental rates and utilization but also on power costs and maintenance expenses. Assessing a GPU's residual value requires asking not merely whether it keeps running, but whether it can continue to earn revenue that exceeds its operating costs.
What matters for companies running GPUs is whether the GPUs they hold meet the required processing quality and speed, and whether they can keep the supporting software in an environment that is still within its support window. Lenders, for their part, must judge whether borrowers meeting those conditions will continue to exist and whether sufficient rental income can be earned. If the effects of software improvements can be verified in actual processing performance and contract terms, decisions to operate older GPUs over the long term will gain concrete grounds beyond the "eight years or more" estimate.