The traditional assumption that advanced computing hardware rapidly loses its usefulness is being challenged by the expanding AI industry. GPUs that were once expected to have a productive lifespan of roughly four years could remain in service for eight years or longer, according to Huang’s recent remarks. Musk, meanwhile, argues that improvements in AI software can increase the economic value of existing computing infrastructure by enabling it to produce more useful results for every watt of electricity consumed. Source: Wccftech
The implications could be significant for cloud computing companies, AI developers and data-centre operators that have committed billions of dollars to specialised hardware. If older GPUs continue generating revenue for longer, operators may be able to spread their initial investments across more years of productive use, potentially improving returns on expensive infrastructure.
Speaking during SpaceX’s third-quarter 2026 earnings call, Musk explained why he believes the value of computing capacity could rise rather than decline as AI technology advances. As models become more capable and software becomes more efficient, the same hardware can potentially deliver greater quantities of useful intelligence without requiring a corresponding increase in electricity consumption.
“The value of compute per watt will increase over time, not decrease,” Musk said, explaining that better AI models can extract more useful output from existing processing resources. He also pointed to the rental market for NVIDIA’s older H100 accelerators, saying they were commanding higher rental rates than when they were new. Source: SpaceX Q3 2026 earnings-call transcript
Musk’s argument rests on a distinction between the age of a computer chip and the economic value of the work it can perform. Although newer processors generally offer improvements in performance, memory capacity, interconnect technology and energy efficiency, older hardware can remain attractive when it is capable of handling commercially useful workloads at competitive prices.
For example, not every AI application requires the newest generation of accelerators. Many inference workloads, which involve running trained AI models rather than building them from scratch, can operate effectively on existing hardware. Businesses may therefore continue renting older GPUs if the cost of running their applications remains competitive and the hardware meets their performance requirements.
The central idea is that a GPU’s value depends not only on its technical specifications but also on how much useful computing work it can deliver relative to its operating cost.
Musk also described a market in which demand for AI computing resources remains exceptionally strong. SpaceX intends to allocate its computing capacity between internal operations and external customers, with a roughly equal split serving as an initial guideline rather than a fixed commitment. The company is using flexible, short-term rental arrangements to respond to changing demand and capture prevailing market prices.
According to Musk, companies ranging from major cloud providers to smaller AI businesses have been seeking access to SpaceX’s computing infrastructure. He said the company could rent out its capacity multiple times over if it agreed to every request. This demand, combined with constrained availability of advanced processors, supports his view that computing resources could remain valuable for an extended period. Source: SpaceX Q3 2026 earnings-call transcript
Huang’s assessment of hardware longevity complements this argument. If GPUs can remain productive for eight years or more, data-centre operators may not need to replace every generation of hardware as quickly as earlier industry assumptions suggested. Older processors could continue serving customers even after newer, faster systems become available.
The shift could also change how companies evaluate capital expenditure. AI infrastructure is expensive to build, with operators investing in accelerators, servers, networking equipment, cooling systems and electricity capacity. Extending the period over which these assets generate revenue can improve the economics of a data centre, provided maintenance costs remain manageable and customers continue to demand its computing services.
Baker, an investor known for his focus on semiconductor and computing infrastructure businesses, has also discussed the rising economic value of GPU capacity. The argument is that lower costs for AI-generated tokens do not necessarily mean lower revenues for computing providers. If improvements in AI systems allow each GPU-hour to generate substantially more useful output, the commercial value of that hour can rise even as the price of individual tokens falls. Source: Wccftech
However, the claim that GPUs are no longer depreciating assets requires an important qualification. A longer useful life does not mean a GPU has stopped losing value. Hardware can remain operational and generate revenue while its resale price declines, its performance falls behind newer products or its operating costs become less competitive.
Accounting depreciation, resale value and commercial usefulness are also different measures. A company may spread the cost of a server over several years for accounting purposes, but that does not guarantee that the equipment will retain its original market value throughout that period. Similarly, a GPU that remains useful for inference may no longer be suitable for the most demanding AI training workloads.
Energy consumption is another consideration. Newer accelerators may deliver more computing performance per watt, making them economically preferable for workloads that require maximum throughput or operate at enormous scale. Older hardware could also face pressure if AI software increasingly relies on features, memory capacities or networking capabilities that previous generations cannot provide.
There is an additional risk that today’s shortage of computing capacity may eventually ease. If major technology companies build enough data centres and new generations of processors become more widely available, rental prices for older GPUs could decline. Hardware longevity alone cannot guarantee strong returns if supply grows faster than demand.
Nevertheless, the current argument offers an important insight into the future of AI infrastructure. The value of computing hardware increasingly depends on the combination of silicon, software efficiency, electricity costs and market demand—not simply on the year in which a chip was released.
For NVIDIA and other hardware suppliers, continued demand for older accelerators could help sustain a broader market for AI computing. For cloud operators, it could create opportunities to extend asset utilisation and generate revenue from equipment that might otherwise be replaced. For AI developers, a wider pool of usable hardware could provide additional flexibility as they scale their applications.
The longer-term outcome will depend on whether demand for AI computing continues to expand and whether software improvements can keep older processors commercially relevant. Huang’s comments suggest that GPUs may have a longer productive lifespan than the industry once assumed, while Musk’s observations illustrate how software advances can increase the value of existing infrastructure.
The emerging possibility is not that GPUs have become immune to depreciation, but that their commercial usefulness may endure far longer than their technological generation would suggest. In an industry where demand for computing power continues to grow, yesterday’s hardware may remain an important part of tomorrow’s AI infrastructure.
