Why AI GPUs Are Becoming Long-Life Assets Instead of Depreciating Hardware
A major shift is happening in the AI hardware market: high-end GPUs are no longer behaving like traditional tech equipment that steadily loses value over time. Instead, the most in-demand AI accelerators are holding their value, and in some cases, becoming more valuable after launch.
That view is now being echoed by some of the most influential voices in technology and investing, including NVIDIA CEO Jensen Huang, Elon Musk, and semiconductor investor Gavin Baker. Their shared message is clear: GPUs used for artificial intelligence are becoming durable, revenue-generating assets with a longer productive life than many previously expected.
The reason comes down to one powerful trend: AI models are getting better, and as they improve, the same GPU can produce more useful output per watt of energy consumed.
Elon Musk recently explained that the value of compute per watt is likely to rise over time rather than fall. As AI software becomes more efficient and models become more capable, each unit of power spent on compute can generate more useful intelligence. In that environment, the hardware running those models becomes more valuable, not less.
This helps explain why older AI GPUs such as NVIDIA’s H100 can still rent at strong prices, sometimes even higher than when they were new. In a conventional hardware cycle, older chips would typically face declining demand as newer products arrive. But AI compute is not following that old pattern.
Jensen Huang has made a similar point, saying that NVIDIA GPUs can remain productive for eight years or more. One reason is that NVIDIA’s hardware is widely used across the AI ecosystem, but another key factor is software. Because NVIDIA’s systems are programmable and receive ongoing software improvements, their performance and usefulness can continue to improve long after the hardware is deployed.
That changes how data centers, cloud providers, and AI companies think about GPU lifecycles. What was once viewed as a four-year asset may now be productive for twice as long, or potentially even longer. These chips are also highly liquid in the sense that if one company no longer needs them, another company often does. Demand exists across many levels of the market, from top-tier AI labs to smaller businesses looking for affordable inference capacity.
The economics of AI inference are playing a major role in this shift. Inference, the process of running AI models to generate responses, images, code, analysis, or other outputs, is becoming one of the biggest drivers of GPU demand. As businesses integrate AI into more products and services, the need for reliable compute continues to rise.
Jensen Huang has said that 1 gigawatt of AI compute infrastructure can cost around $50 billion to build, while generating roughly $40 billion in annual revenue. That implies a payback period of a little over one year, an unusually fast return for infrastructure of that scale.
Musk has also pointed to similar market conditions, noting that AI compute can be valued at tens of billions of dollars per gigawatt, with payback periods of less than a year in some cases. He also suggested that short-term compute rental agreements are attractive because they can renew at higher prices as demand increases.
This is one reason GPU rental pricing has strengthened even as the cost of individual AI tokens has declined. As models become more efficient, each GPU hour can generate more useful tokens or outputs. In other words, the customer may pay less per token, but the GPU itself becomes more productive per hour, making that hour of compute more valuable.
Gavin Baker has described this dynamic as a major reason GPU rental prices have surged. When a single GPU hour can produce more useful AI output, demand for that hour increases. That supports stronger pricing and changes the traditional assumptions about hardware depreciation.
Recent market data also supports this view. Older NVIDIA A100 and H100 GPUs are reportedly holding residual values far above what standard three-year or five-year depreciation schedules would suggest. Even more striking, NVIDIA’s B200 GPUs have been reported to carry a residual value of about 158 percent of their original cost, meaning they may trade at a 58 percent premium to launch pricing.
That kind of pricing behavior is rare in enterprise hardware. Servers, processors, storage systems, and networking equipment usually lose value as newer generations arrive. But AI GPUs are benefiting from a supply-demand imbalance, rapid AI adoption, and the growing usefulness of AI models.
The bigger idea is simple: when a company retires a GPU from its own fleet, that chip does not become worthless. It moves into the secondary market, where another buyer or renter can use it for a different workload. A cutting-edge AI lab may upgrade to the newest GPU generation, but the previous generation can still be valuable for inference, fine-tuning, research, enterprise AI tools, or smaller-scale model deployment.
This creates a layered market where GPUs continue producing economic value at different stages of their lifespan. The most advanced companies may want the newest chips, while other customers are eager to use proven hardware at lower cost. That makes AI GPUs more durable, more fungible, and more financially attractive than many traditional computing assets.
For investors, data center operators, and AI companies, this trend could reshape how hardware spending is evaluated. If GPUs remain productive for eight years or more, and if resale values stay strong, the economics of AI infrastructure look very different from the older model of rapid depreciation.
The AI boom has already transformed demand for advanced chips. Now it may be transforming the very definition of a GPU asset. Instead of being seen as equipment that steadily loses value, high-performance AI GPUs are increasingly viewed as long-life infrastructure capable of generating revenue across multiple owners, markets, and use cases.
As AI adoption grows, the value of compute may depend less on the age of the hardware and more on what that hardware can produce. And as long as GPUs keep delivering useful intelligence efficiently, their role in the AI economy is likely to remain both powerful and profitable.
