The question everyone in AI is asking: How long before a GPU depreciates?

As some of the world's leading corporations plan to invest a staggering $1 trillion over the next five years into artificial intelligence data centers, one critical issue is capturing the attention of both executives and investors: depreciation. In the realm of finance, depreciation refers to the allocation of the cost of tangible assets throughout their anticipated useful life. This concept has become increasingly significant in the technology sector, particularly as firms assess the longevity and value retention of the vast number of Nvidia graphics processing units (GPUs) they are acquiring. Major players such as Google, Oracle, and Microsoft have projected that their servers could be operational for up to six years. However, there are concerns that they may lose value much sooner. Microsoft's recent annual report indicated that its computer hardware has a lifespan ranging from two to six years. This uncertainty poses a considerable challenge for the investors and lenders funding these massive AI infrastructure expansions, as the longer the equipment remains valuable, the more manageable the depreciation impact on profits. The challenge with AI GPUs is their relative novelty; Nvidia's earliest AI-targeted processors for data centers were introduced around 2018, while the current surge in AI interest began with the launch of ChatGPT in late 2022. Since that pivotal moment, Nvidia's data center revenues have skyrocketed from $15 billion to an impressive $115 billion by January. Industry expert Haim Zaltzman, vice chair of Latham & Watkins' emerging companies and growth practice, noted the lack of historical data on GPU longevity compared to other established heavy machinery. "Is it three years, is it five, or is it seven?" he posed, highlighting the significant implications for financing outcomes. Some Nvidia clients, however, are optimistic about the longevity of AI chips, suggesting that older processors will continue to have value for various tasks. CoreWeave, a company that purchases GPUs to lease them to customers, has adopted a six-year depreciation cycle since 2023. CEO Michael Intrator shared insights following a recent earnings report, indicating that they are relying on data to make informed decisions about GPU lifespan. He noted that CoreWeave's Nvidia A100 chips, launched in 2020, are fully booked, and a recent availability of Nvidia H100 chips from 2022 saw them rented out at 95% of their original price. Despite this optimism, CoreWeave's stock fell 16% after the earnings announcement, attributed to delays from a third-party data center developer affecting their full-year outlook. The share price has plummeted 57% since June, reflecting broader concerns regarding AI-related spending. Similarly, Oracle's stock has decreased by 34% from its peak in September. Skepticism about the AI market is growing, with notable short seller Michael Burry recently revealing his positions against Nvidia and Palantir. He has expressed concerns that companies like Meta, Oracle, Microsoft, Google, and Amazon may be overstating the useful life of their AI chips while downplaying depreciation, suggesting a more realistic lifespan of two to three years for server equipment. AI chips face the risk of rapid depreciation due to wear and tear or the emergence of newer models. While they may still be capable of handling certain workloads, the economic viability could diminish significantly. Nvidia CEO Jensen Huang hinted at this reality during a recent event, joking that the release of the new Blackwell chip might render its predecessor, the Hopper, nearly worthless. In February, Amazon revised its useful life estimates for a segment of its servers from six years to five, citing rapid advancements in AI technology as a reason for the change. On the other hand, some hyperscalers are extending their GPU lifespan projections for new server equipment. Despite Microsoft's aggressive AI infrastructure plans, CEO Satya Nadella emphasized the importance of pacing their GPU acquisitions to avoid overcommitting to a single generation. Dustin Madsen, vice president of the Society of Depreciation Professionals, explained that depreciation is ultimately a management estimate, subject to change with developments in the fast-evolving tech landscape. These estimates account for factors like technological obsolescence and historical lifespans, and they must be substantiated to auditors who will review engineering data validating the projected lifespan. As the AI landscape continues to evolve, the question of GPU value retention remains a critical topic for investors and companies alike.