In finance, depreciation refers to the method of distributing the cost of a tangible asset over its anticipated useful lifespan. This concept is gaining significance in the technology sector as companies estimate how long the numerous Nvidia graphics processing units, or GPUs, will remain advantageous or maintain their worth.
Infrastructure leaders such as Google, Oracle, and Microsoft have indicated their servers might be applicable for up to six years. However, they could also depreciate at a much faster rate. Microsoft’s latest annual filing noted that its computer equipment has a lifespan ranging from two to six years.
This presents significant considerations for the investors and lenders backing the massive AI expansions, as the longer equipment retains its value, the more opportunities a company has to prolong depreciation, minimizing its negative impact on profits.
AI GPUs pose a unique challenge as they are relatively novel in the marketplace. Nvidia’s initial AI-oriented processors for data centers were launched around 2018. The current surge in AI began with the introduction of ChatGPT in late 2022. Since that time, Nvidia’s yearly data center revenue has surged from $15 billion to $115 billion for the fiscal year concluding in January.
There is no established precedent for the lifespan of GPUs when juxtaposed with other forms of heavy machinery that industries have relied on for years, stated Haim Zaltzman, vice chair of Latham & Watkins’ emerging companies and growth sector.
“Is it three years, is it five, or is it seven?” asked Zaltzman, who focuses on GPU financings, during an interview. “It’s a significant distinction in terms of financing success.”
Some of Nvidia’s clients assert that AI chips will maintain their value for an extended period and that customers will continue to invest in older processors given their ongoing utility for various tasks. CoreWeave, which procures GPUs and rents them to customers, has implemented six-year depreciation schedules for its infrastructure since 2023.
CoreWeave CEO Michael Intrator informed CNBC this week, post quarterly earnings, that his firm is taking a “data-driven” approach regarding GPU lifespan.
Intrator noted that CoreWeave’s Nvidia A100 chips, introduced in 2020, are all fully subscribed. Additionally, a batch of Nvidia H100 chips from 2022 became accessible due to a contract termination, and were promptly reserved at 95% of their initial cost.
“Every piece of data I’m receiving suggests that the infrastructure holds its value,” Intrator stated.
Nonetheless, CoreWeave experienced a 16% drop in shares following its earnings report due to setbacks at a third-party data center developer impacting full-year projections. The stock has declined 57% from its peak in June, part of a broader market sell-off that reflects trepidation about overspending in AI. Oracle’s stock has fallen 34% from its highest mark in September.
Among the most vocal critics of the AI market is short seller Michael Burry, who recently revealed positions against Nvidia and Palantir.
Burry suggested this week that corporations including Meta, Oracle, Microsoft, Google, and Amazon are exaggerating the lifespan of their AI chips and underreporting depreciation. He estimates that the actual usable life of server hardware is around two to three years, indicating companies are inflating their earnings in consequence.
Amazon and Microsoft opted not to comment. Meta, Google, and Oracle have yet to respond to requests for comments.
‘You couldn’t give Hoppers away’
There are various scenarios where AI chips may lose value before reaching six years. They could become inoperative and fail, or they might be rendered obsolete as newer GPUs are introduced. They could still serve certain workloads, although the economic viability could be significantly reduced.
Nvidia CEO Jensen Huang has suggested as much. When Nvidia unveiled a new Blackwell chip earlier this year, he remarked humorously that the worth of its predecessor, the Hopper, would diminish.
“Once Blackwell begins to ship in significant quantities, you couldn’t even give Hoppers away,” Huang stated in March during Nvidia’s AI conference.
“There are conditions where Hopper is acceptable,” he added. “But few.”
Nvidia now releases new AI chips annually, changing from its previous two-year schedule. Advanced Micro Devices, its nearest competitor in GPUs, has likewise adopted this approach.
Nvidia will report its quarterly results next week.
Amazon, in a February filing, indicated it has reduced the useful life for a segment of its servers from six years to five years due to findings from a study highlighting “an accelerated pace of technological advancement, especially in artificial intelligence and machine learning.”
At the same time, other major data centers are increasing their GPU useful life estimates for newer server models
“One of the crucial lessons we learned even with Nvidia is that their rate of innovation increased,” Nadella shared. “That was a significant consideration. I wanted to avoid being burdened with four or five years of depreciation on one generation.”
Nvidia chose not to comment.
Dustin Madsen, vice president of the Society of Depreciation Professionals and founder of Emrydia Consulting, stated that depreciation is a managerial financial estimate, and advancements in a swiftly evolving sector like technology can alter initial forecasts.
Depreciation assessments, Madsen stated, typically incorporate factors such as technological obsolescence, upkeep, historical longevity of similar apparatus, and internal engineering evaluations.
“You must persuade an auditor that your proposed lifespan is genuinely its lifespan,” Madsen remarked. “They will scrutinize all relevant elements, including your engineering evidence suggesting that these assets’ lifespan is close to six years, and conduct a thorough audit.”
