Several of the world's most valuable companies plan to invest $1 trillion over the next five years in building artificial intelligence data centers, and executives and investors are focused on one key item: depreciation.
In accounting, depreciation refers to the process of amortizing the cost of fixed assets over their estimated useful life. This concept is becoming increasingly important in the tech industry, as companies need to predict how long the hundreds of thousands of Nvidia graphics cards (GPUs) they purchase will remain available or valuable.
Infrastructure giants like Google, Oracle, and Microsoft say their servers have a lifespan of up to six years. But their depreciation cycles can also be much shorter _ Microsoft stated in its latest annual report that its computer equipment has a lifespan of two to six years.
This is a key factor for investors and lenders funding this massive AI infrastructure: the longer the equipment retains its value, the longer the depreciation period for the company, and the smaller the impact on profits.
The depreciation problem of AI graphics cards
AI graphics cards present a unique challenge: they are still relatively new to the market. Nvidia's first AI-dedicated processor for data centers launched around 2018, while the current AI boom began with the release of ChatGPT in late 2022. Since then, Nvidia's data center revenue has surged from $15 billion to $115 billion in the fiscal year ending January this year.
Heim Zaltman, vice president of the Emerging Companies and Growth practice at Riessen LLP, stated that compared to other heavy equipment that enterprises have used for decades, there is no real historical reference for the lifespan of graphics cards.
"Is it three years, five years, or seven years?" Zaltman, who is involved in graphics card financing, said in an interview. "That makes a world of difference to the success of the financing." Some of Nvidia's customers believe that AI chips will retain their value in the long term, and they will continue to pay for older processors because they remain useful for other tasks. CoreWeave, a company that specializes in purchasing and leasing graphics cards to customers, has set a six-year depreciation period for its infrastructure since 2023.
Following the release of its quarterly earnings report, CoreWeave CEO Michael Intertall told CNBC this week that the company takes a "data-driven" approach to assessing the lifespan of its graphics cards.
Intertall stated that the Nvidia A100 chips released in 2020 are currently fully pre-ordered. He also mentioned that a batch of Nvidia H100 chips for 2022 became available again due to contract expiration and were immediately pre-ordered at 95% of the original price.
"All the data I have is indicating that this infrastructure will retain its value," Intertall said.
Despite this, CoreWeave's stock price plummeted 16% after the earnings release, impacted by delays from a third-party data center developer affecting its full-year guidance. The stock has fallen 57% from its June high, a drop that also reflects market concerns about overspending in the AI sector and is part of a broader sell-off. Oracle's stock price has also plummeted 34% since hitting an all-time high in September.
One of the most outspoken skeptics of the AI trade is short seller Michael Burry, who recently disclosed his short bets on Nvidia and Palantir.
This week, Burry pointed out that companies like Meta, Oracle, Microsoft, Google, and Amazon have overstated the lifespan of their AI chips while underestimating depreciation costs. He believes the actual lifespan of server equipment is about two to three years, and these companies have therefore inflated their earnings.
Amazon and Microsoft declined to comment, while Meta, Google, and Oracle did not respond to requests for comment.
"Hoppers chips will be unwanted"
AI chips can depreciate in several ways within six years: they may fail due to wear and tear, or they may become obsolete with the release of new graphics cards. They might still be usable for some workloads, but their economic efficiency would be significantly reduced.
Nvidia CEO Jensen Huang has hinted at this. Earlier this year, when Nvidia released its new Blackwell chips, he joked that the value of the previous generation of Hoppers chips would shrink dramatically.
"When Blackwell chips start shipping in large quantities, Hoppers chips won't be worth anything, even if they were given away for free," Huang said at Nvidia's AI conference in March.
"There are some situations where Hoppers chips can still be used," he added, "but those are rare."
Nvidia now releases a new AI chip every year, whereas previously the update cycle was two years. Its closest graphics card competitor, AMD, is following suit.
Nvidia will release its quarterly earnings report next week.
In a February filing, Amazon stated that it had shortened the lifespan of some servers from six years to five years, citing a study that found "the pace of technological advancement is accelerating, particularly in artificial intelligence and machine learning."
Meanwhile, other hyperscale tech companies are extending the projected lifespan of graphics cards in their new server equipment.
Despite Microsoft's plans to heavily invest in AI infrastructure, CEO Satya Nadella stated this week that the company is trying to stagger its AI chip procurement timeline to avoid over-investing in a single generation of processors. He also noted that the biggest competitor to any new Nvidia AI chip is its predecessor.
"Even with Nvidia, one of our biggest lessons has been how fast their technology iterates," Nadella said. "That's a key factor, and I don't want to be stuck with a generation of products that needs four or five years of depreciation." Nvidia declined to comment.
Dustin Madsen, vice president of the Depreciation Professionals Association and founder of Emrydia Consulting, stated that depreciation is a financial estimate made by management, and the dynamics of fast-moving industries like technology can change initial forecasts.
Madsen stated that depreciation estimates typically consider multiple assumptions, including the rate of technological obsolescence, maintenance needs, the historical useful life of similar equipment, and internal engineering analysis.
"You have to convince the auditors that the asset useful life you're claiming is indeed realistic," Madsen said. "They'll check all those factors; for example, if your engineering data shows the asset's useful life is about six years, they'll do a very detailed audit of that."
Source: Content from CNBC
