How long before a GPU depreciates?

As a handful of the world’s most valuable companies prepare to spend $1 trillion on AI data centers over the next five years, executives and investors have one item on their minds: depreciation.

In accounting, depreciation is the act of allocating the cost of a tangible asset over its expected useful life. This is becoming an increasingly important concept in the technology industry as companies estimate how long hundreds of thousands of users will be making transactions. Nvidia The graphics processing units or GPUs they purchase will remain useful or retain their value.

Infrastructure giants are like this Google, Seer And Microsoft They said their servers could be useful for up to six years. However, they can lose value much sooner. Microsoft said in its latest statement: annual filing computer equipment lasts two to six years.

This is a lot to consider for investors and lenders financing massive AI builds; because the longer equipment remains valuable, the longer a company can extend depreciation and the less it hurts profits.

AI GPUs pose a particular challenge because they are relatively new to the market. Nvidia’s first AI-focused processors for data centers came out around 2018. The current AI boom began with the launch of ChatGPT in late 2022. Since then, Nvidia’s annual data center revenue has grown from $15 billion to $115 billion in the fiscal year that ended in January.

There’s no real track record for how long GPUs can last compared to other types of heavy equipment that businesses have used for decades, said Haim Zaltzman, vice president of Latham & Watkins’ emerging companies and growth practice.

“Three years, five or seven?” Zaltzman, who is working on GPU financing, said in an interview. “There’s a big difference in terms of how successful it is in terms of funding.”

Some of Nvidia’s customers say its AI chips will retain their value for a long time, and that customers will continue to pay for access to older processors because they will still be useful for other tasks. CoreWeaveIt has been using six-year depreciation cycles for its infrastructure since 2023, which purchases GPUs and leases them to customers.

CoreWeave CEO Michael Intrator told CNBC this week following quarterly earnings that his company is “data-driven” on GPU shelf life.

CoreWeave is fully stocked with Nvidia A100 chips announced in 2020, Intrator said. He also added that a batch of Nvidia H100 chips for 2022 were made available for sale as the contract expired, and they were immediately reserved at 95% of the original price.

“All the data points I get show me that the infrastructure is holding its value,” Intrator said.

Still, CoreWeave shares tumbled 16% following the earnings report as delays at a third-party data center developer hit its full-year target. The stock has fallen 57% from its peak in June, part of a broader selloff reflecting concerns about overspending on artificial intelligence. Oracle shares have fallen 34% since their record high in September.

One of the most vocal skeptics of the AI ​​trade is short seller Michael Burry, who recently announced bets against Nvidia. palantir.

bury this week recommended including these companies MetaOracle, Microsoft, Google and Amazon They overestimate the lifespan of AI chips and understate their depreciation. He estimates the real lifespan of server equipment to be about two to three years, and says companies inflate their earnings as a result.

Amazon and Microsoft declined to comment. Meta, Google and Oracle did not respond to requests for comment.

There are several ways AI chips could lose value before six years. They may wear out and break, or become obsolete when new GPUs are released. They may still be useful for running certain workloads, but the economics could be much worse.

Nvidia CEO Jensen Huang also hinted at this. When Nvidia announced a new Blackwell chip earlier this year, it joked that its predecessor, the Hopper, would fall in value.

“You can’t give Hoppers away once Blackwell starts shipping in volume,” Huang said at Nvidia’s AI conference in March.

“There are situations where Hopper is good,” he continued. “Not a lot.”

Nvidia is now releasing new AI chips on an annual basis, compared to the previous two-year pace. Advanced Micro DevicesIts closest GPU competitor is .

Amazon, in February applicationIt said it had reduced the lifespan of a subset of its servers from six to five years because it had conducted a study that found “the pace of technology development, particularly in artificial intelligence and machine learning, is accelerating.”

Meanwhile, other hyperscalers are expanding GPU lifetime estimates for newer server equipment.

Although Microsoft plans to aggressively build its AI infrastructure, CEO Satya Nadella in question He said this week that his company is trying to space out its AI chip purchases and avoid overinvesting in single-generation processors. He added that the biggest competition for any new Nvidia AI chip is its predecessor.

“One of the biggest learnings we had from Nvidia was the increased speed of switching,” Nadella said. “That was a big factor. I didn’t want to get stuck with four or five years of attrition in a generation.”

Nvidia declined to comment.

Depreciation is a financial estimate made by management, and developments in a fast-moving industry like technology can change initial estimates, said Dustin Madsen, vice president of the Association of Depreciation Professionals and founder of Emrydia Consulting.

Madsen said depreciation estimates often take into account assumptions such as technological obsolescence, maintenance, past lives of similar equipment and in-house engineering analyses.

“You’re going to have to convince an auditor that what you’re suggesting his life will be like is actually his life,” Madsen said. “They’ll look at all of these factors, like your engineering data that suggests the lifespan of these assets is about six years, and they’ll audit that at a very granular level.”

— CNBC’s Jordan Novet contributed to this story.

WRISTWATCH: Chris Wood: We removed Nvidia from our portfolio, preferred Chinese AI names