How long will it take for my GPU to depreciate?

As some of the world’s most valuable companies look to spend $1 trillion over the next five years on data centers for artificial intelligence, one of the concerns for executives and investors is depreciation.

In accounting, depreciation is the act of allocating the cost of a hard asset over its expected useful life. It’s an increasingly important concept in the technology industry as companies predict how long hundreds of thousands of services will last. Nvidia The graphics processing units you have purchased will continue to work or retain their value.

Major infrastructure companies are: google, oracle and microsoft The company says its servers can last up to six years. However, it may also depreciate more quickly. Microsoft said in its latest annual report that its computer equipment has a lifespan of two to six years.

That’s a lot to consider for investors and lenders funding big AI buildouts. This is because the longer the equipment lasts in value, the longer the company can extend the depreciation period and the less negative impact it will have on profits.

AI GPUs pose particular challenges because they are still relatively new to the market. Nvidia’s first AI-focused data center processors arrived around 2018. The current AI boom began with the launch of ChatGPT in late 2022. Since then, Nvidia’s annual data center revenue has jumped from $15 billion to $115 billion in the year ending in January.

Chaim Salzman, vice chairman of Latham & Watkins’ emerging and growth practice, said there is no real track record of how long GPUs last compared to other types of heavy equipment that companies have used for decades.

“Is it three years, five years or seven years?” Salzman, who works on GPU funding, said in an interview. “It’s a huge difference in terms of how successful you are for fundraising purposes.”

Some of Nvidia’s customers say they will continue to pay for access to older processors because AI chips retain their value over time and older processors are still useful for other tasks. coreweaveThe company, which buys GPUs and rents them to customers, has a six-year depreciation cycle for its infrastructure starting in 2023.

CoreWeave CEO Michael Intrator told CNBC this week following quarterly results that the company is “data-driven” in its approach to GPU expiration.

Intrater said all of CoreWeave’s Nvidia A100 chips announced in 2020 are fully booked. He also added that as the contract expired, a batch of Nvidia H100 chips from 2022 became available and were immediately reserved for 95% of the original price.

“Every data point I have shows that the infrastructure holds value,” Intrater said.

Still, Coreweave’s stock fell 16% after the earnings report as the third-party data center developer’s delays hit its full-year outlook. The stock is down 57% from its June high as part of a broader selloff reflecting concerns about overspending in the AI ​​sector. Oracle stock has plunged 34% from its all-time high in September.

Among the most vocal skeptics of AI trading is short seller Michael Varley. He recently revealed his bets on Nvidia and America. Palantir.

Barry suggested this week that the following companies: metaOracle, Microsoft, Google, Amazon They overstate the useful life of AI chips and underestimate their depreciation costs. He said the actual useful life of server equipment is about two to three years, and companies are inflating their profits as a result.

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

There are many possibilities for AI chips to depreciate within six years. They can wear out and break, or become obsolete when new GPUs are released. These are still useful for running certain workloads, but the economics are much worse.

Nvidia CEO Jensen Huang has hinted as much. When Nvidia announced its new Blackwell chip earlier this year, he joked that it would devalue its predecessor, Hopper.

“Once Blackwell starts shipping in volume, we can’t afford to let go of Hopper,” Huang said at NVIDIA’s AI conference in March.

“There are situations where Hopper is fine,” he continued. “Not too many.”

Nvidia used to release new AI chips every two years, but now it releases them every year. advanced micro deviceIts closest GPU competitors followed suit.

Nvidia will report its quarterly results next week.

Amazon said in a February filing that it shortened the useful life of some of its servers from six years to five years after an investigation found “the pace of technology development has increased, particularly in the areas of artificial intelligence and machine learning.”

Meanwhile, other hyperscalers are extending GPU lifetime estimates for new server equipment.

Microsoft plans to aggressively build out its AI infrastructure, but CEO Satya Nadella said this week that the company is striving to space out AI chip purchases and avoid overinvesting in a single generation of processors. He added that the new Nvidia AI chip’s biggest competition is its predecessor.

“One of the biggest learnings we’ve had with Nvidia as well is that the pace of migration has increased,” Nadella said. “That was a big factor. We didn’t want to have a problem with four to five years of depreciation in one generation.”

Nvidia declined to comment.

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

Madsen said depreciation estimates typically take into account assumptions such as technological obsolescence, maintenance, historical useful life of similar equipment, and internal engineering analysis.

“You have to convince the auditors that what they’re suggesting is what that lifespan is going to be,” Madsen said. “They look at all of these factors, including engineering data that suggests the useful life of these assets is about six years, and they audit that at a very detailed level.”

–CNBC’s Jordan Nobe contributed to this article.

Featured: Chris Wood: Removed Nvidia from portfolio. Prefer Chinese AI names