The big wrinkle in the multitrillion-dollar AI buildout

By Clare Duffy, CNN

New York (CNN) — There’s a giant question hanging over the tech industry: How long will its massive investments in AI infrastructure really last?

Tech giants are shelling out hundreds of billions of dollars on artificial intelligence infrastructure — mainly, data centers and the chips that power them. It’s an investment they say will set the stage for AI to overhaul our economy, our jobs and even our personal relationships.

This year alone, tech firms are expected to pour $400 billion into AI-related capital expenditures.

A portion of that will almost certainly put a recurring strain on companies’ balance sheets. And for companies hinging their future on AI, the question of how frequently they’ll have to upgrade or replace advanced chips is a critical one — especially since there’s growing skepticism of whether AI will produce returns large or quickly enough to recoup both existing investments and cover future infrastructure costs.

That’s fueling concerns around an AI bubble — worries that the hype around and spending on AI is out of sync with its true value. Those worries come as the “Magnificent Seven” tech stocks make up around 35% of the value of the S&P 500, raising questions about what an AI crash would mean for the economy.

“The extent to which all of this build out is a bubble partially depends on the lifespan of these investments,” said Tim DeStefano, associate research professor at Georgetown’s McDonough business school.

Chip lifecycles

It’s unclear how long top-of-the-line graphics processing units (GPUs), the chips most often used for AI training and processing, will remain useful.

Several tech experts told CNN that they estimate AI chips can be used to train large language models between 18 months and three years. But the chips could continue being used for less intensive tasks for several more years, they added.

In contrast, central processing units (CPUs) used in traditional non-AI data centers are typically replaced every five to seven years, the experts said.

That’s partially because training AI models exposes the chips to significant strain and heat, wearing them down faster. About 9% of GPUs will fail over the course of a year, compared with around 5% of CPUs, said David Bader, professor of data science at the New Jersey Institute of Technology.

Subsequent generations of AI chips are also rapidly improving and becoming more efficient, meaning it might not be economical to continue running AI workloads on older chips even if they’re functional.

Different experts offer slightly different estimates. DeStefano said AI chips will likely break down after about five to 10 years of use, but their economic lifespan is only around three to five years.

Meanwhile, Bader estimates GPUs can be used to train AI models for 18 to 24 months. But he said older chips can still handle tasks like processing users’ AI queries, known as inference, for around five more years, extending their value.

Nvidia, the largest provider of AI chips, says its CUDA software system enables customers to update existing chips’ software, potentially delaying the need to upgrade to the latest product.

Nvidia CFO Colette Kress said on the company’s latest earnings call last month that GPUs “shipped six years ago are still running at full utilization today” because of its CUDA system.

But whether chips last two years or six years, tech companies still face the same question: “Where’s the revenue going to come in that’s going to allow you to rebuild at that scale?” said Mihir Kshirsagar, director of the technology policy clinic at Princeton’s Center for Information Technology Policy.

What does this have to do with the AI bubble?

The faster chips degrade, the more companies will feel pressure to see returns on AI to fund their replacement.

And long-term demand for AI remains unclear, especially in light of reports this year that most companies implementing the technology haven’t yet seen benefits to their bottom lines. Corporate customers are going to be the real moneymakers for AI companies, but those firms are still figuring out how to the technology use to generate revenue or reduce costs, DeStefano said.

“There’s the demand for generative AI from individual users … but that’s not enough for these large AI companies to recoup their investment costs,” he said.

Michael Burry, the famed investor behind “The Big Short,” recently warned of an AI bubble. His argument is based in part on the prediction that tech firms are overestimating the valuable life of their chip investments, which could eventually weigh on their earnings.

AI leaders are also starting to talk more openly about the question.

Microsoft CEO Satya Nadella said in a podcast interview last month that the company has begun spacing out its infrastructure investments so that its data center chips don’t become obsolete at the same time.

And OpenAI CFO Sarah Friar raised alarms last month when she said the company’s role as a frontier AI model maker depends on whether the most advanced chips last “three years, four years, five years or even longer.”

If that lifecycle is shorter, she suggested the company might need the US government to “backstop” the debt it’s taking on to finance its aggressive infrastructure commitments. (OpenAI quickly tried to walk back the comment, saying it was not seeking a government backstop.)

In previous market bubbles, infrastructure built during the hype cycle that sat dormant after the burst was still usable years later. Fiber optic cables laid during the dot-com bubble of late 1990s, for example, now provide the foundation for today’s internet.

But the AI bubble — if it’s real — would be a different situation, said Paul Kedrosky, managing partner at investment firm SK Ventures. He argued AI data centers won’t retain the same potential for use over time without ongoing investments in new chips. And the ramifications could extend well beyond tech giants’ balance sheets and share prices.

“Not only are we building these data centers, (tech firms) are pushing to build electricity plants to support all of it,” Kshirsagar said. “If the economics don’t work out, there are some very big societal questions.”

CNN’s Krishna Andavolu contributed to this report.

The-CNN-Wire
™ & © 2025 Cable News Network, Inc., a Warner Bros. Discovery Company. All rights reserved.