Artificial intelligence faces an unexpected challenge: chip durability.

Analysts warn that companies are making overly optimistic estimates about the lifespan of these specialized chips before they become obsolete.

In pursuit of the AI ​​dream, the tech industry has poured approximately $400 billion this year into specialized chips and data centers, but there are growing questions about the wisdom of this unprecedented level of investment.

At the heart of these doubts are overly optimistic estimates about the lifespan of these specialized chips before they become obsolete.

With persistent concerns about an AI bubble and much of the U.S. economy now reliant on the AI ​​boom, analysts warn that the wake-up call could be harsh and costly.

"Fraud" is how renowned investor Michael Burry, famous for the film "The Big Short," described the situation on social media platform X in early November.

Prior to the wave of AI created by ChatGPT, major cloud computing companies typically assumed their chips and servers would have a lifespan of around six years.

However, Mihir Kshirsagar of Princeton University's Center for Information Technology Policy argues that "the combination of wear and tear and technological obsolescence makes the six-year lifespan assumption difficult to sustain."

Another issue: chip manufacturers—with Nvidia as the undisputed leader—are releasing new processors that are far more powerful than before.

Less than a year after launching its flagship Blackwell chip, Nvidia announced that Rubin will launch in 2026 with 7.5 times higher performance.

At this rate, chips will lose 85 to 90 percent of their market value within three to four years, warns Gil Luria of financial advisory firm DA Davidson.

Nvidia CEO Jensen Huang himself articulated this view in March, explaining that once Blackwell was released, nobody wanted to use the previous generation of chips anymore.

The ChatGPT application. (Source: AFP/VNA)

"There are instances where Hopper chips still work well," he added, referring to the older chips. "But not many."

AI processors are also experiencing problems more frequently than before, Luria noted. "They overheat so much that sometimes the devices burn out."

A recent study by Meta on their Llama AI model showed an annual failure rate of 9%.

According to both Kshirsagar and Burry, the actual lifespan of these AI chips is only two or three years.

Nvidia countered in an unusual statement in November, defending the industry's four-to-six-year estimate based on factual evidence and usage trends.

But Kshirsagar believes these optimistic assumptions mean an AI boom based on "artificially low" costs – and the consequences are inevitable.

If companies are forced to shorten depreciation periods, "it will immediately impact profitability" and cut profits, warns Jon Peddie of Jon Peddie Research. "This is where companies get into trouble with fraudulent accounting practices."

Analysts warn that the consequences could be far-reaching across an economy increasingly reliant on AI.

Luria wasn't worried about giants like Amazon, Google, or Microsoft, companies with diversified revenue streams. His focus was on Oracle and CoreWeave.

Both companies are burdened with massive debt while racing to buy more chips to compete for cloud computing customers.

Building a data center requires raising a significant amount of capital, Luria pointed out.

"If they appear less profitable" because the equipment has to be replaced more frequently, "then raising capital will become more expensive."

The situation is particularly precarious because some loans use the chips themselves as collateral.

Some companies hope to mitigate the damage by reselling the old chips or using them for less demanding tasks than advanced artificial intelligence.

A chip from 2023, "if economically feasible, could be used for secondary issues and as a backup," Peddie said.

(Vietnam+)

Source: https://www.vietnamplus.vn/tri-tue-nhan-tao-gap-phai-mot-thach-thuc-khong-ngo-do-ben-cua-chip-post1083040.vnp