The CEO of considered one of China’s prime chipmakers simply despatched a stark warning in opposition to huge tech’s unprecedented race to construct out as many AI knowledge facilities as it will probably.
“Corporations would like to construct 10 years’ price of information middle capability inside one or two years,” Semiconductor Manufacturing Worldwide Corp.’s Co-CEO Zhao Haijun mentioned Wednesday on a name with analysts, per Bloomberg. “As for what precisely these knowledge facilities will do, that hasn’t been absolutely thought by means of.”
Simply Alphabet, Microsoft, Meta, and Amazon alone are anticipated to spend almost $700 billion on AI this 12 months, repeatedly assuring traders that demand outpaces provide. Over the following three years, the value tag of the information middle buildout is anticipated to exceed $3 trillion.
These monetary commitments have began to spook traders. In the meantime, the image that’s step by step forming has led to the present AI hype being in comparison with the dot-com bubble.
Invigorated by the surplus funding flowing into constructing out the web and anticipated demand for web providers, the telecommunications trade spent billions of {dollars} to put out fiber optic cables within the late Nineties. By 2002, the dot-com bubble had lengthy burst, and nonetheless lower than 5% of the fiber optic community was reportedly in use. A telecoms crash rapidly adopted the dot-com bubble burst, and these unused fiber networks, referred to as “darkish fiber,” sat idle for years.
Ultimately, although, the demand for the web did pan out as anticipated, simply years late. As demand for the web grew exponentially, a number of the infrastructure that was constructed out within the late 90s was finally used.
For AI, although, the state of affairs is likely to be a little bit completely different: AI chips utilized in these knowledge facilities have a comparatively clear expiration date, and if demand doesn’t pan out as anticipated earlier than these chips expire, then which means these firms can have flushed an entire lot of cash and sources down the drain.
Meta says its chips at the moment are good for roughly 5 and a half years, up from a earlier estimate of 4 years. Nvidia executives have claimed that the corporate’s chips that had been shipped out six years in the past are nonetheless in full use.
However even when the chips themselves are wonderful, their worth additionally depreciates as new, higher fashions get churned out. At its present charge, Nvidia releases a brand new flagship AI chip yearly.
The depreciation of the worth of chips is factored into firm earnings, however trade consultants are having a tough time seeing eye to eye on whether or not the present estimates are practical.
Some argue {that a} six-year depreciation cycle is completely reasonable, and older GPUs will nonetheless be fascinating as cheaper options when newer, extra superior variations come out. However, you will have traders like Michael Burry (of “The Massive Quick” fame) who claim that the precise helpful lifetime of an AI chip is not more than 2-3 years.
“By my estimates, they’ll understate depreciation by $176 billion 2026-2028,” Burry mentioned in a post on X.
In its newest annual report, Microsoft mentioned its “pc gear” had an estimated helpful life ranging anyplace from two to 6 years.
