China's Top Chipmaker Warns AI Chips and Data Centers Could Sit Idle Amid $3 Trillion Spending

China's Top Chipmaker Warns Rushed AI Capacity Could Sit Idle

China's top chipmaker has warned that breakaway spending on artificial intelligence chips is bringing forward years of future demand, raising the risk that some data centers could sit idle. "Companies would love to build 10 years' worth of data center capacity within one or two years," Semiconductor Manufacturing International Corp.'s Co-Chief Executive Officer Zhao Haijun said Wednesday on a call with analysts. "As for what exactly these data centers will do, that hasn't been fully thought through." AI-related infrastructure investment is projected to exceed $3 trillion over the next five years, according to Moody's Ratings, as developers pour eye-watering sums into data centers to house training and inference chips designed by companies including Nvidia Corp., Advanced Micro Devices Inc. and Huawei Technologies Co. In 2026 alone, the combined capital expenditure of Alphabet Inc., Amazon.com Inc., Meta Platforms Inc. and Microsoft Corp. is on track to reach $650 billion, driven by their costly AI arms race. Get the Tech Newsletter bundle. Get the Tech Newsletter bundle. Get the Tech Newsletter bundle. Bloomberg's subscriber-only tech newsletters, and full access to all the articles they feature. Bloomberg's subscriber-only tech newsletters, and full access to all the articles they feature. Bloomberg's subscriber-only tech newsletters, and full access to all the articles they feature. Bloomberg may send me offers and promotions. Plus Signed UpPlus Sign UpPlus Sign Up By submitting my information, I agree to the Privacy Policy and Terms of Service. China's leading AI developers, including Alibaba Group Holding Ltd., Tencent Holdings Ltd. and ByteDance Ltd., are also investing heavily in AI infrastructure equipped with both Nvidia chips and domestically produced alternatives. SMIC operates chipmaking plants from Beijing to Shanghai and Shenzhen, but it can only manufacture less advanced AI chips compared with those produced by Nvidia and its contract manufacturer, Taiwan Semiconductor Manufacturing Co., due to US export restrictions that limit access to cutting-edge equipment. The surge in spending has also triggered a shortage of high-bandwidth memory, a critical high-end component that enables advanced AI computing. The tight supply of HBM could persist for years, as new capacity takes time to build and qualify, Zhao said. SMIC's domestic clients, including Huawei and Cambricon Technologies Corp., are aiming for a rapid ramp-up of their silicon production to meet China's AI needs. "It's like building high-speed rail stations and highways -- even if there aren't that many cars today, you still want to complete 10 years' worth of infrastructure in just two years," Zhao said.

Top Chinese Chipmaker Warns Rapid Data Center Buildout Plan Is Half-Baked

The CEO of one of China’s top chipmakers just sent a stark warning against big tech's unprecedented race to build out as many AI data centers as it can. “Companies would love to build 10 years’ worth of data center capacity within one or two years,†Semiconductor Manufacturing International Corp.’s Co-CEO Zhao Haijun said Wednesday on a call with analysts, per Bloomberg. “As for what exactly these data centers will do, that hasn’t been fully thought through.†Just Alphabet, Microsoft, Meta, and Amazon alone are expected to spend nearly $700 billion on AI this year, repeatedly assuring investors that demand outpaces supply. Over the next three years, the price tag of the data center buildout is expected to exceed $3 trillion. Those financial commitments have started to spook investors. Meanwhile, the picture that is gradually forming has led to the current AI hype being compared to the dot-com bubble. Invigorated by the excess investment flowing into building out the internet and anticipated demand for internet services, the telecommunications industry spent billions of dollars to lay out fiber optic cables in the late 1990s. By 2002, the dot-com bubble had long burst, and still less than 5% of the fiber optic network was reportedly in use. A telecoms crash quickly followed the dot-com bubble burst, and these unused fiber networks, called "dark fiber," sat idle for years. Eventually, though, the demand for the internet did pan out as expected, just years late. As demand for the internet grew exponentially, a lot of the infrastructure that was built out in the late 90s was eventually used. For AI, though, the situation might be a little different: AI chips used in these data centers have a relatively clear expiration date, and if demand doesn't pan out as expected before these chips expire, then that means these companies will have flushed a whole lot of money and resources down the drain. Meta says its chips are now good for roughly five and a half years, up from a previous estimate of four years. Nvidia executives have claimed that the company's chips that were shipped out six years ago are still in full use. But even if the chips themselves are fine, their value also depreciates as new, better models get churned out. At its current rate, Nvidia releases a new flagship AI chip every year. The depreciation of the value of chips is factored into company earnings, but industry experts are having a hard time seeing eye to eye on whether the current estimates are realistic. Some argue that a six-year depreciation cycle is perfectly reasonable, and older GPUs will still be desirable as cheaper alternatives when newer, more advanced versions come out. On the other hand, you have investors like Michael Burry (of "The Big Short" fame) who claim that the actual useful life of an AI chip is no more than 2-3 years. "By my estimates, they will understate depreciation by $176 billion 2026-2028," Burry said in a post on X. In its latest annual report, Microsoft said its "computer equipment" had an estimated useful life ranging anywhere from two to six years.