How long can Big Tech’s AI bet really last?

Big Tech is spending at a scale rarely seen outside wartime or space races, pouring hundreds of billions of dollars into artificial intelligence chips and data centres it says will reshape the global economy. But behind the bold promises and soaring stock prices lies an uncomfortable question with worldwide implications: if the hardware powering the AI boom ages faster than expected, could today’s historic investments become tomorrow’s financial burden and trigger a tech bubble with consequences far beyond Silicon Valley?

This year alone, tech firms are expected to spend about $400 billion on AI-related capital expenditures, largely on data centres and advanced chips used to train and run AI models. While companies argue these investments are essential for long-term growth, analysts warn that the short lifespan of AI hardware could strain balance sheets and intensify concerns about an emerging AI bubble, CNN Business writes. 

“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 University’s McDonough School of Business.

At the centre of the debate are graphics processing units (GPUs), the specialised chips that power most modern AI systems. Several experts estimate that top-tier GPUs can be used to train large language models for only 18 months to three years. After that, they may still be usable for less demanding tasks, such as processing user queries, for several more years.

By comparison, traditional central processing units (CPUs) in non-AI data centres are typically replaced every five to seven years. AI chips wear out faster because of the intense heat and strain involved in training models. About 9% of GPUs fail each year, compared with roughly 5% of CPUs, according to David Bader, a professor of data science at the New Jersey Institute of Technology.

Rapid improvements in new generations of chips further shorten the economic life of existing hardware. Even if older chips still function, it may not make financial sense to keep using them. DeStefano estimates that while AI chips may physically last five to 10 years, their economic lifespan is closer to three to five years. Bader puts the useful life for training at 18 to 24 months, with another five years of value for inference tasks.

Nvidia, the world’s largest supplier of AI chips, argues that software upgrades can extend hardware life. The company says its CUDA system allows customers to update existing chips rather than immediately replace them. Nvidia CFO Colette Kress recently said that GPUs “shipped six years ago are still running at full utilisation today.”

Still, uncertainty over returns remains.

“Where’s the revenue going to come in that’s going to allow you to rebuild at that scale?” asked Mihir Kshirsagar, director of the technology policy clinic at Princeton University’s Centre for Information Technology Policy.

Skepticism is growing over whether AI demand will be strong and sustained enough, particularly among corporate customers who are expected to drive most future revenue. Many companies experimenting with AI have yet to see meaningful gains.

“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,” DeStefano said.

Investor Michael Burry has warned of an AI bubble, arguing that companies may be overestimating the useful life of their chip investments. Industry leaders are also voicing caution. Microsoft CEO Satya Nadella has said the company is spacing out infrastructure spending to avoid having chips become obsolete simultaneously.

OpenAI CFO Sarah Friar has questioned whether advanced chips will last “three years, four years, five years or even longer,” noting that shorter lifecycles could affect financing plans.

Unlike past tech bubbles, such as the dot-com era when unused fibre-optic cables later found value, AI infrastructure may require constant reinvestment.

“Not only are we building these data centres, (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.”

By Sabina Mammadli