As tech giants race to build data centers targeting artificial general intelligence (AGI), IBM CEO Arvind Krishna cautions that the industry is on a costly path that's difficult to recoup.
Speaking on a recent podcast on The Verge, Krishna estimated that building a single 1GW AI data center costs about US$80 billion. As AI data center projects that have been announced around the world add up to nearly 100GW in capacity, it means total capital expenditure commitments reach roughly US$8 trillion.
This scale demands annual profits of at least US$800 billion just to break even. Pressure comes largely from the short five-year lifespan of accelerator hardware. The rapid depreciation cycle forces operators to replace entire fleets frequently, causing data center costs to multiply quickly.
American investor Michael Burry recently questioned cloud service providers' (CSP) assumptions, suggesting actual AI server lifespans may be only two to three years. Moreover, as device performance improves and model sizes grow, older GPUs are retired faster, making it uncertain whether cloud firms can extend equipment usage effectively.
On AI's economic impact, Krishna acknowledged that current generative AI tools boost employee productivity but said true AGI remains distant. He rated the chance of achieving AGI with today's large language model (LLM) technology at just 0-1%, unless there are breakthroughs that enable integration of more diverse knowledge forms.
Business Insider reported that Safe Superintelligence (SSI) CEO Ilya Sutskever also noted in a recent podcast that gains from simply scaling compute and data have plateaued, urging the industry to refocus research on more efficient use of computing power.
Article edited by Jack Wu
