"Arvind Krishna, who has been at the helm of the legacy tech company since 2020, said even a simple calculation reveals there is "no way" tech companies' massive data center investments make sense. This is in part because data centers require huge amounts of energy and investment, Krishna said on the Decoder podcast. Goldman Sachs estimated earlier this year that the total power usage by the global data center market stood at around 55 gigawatts, of which only a fraction (14%) is dedicated to AI."
"Yet, building out a data center that uses merely one gigawatt costs a fortune-an estimated $80 billion in today's dollars, according to Krishna. If a single company commits to building out 20 to 30 gigawatts then that would amount to $1.5 trillion in capital expenditures, Krishna said. That's an investment about equal to Tesla's current market cap. All the hyperscalers together could potentially add about 100 gigawatts, he estimated."
""It's my view that there's no way you're going to get a return on that because $8 trillion of CapEx means you need roughly $800 billion of profit just to pay for the interest," he said. Moreover, thanks to technology's rapid advance, the chips powering your data center could quickly become obsolete. "You've got to use it all in five years because at that point, you've got to throw it away and refill it," he said."
Global data center power usage is roughly 55 gigawatts with about 14% currently dedicated to AI, and AI demand could push total usage to around 84 gigawatts by 2027. Building a one-gigawatt data center can cost about $80 billion, so deploying 20–30 gigawatts could require roughly $1.5 trillion in capital expenditures. Adding around 100 gigawatts across hyperscalers could imply about $8 trillion of investment, which would demand extremely large profits just to service interest. Rapid technology advances can make data-center chips obsolete within five years, increasing replacement and operating risk. AGI competition is accelerating these investments.
Read at Fortune
Unable to calculate read time
Collection
[
|
...
]