IBM CEO says the math behind trillion-dollar AI data centers does not work today

Napkin math that stops the room

IBM CEO Arvind Krishna recently walked through some back-of-the-envelope math on AI data centers and essentially stated that the numbers do not add up.

Krishna framed the arithmetic this way: if the industry were to spend about eight trillion dollars in cumulative capital on AI compute, then, assuming a roughly ten percent annual return requirement to cover the cost of capital, that would imply on the order of eight hundred billion dollars in annual profit to make the math work.

How one gigawatt turns into a massive bill

Krishna used a rule of thumb that equipping a one-gigawatt AI training site with compute and supporting infrastructure costs roughly eighty billion dollars at current prices.

Now imagine a single hyperscaler committing 20 to 30 gigawatts over time. You are suddenly staring at roughly $1.5 trillion in capex for just one company. Multiply that across all the big AI contenders, and the total becomes staggering.

Depreciation makes the treadmill even faster

The hardware inside these AI factories does not last forever. Krishna notes that accelerators need to be written off in about five years, after which they are effectively replaced, and the racks refilled.

That means the industry is not just buying once, but is also locking into a repeating refresh cycle. To merely stand still on performance, you are constantly racing against depreciation and obsolescence.

Scaling to one hundred gigawatts changes the stakes

Examining public announcements, Krishna estimates that total AI compute commitments worldwide amount to around one hundred gigawatts. Using his $8 billion per gigawatt rule of thumb, that implies roughly 8$ trillion of future capex.

He is blunt about the implications, stating that he sees no realistic way to earn an adequate return on that much spending under current pricing and demand assumptions.

Why $8 trillion demands impossible profits

Here is where his skepticism really bites. $8 trillion in capital, financed in today’s world, would require about ten percent of that in yearly profit to cover the cost of capital.

That is roughly $800 billion in annual profit from AI data center investments alone. For perspective, that is like minting a new mega tech giant’s worth of earnings out of thin air.

OpenAI’s giant ambitions meet IBM caution

Krishna’s view quietly clashes with the ultra-bullish stance from Sam Altman and other AI leaders. OpenAI has publicly disclosed infrastructure commitments in the roughly $1.4 trillion range and argues those investments will be economically justified over time.

Krishna frames that as a belief that one company can capture almost the entire market. He understands the ambition but stops short of sharing the conviction.

Energy and infrastructure needs skyrocket

The financial math is only half the story. To feed those 100 gigawatts of AI compute, you need enormous power and cooling. Altman has even suggested adding 100 gigawatts of new energy capacity per year to support future AI.

Krishna’s comments imply a more grounded view that the physical and financial constraints are tightly coupled, and neither looks trivial to overcome.

IBM skepticism extends to AGI timelines

Underpinning Krishna’s caution is a deeper doubt about artificial general intelligence. He pegs the probability of reaching true AGI with today’s techniques at between zero and one percent without a significant breakthrough.

In other words, he does not believe simply piling on more compute, more models, and more data centers will magically cross some intelligence threshold anytime soon.

Other AI leaders are cooling the AGI hype

Krishna is not alone here. Other prominent voices have started calling the AGI narrative overhyped, even describing it as bordering on mass hypnosis.

The emerging consensus among skeptics is that current large language models are powerful tools, but not, by themselves, the seeds of human-level cognition.

If Krishna’s arithmetic holds, then many of the most aggressive capital expenditure plans will require either much higher prices or sustained demand growth to be viable.

Scaling is giving way to a new research phase

Even inside frontier labs, the mood is shifting. Researchers like Ilya Sutskever have argued that simply scaling model size by one hundred times from here will not deliver the mythical step change people expect.

Krishna echoes that sentiment, saying the path to something closer to AGI likely involves fusing hard symbolic knowledge with LLMs and other approaches, not just building bigger prediction engines.

Krishna still sees trillions in real enterprise value

Despite all the skepticism, he is far from bearish on AI itself. Krishna expects current-generation systems to unlock trillions of dollars in productivity across enterprises, automating workflows, augmenting workers, and reshaping software.

His point is not that AI is a fad. It is that being transformative does not automatically justify every eye-watering infrastructure bet being floated in today’s hype cycle.

The bubble question and IBM’s long game

When pressed on whether this resembles another dot-com-style bubble, Krishna draws a nuanced line. He agrees some players will lose money chasing impossible returns, but he does not see the entire field as doomed.

IBM’s strategy is quieter, focusing on tools, services, and even quantum research rather than building the most expensive AI factories on Earth and hoping it all pays off.

And if you want to see how other tech giants are rethinking their infrastructure strategies, take a look at Microsoft, which has introduced a powerful two-state AI engine for merging data centers.

Why his warning matters for the next AI decade

Krishna’s napkin math is more than a contrarian hot take. It serves as a reminder that physics, finance, and timelines must align for this AI boom to be sustainable.

If the economics of trillion-dollar data centers truly do not work at today’s rates, something has to give: prices, expectations, or architectures. For investors, builders, and policymakers, ignoring that mismatch would be the real fantasy.

And if you want to see how some companies are racing ahead despite those concerns, take a look at how Google is accelerating its AI ambitions with a $40 billion investment in Texas data centers.

What do you think about IBM’s CEO’s verdict that even spending trillions on a data center, it won’t work today? Please share your thoughts and drop a comment.

Read More From This Brand:

• Can America keep up with AI’s growing data center demands?
• OpenAI’s massive data centers set to outpower New York City, warns Sam Altman
• JPMorgan sees a five trillion AI data center surge draining capital from all major debt sectors

Don’t forget to follow us for more exclusive content on MSN.

If you liked this story, you’ll LOVE our FREE emails. Join today and be the first to get stories like this one.

This slideshow was made with AI assistance and human editing.