IBM's CEO says the AI infrastructure math doesn't work
December 2025 Update: IBM CEO Arvind Krishna warning 100GW of planned AI infrastructure requires $8T CapEx and $800B annual profit to justify—"There's no way you're going to get a return on that." Krishna estimating 0-1% chance current LLM architectures achieve AGI. Equipment depreciation forcing replacement every 5 years compounds the challenge. Enterprise AI succeeding; speculative AGI infrastructure struggling.
Arvind Krishna, IBM's chief executive, delivered a sobering assessment on The Verge's Decoder podcast: the trillion-dollar AI infrastructure buildout cannot generate returns under current economics.¹ Krishna's analysis relies on straightforward arithmetic that the industry has largely chosen to ignore.
The numbers work like this. A one-gigawatt AI data center requires approximately $80 billion in compute hardware.² Current plans across public and private announcements point to roughly 100 gigawatts of AGI-focused capacity.³ That totals $8 trillion in capital expenditure. To justify that investment, operators need approximately $800 billion in annual profit just to service the cost of capital.⁴
"There's no way you're going to get a return on that in my view," Krishna stated plainly.⁵
The depreciation problem nobody discusses
Krishna identified depreciation as the calculation most underappreciated by investors and analysts. AI accelerators depreciate over five years. The pace of architectural change means operators must replace entire fleets rather than extend them. Equipment purchased today becomes obsolete before it pays for itself.
"You've got to use it all in five years because at that point, you've got to throw it away and refill it," Krishna explained.⁶ Every hardware generation delivers substantial performance improvements. Running inference on five-year-old GPUs means paying more for electricity than the compute produces in value. The depreciation cycle forces continuous reinvestment that compounds the capital requirements.
The hyperscalers building these facilities understand the depreciation math. They build anyway, betting that AI revenue growth will outpace infrastructure costs. Krishna questions whether that bet can pay off at the scale currently planned. The gap between capital requirements and plausible revenue remains too wide.
Krishna diverges from the AGI believers
The IBM CEO expressed skepticism about artificial general intelligence reaching fruition through current approaches. Krishna estimated the likelihood that existing large language model architectures achieve AGI at somewhere between zero and one percent.⁷ AGI "will require more technologies than the current LLM path," he argued.
When Decoder host Nilay Patel noted that OpenAI CEO Sam Altman believes OpenAI can generate returns on its capital expenditures, Krishna offered a direct response: "That's a belief. That's what some people like to chase. I understand that from their perspective, but that's different from agreeing with them."⁸
The disagreement highlights a fundamental divide in the AI industry. One camp believes AGI will emerge from scaling current architectures, justifying any level of infrastructure investment. The other camp, including Krishna, believes the path to AGI requires breakthroughs beyond larger models and more compute. That belief gap drives wildly different investment calculations.
The enterprise opportunity remains real
Krishna's skepticism about AGI infrastructure economics does not extend to enterprise AI. The IBM CEO maintains that AI "is going to unlock trillions of dollars of productivity in the enterprise."⁹ The distinction matters: enterprise AI operates at scales where infrastructure costs generate clear returns.
A company deploying AI to automate customer service, optimize supply chains, or accelerate drug discovery can calculate return on investment with reasonable precision. The compute requirements remain bounded. The value generated flows directly to the organization deploying the technology. Enterprise AI economics work at human scale.
AGI economics operate differently. Building infrastructure to achieve artificial general intelligence requires massive upfront investment with uncertain payoff. The timeline extends beyond normal investment horizons. The competitive dynamics encourage overbuilding, as each player races to achieve breakthrough capabilities first. Rational investment analysis gives way to existential competition.
IBM's strategic positioning reflects Krishna's view. The company focuses on enterprise AI deployment rather than frontier model development. The $11 billion Confluent acquisition announced days after Krishna's podcast appearance reinforces that focus.¹⁰ IBM builds the data infrastructure enterprises need to deploy AI, leaving AGI pursuit to others willing to accept longer odds.
What if Krishna is right?
The AI infrastructure buildout continues regardless of Krishna's warnings. Microsoft, Google, Amazon, and Meta collectively plan hundreds of billions in data center investment. OpenAI, through its Stargate partnership, targets $500 billion in infrastructure spending.¹¹ The capital keeps flowing.
If Krishna's analysis proves correct, the reckoning arrives within the depreciation cycle. Facilities built in 2025 must generate returns by 2030 or face write-downs. Companies that overbuilt will rationalize capacity. The AI infrastructure market will consolidate around survivors with sustainable economics.
The scenario does not require AI to fail. AI can deliver enormous value while AGI-scale infrastructure investment still proves uneconomic. Enterprise AI flourishes while AGI pursuit consumes capital without adequate returns. The distinction between valuable technology and sound investment remains important.
If Krishna's analysis proves wrong, the AGI believers win the biggest bet in technology history. Successful AGI would justify any infrastructure investment and reward early builders with transformative capabilities. OpenAI, xAI, and other AGI-focused organizations operate on that thesis.
The market will decide
IBM occupies a different position than the companies Krishna implicitly criticizes. Big Blue lacks the resources to compete in the AGI infrastructure race and has chosen not to try. Krishna's warnings could reflect strategic positioning as much as analytical conviction. IBM benefits if enterprise AI captures more attention and capital than AGI pursuit.
That positioning does not invalidate the analysis. The arithmetic Krishna presents requires refutation, not dismissal. Eight trillion dollars in capital requires hundreds of billions in annual returns. Current AI revenue, while growing rapidly, falls far short. The gap must close for the investments to pay off.
Krishna's willingness to voice skepticism publicly provides a useful counterweight to the prevailing enthusiasm. The AI industry benefits from executives who question assumptions and demand rigorous financial analysis. Whether Krishna proves right or wrong, the discipline of answering his challenge strengthens investment decisions across the ecosystem.
The trillion-dollar question remains open: can AI infrastructure generate returns at the scale currently planned? Krishna says no. The builders disagree. The next five years will reveal which view proves correct.
References
-
Fortune. "IBM CEO warns there's 'no way' hyperscalers like Google and Amazon will be able to turn a profit at the rate of their data center spending." December 3, 2025. https://fortune.com/2025/12/03/ibm-ceo-no-way-hyperscalers-google-amazon-turn-profit-data-center-spending/
-
Tom's Hardware. "IBM CEO warns that ongoing trillion-dollar AI data center buildout is unsustainable." December 4, 2025. https://www.tomshardware.com/tech-industry/ibm-ceo-warns-trillion-dollar-ai-boom-unsustainable-at-current-infrastructure-costs
-
Tom's Hardware. "IBM CEO warns that ongoing trillion-dollar AI data center buildout is unsustainable."
-
Fortune. "IBM CEO warns there's 'no way' hyperscalers will be able to turn a profit."
-
Fortune. "IBM CEO warns there's 'no way' hyperscalers will be able to turn a profit."
-
TechRadar. "IBM CEO warns trillion-dollar AI data center expansions risk catastrophic losses if hardware refresh cycles stay relentlessly fast." December 2025. https://www.techradar.com/pro/there-is-no-way-ibm-ceo-says-current-ai-data-center-trends-are-unsustainable-and-he-would-know
-
CXO Digital Pulse. "IBM's Arvind Krishna Warns of Unsustainable Economics Behind AGI-Scale Data Center Race." December 2025. https://www.cxodigitalpulse.com/ibms-arvind-krishna-warns-of-unsustainable-economics-behind-agi-scale-data-center-race/
-
Fortune. "IBM CEO warns there's 'no way' hyperscalers will be able to turn a profit."
-
Fortune. "IBM CEO warns there's 'no way' hyperscalers will be able to turn a profit."
-
IBM. "IBM to Acquire Confluent to Create Smart Data Platform for Enterprise Generative AI." IBM Newsroom, December 8, 2025. https://newsroom.ibm.com/2025-12-08-ibm-to-acquire-confluent-to-create-smart-data-platform-for-enterprise-generative-ai
-
OpenAI Stargate Partnership announcements, 2025.
Key takeaways
For strategic planners: - 1GW AI data center requires ~$80B in compute hardware; 100GW of planned AGI capacity = $8T capex - $800B annual profit needed just to service cost of capital on $8T infrastructure investment - Krishna estimates 0-1% probability current LLM architectures achieve AGI
For finance teams: - AI accelerators depreciate over 5 years; architectural changes force full fleet replacement - Running inference on 5-year-old GPUs costs more in electricity than compute produces in value - The gap between capital requirements and plausible revenue remains too wide per Krishna
For enterprise AI teams: - Krishna distinguishes AGI pursuit from enterprise AI: bounded compute requirements, calculable ROI - IBM's $11B Confluent acquisition reinforces focus on enterprise data infrastructure over frontier models - Enterprise AI economics work at human scale; AGI economics require massive upfront investment with uncertain payoff
For investors: - If Krishna proves correct, reckoning arrives within 5-year depreciation cycle (facilities built 2025 must return by 2030) - OpenAI Stargate partnership targets $500B infrastructure; Microsoft/Google/Amazon plan hundreds of billions more - IBM positioned for enterprise AI deployment rather than frontier model development
