Arvind Krishna, the CEO of IBM, recently raised a sharp warning about the scale and economics of the current AI infrastructure buildout. He argued that the pace and size of investment in hyperscaler data centers could produce returns that do not match the astronomical capital being committed. The core of the argument is straightforward: at current infrastructure costs, scaling compute to the levels some now imagine may require capital outlays in the trillions of dollars, while hardware depreciation cycles and uncertain revenue per unit of compute make predictable profitability much harder to achieve.

The exchange has catalyzed a broader conversation about data center ROI, AI infrastructure costs, and how companies, investors, and policymakers should weigh near-term excitement against long-term economics and sustainability.

Why the Warning Matters

Krishna’s public comments matter for three reasons. First, IBM is a legacy infrastructure and enterprise software company; its leaders have long experience with the economics of large scale compute deployments. Second, the numbers involved are very large and affect supply chains, energy systems, and broader corporate capital allocation priorities. Third, when a prominent industry leader expresses skepticism about the financial case for widespread buildouts, it shapes investor expectations and can slow or reshuffle planned projects across the sector.

What Krishna Actually Said and the Napkin Math

On a recent tech podcast, Krishna ran through some rough, but striking, calculations. He estimated that a one gigawatt data center buildout could amount to roughly eighty billion dollars in capital, and that scaling to multiple tens of gigawatts globally would multiply that figure into the trillions. He also emphasized that AI accelerators and other specialized chips depreciate rapidly and often require major refresh cycles every four to five years, which compounds the cost challenge. Those two observations together create a scenario where capital intensity and replacement cycles push required returns very high, making traditional ROI models look fragile.

Putting the Numbers in Context

To make sense of this, consider two dynamics. First, the unit economics of AI compute depend on both the price and utilization of machines. If large fleets sit idle or are underused while waiting for applications that generate high-margin revenue, the capital does not earn its cost of capital. Second, hardware obsolescence is not hypothetical. New chip generations can dramatically shift performance per watt and per dollar, which means older fleets can quickly lose economic relevance even if they remain operational. The combination of high upfront capital and rapid obsolescence widens the gap between expenditure and reliable revenue.

The Debate: Hype versus Productivity Gains

Not everyone agrees with Krishna’s alarm. Some industry leaders and investors argue that the productivity gains enabled by modern AI systems will justify large-scale infrastructure investment. They point to enterprise adoption, new revenue streams, and productivity increases that may accrue over years. Krishna acknowledged the productivity upside for enterprises, but he also stressed that claims about achieving artificial general intelligence or other transformational outcomes are speculative and should not be treated as a certain justification for unlimited capital deployment. The debate is therefore partly empirical and partly philosophical about how to value future productivity and risk.

Regional and Policy Implications for Hyperscaler Data Centers

Hyperscaler data centers are not just private investments; they interact with local grids, water systems, and permitting processes. Large projects can overwhelm local energy capacity and raise environmental questions, which in turn can slow deployments and add cost. Policy makers should therefore consider whether incentives, grid upgrades, or permitting reforms are necessary to manage the societal cost of very large data center programs. If governments choose to subsidize parts of the stack, they should do so transparently and with clear guardrails to ensure public value.

Risks for Companies and Investors

There are a few tangible risks to watch. First, companies that front-load capital to build gigantic proprietary compute fleets may find it difficult to re-role those assets if software and architectures evolve. Second, investors who fund hypergrowth with expectations of steadily rising margins could face sharp corrections if realization of those margins lags or if oversupply depresses pricing. Third, short-term investor euphoria following product demos or model breakthroughs can obscure the underlying capital servicing burden. For risk-conscious boards and CFOs, a disciplined approach to incremental capacity investment tied to clear utilization metrics is safer than all-in capital bets.

Operational Realities: Depreciation, Refresh, and Time Horizon

Operationally, the refresh cycle is the engine that multiplies costs. If a company must replace specialized accelerators every four to five years to remain competitively performant, that becomes a recurring capital requirement. When that recurring cost is multiplied across a fleet sized in gigawatts, the annualized capital servicing requirement may rival revenues unless the company can monetize compute at very favorable rates. This point makes sustainable unit economics much harder to engineer and is central to Krishna’s argument about data center ROI.

How Companies Can Respond Practically

Leaders should consider a few pragmatic responses. First, adopt a pay-as-you-grow approach to capacity, pairing capital deployment with confirmed bookings or firm partnerships. Second, diversify compute sourcing between owned capacity and third-party cloud or co-location to reduce stranded asset risk. Third, invest in higher system utilization through workload consolidation and specialization. Fourth, prioritize energy efficiency and next-generation cooling and power systems to lower total cost of ownership. Finally, build scenario analysis into capital planning to stress-test ROI under slower-than-expected monetization cases.

What Policymakers Should Consider

Policymakers should encourage transparent assessments of social and environmental impact, link incentives to measurable public benefit, and support grid modernization to avoid hidden public costs. They should also consider whether tax and subsidy programs inadvertently encourage overbuilding without guaranteeing domestic economic returns. A balanced public policy can encourage innovation while limiting downside exposure for taxpayers and communities.

Measured Ambition Rather Than Reckless Scale

Arvind Krishna’s warning is not an argument against AI or against building compute capacity. It is a caution about scale and timing. AI data center investment can deliver significant value, but only when matched with realistic monetization paths, operational discipline, and attention to refresh cycles and energy constraints. For companies, investors, and policy makers the sane path is a measured ambition that tests demand, validates business models, and staggers capital deployments so that infrastructure grows with clear utilization and revenue signals. That approach will protect balance sheets and ensure that the AI revolution yields durable rather than fleeting gains.