IBM CEO Arvind Krishna has issued a stark warning regarding the accelerating pace and financial scale of AI data center expansion, declaring the current trends unsustainable under existing economic models. His assessment arrives amid an industry-wide rush to build massive computing campuses dedicated to advanced model training, which he suggests is driven more by competitive pressure than sound financial logic.
Krishna provides alarming statistics to illustrate the magnitude of the problem. He estimates that outfitting a single one-gigawatt (1GW) AI facility with the necessary specialized compute hardware now approaches a staggering $80 billion.
When projected against the industry’s planned capacity—estimated to be nearing 100GW in future AI facilities—the implied financial exposure across the sector begins to approach $8 trillion. This unprecedented level of capital expenditure is challenging the conventional limits of enterprise spending.
The core of Krishna’s concern lies in the refresh cycle governing high-end AI accelerators, primarily GPUs.1 Unlike traditional enterprise hardware, the specialized silicon deployed in these centers is typically replaced in full every five years. This cycle is not an option but a compounding, repeating obligation.
This mandatory replacement is driven by rapid architectural change. Krishna argues that the performance leaps achieved by each new generation arrive faster than depreciation schedules can comfortably absorb. Hardware that is physically functional becomes economically obsolete long before its financial write-downs are complete, forcing early and expensive churn.2
This shift has materially altered the definition of scale. Specialized accelerators for parallel workloads now sit at the center of spending decisions, moving capital requirements far beyond what traditional, CPU-centric enterprise data centers ever demanded.
The CEO asserts that depreciation is the most misunderstood factor by market participants. The burden on these new facilities no longer rests primarily with energy consumption or land acquisition, but with the forced turnover of increasingly costly hardware stacks, a concern echoed by investors like Michael Burry.
Krishna calculates that for these multi-gigawatt campuses, the required hundreds of billions of dollars in annual profit simply to service the cost of capital is based on present hardware economics, not speculative long-term efficiency gains.
These massive buildout proposals are escalating, with some new AI campuses rivaling the electricity demand of entire nations, prompting parallel concerns about grid capacity and long-term energy pricing stability.
The CEO also casts doubt on the technological necessity of this spending, estimating near-zero odds that today’s large language models (LLMs) will reach general intelligence merely on the next hardware generation without a fundamental change in knowledge integration.
The interpretation is difficult to avoid: the massive buildout assumes future revenues will scale proportionately to match unprecedented investment, even as depreciation cycles shorten and power limits tighten, leading to significant financial risk.