- Populating a single one-gigawatt AI facility costs nearly $80 billion
- Planned AI capacity across the industry could total 100GW
- High-end GPU hardware must be replaced every five years without extension
IBM chief executive Arvind Krishna questions whether the current pace and scale of AI data center expansion can ever remain financially sustainable under existing assumptions.
He estimates that populating a single 1GW site with compute hardware now approaches $80 billion.
With public and private plans indicating close to 100GW of future capacity aimed at advanced model training, the implied financial exposure rises toward $8 trillion.
Economic burden of next-generation AI sites
Krishna links this trajectory directly to the refresh cycle that governs today’s accelerator fleets.
Most of the high-end GPU hardware deployed in these centers depreciates over roughly five years.
At the end of that window, operators do not extend the equipment but replace it in full. The result is not a one-time capital hit but a repeating obligation that compounds over time.
CPU resources also remain part of these deployments, but they no longer sit at the center of spending decisions.
The balance has shifted toward specialized accelerators that deliver massive parallel workloads at a pace unmatched by general-purpose processors.
This shift has materially altered the definition of scale for modern AI facilities and pushed capital requirements beyond what traditional enterprise data centers once demanded.
Krishna argues that depreciation is the factor most often misunderstood by market participants.
The pace of architectural change means performance jumps arrive faster than financial write-downs can comfortably absorb.
Hardware that is still functional becomes economically obsolete long before its physical lifespan ends.
Investors such as Michael Burry raise similar doubts about whether cloud giants can keep stretching asset life as model sizes and training demands grow.
From a financial perspective, the burden no longer sits with energy consumption or land acquisition, but with the forced churn of increasingly expensive hardware stacks.
In workstation-class environments, similar refresh dynamics already exist, but the scale is fundamentally different inside hyperscale sites.
Krishna calculates that servicing the cost of capital for these multi-gigawatt campuses would require hundreds of billions of dollars in annual profit just to remain neutral.
That requirement rests on present hardware economics rather than speculative long-term efficiency gains.
These projections arrive as leading technology firms announce ever larger AI campuses measured not in megawatts but in tens of gigawatts.
Some of these proposals already rival the electricity demand of entire nations, raising parallel concerns around grid capacity and long-term energy pricing.
Krishna estimates near-zero odds that today’s LLMs reach general intelligence on the next hardware generation without a fundamental change in knowledge integration.
That assessment frames the investment wave as driven more by competitive pressure than by validated technological inevitability.
The interpretation is difficult to avoid. The buildout assumes future revenues will scale to match unprecedented spending.
This is happening even as depreciation cycles shorten and power limits tighten across multiple regions.
The risk is that financial expectations may be racing ahead of the economic mechanisms required to sustain them over the full lifecycle of these assets.
Via Tom’s Hardware
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!
And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.
