IBM CEO Arvind Krishna does not see the current AI wave as a bubble. In The Verge’s Decoder podcast, he argues that generative AI and large language models represent a structural technological shift, especially in the business world.
According to Krishna, AI is not growing on speculation, but on actual value creation. While he believes the technology is sustainable, he also sees a notable financial threat: the enormous investments in AI data centers are developing at a pace that is difficult to sustain economically.
The core of Krishna’s reasoning is that AI now delivers direct productivity gains. Companies are deploying systems for automation, software development, analysis, and internal efficiency improvements. This, he argues, makes this phase fundamentally different from previous waves of hype. He sees no signs of a bubble, because the underlying demand is driven by concrete business applications rather than speculative expectations.
At the same time, Krishna points to risks outside the technology itself. In an analysis highlighted by Tom’s Hardware, among others, he points to the capital expenditures planned by major AI players. According to his calculations, filling a one-gigawatt AI data center costs about $80 billion in hardware. Large companies are now working on plans for tens to even hundreds of gigawatts of future capacity. In total, this could theoretically amount to around $8 trillion in investments.
According to Krishna, that amount is arduous to recoup. To cover the capital costs, companies would have to generate about $800 billion in profits per year collectively. In addition, current AI accelerators are typically depreciated over five years. Due to rapid technological progress, Krishna expects that replacing virtually the entire hardware fleet will remain necessary. He argues that the equipment must be fully utilized during that period, as it will become obsolete afterward and have to be repurchased.
Many data center plans are economically questionable, according to Krishna
This combination of short replacement cycles and enormous economies of scale means that, according to Krishna, some of the current data center plans are economically questionable. This is separate from the question of whether AI itself is a bubble. Krishna makes a clear distinction here. In his view, the technology is sustainable, and demand from enterprises continues to grow. The infrastructure race between some players may not be.
IBM has therefore opted for a different approach. Krishna acknowledges that the first generation of IBM’s AI, such as the original Watson, was too monolithic and consequently difficult to scale. He believes that the new modular approach is better suited to business practices and makes AI applicable across a variety of scenarios without requiring the replacement of entire systems.
Krishna expects generative AI to remain an important driver of productivity growth in the coming years, while quantum computing could be the next step in the longer term. He believes that the technology itself will remain robust, but that the economic basis for ever-larger data centers requires more realism.