IBM’s Krishna: Trillions in AI spend, “Near Zero” Odds of AGI from Today’s Tech
IBM CEO Arvind Krishna has poured cold water on some of the biggest assumptions behind the current AI boom, even as he insists it is not a bubble and continues to invest heavily in generative AI and quantum computing. He was speaking in a long-form Decoder interview with Nilay Patel for The Verge.
The Real Economics of AI Data Centres
Krishna’s most striking point is about the economics of the AI data centre build-out. He estimates that filling a 1 GW AI data centre with current-generation accelerators costs around 80 billion US dollars. Scaled up to roughly 100 GW of capacity, as implied by today’s headline announcements, he puts the total bill at about 8 trillion dollars. On his numbers, that would require about 800 billion dollars in annual profit just to service the capital, which he basically says is not going to happen. He expects some equity investors to do well, but a lot of debt will never be paid back.
At the same time, he argues that the underlying AI technology is a clear improvement over earlier deep learning systems and will unlock “trillions” in enterprise productivity once hardware, model architectures and software optimisation cut costs by roughly a factor of 30 over about five years. His view is that the consumer platforms chasing massive scale may see very mixed returns, but the technology itself is not a bubble.
Zero-Chance AGI?
On artificial general intelligence, Krishna is far more sceptical than many of today’s AI leaders. He puts the probability that current large language model-centric approaches will ever deliver AGI in the range of zero to one percent. LLMs, in his view, are powerful statistical engines that are “here to stay”, but they are not enough on their own. He expects any path to AGI to require fusing LLM-style models with more explicit forms of knowledge and reasoning, and he thinks the next breakthrough is more likely to come from academia and national labs than from commercial labs doubling down on ever-larger transformers.
Krishna is also surprisingly candid about IBM’s past AI missteps. He describes the original Watson push into healthcare as “inappropriate” and admits that IBM tried to sell AI as a monolithic system in one of the toughest and most regulated possible markets. He now frames Watsonx as a rebuilt, modular AI stack aimed at enterprise workloads, where customers want building blocks they can tune and integrate, rather than a single closed system.
On jobs, Krishna expects AI to displace up to about 10 percent of roles, concentrated in specific functions, but he argues that the bigger opportunity is to use AI to make junior staff perform like experienced experts. Internally, he says a 6 000-person software team that adopted IBM’s own AI coding assistant saw productivity gains of around 45 percent in four months, and IBM is using that as a reason to hire more engineers and ship more products, rather than cut headcount.
Quantum Computing
Looking beyond GPUs, Krishna lays out IBM’s case for quantum computing as a third compute pillar alongside CPUs and GPUs. He positions quantum processors as an additional accelerator for problems that are effectively intractable on classical hardware, not as a replacement. Citing outside consultancy work, he says utility-scale quantum systems could enable 400 to 700 billion dollars of annual value, with the tech industry capturing perhaps 20 to 30 percent of that. He talks about a three- to five-year window to reach “utility” systems, with early work already under way in areas such as finance, batteries and materials.
Strategically, Krishna reiterates IBM’s commitment to hybrid and regulated environments. He does not believe most large customers will move entirely to a single public cloud, especially outside the United States or in sectors like finance and healthcare. IBM’s plan is to be the technology partner for those hybrid, compliance-heavy deployments, while betting that AI and, later, quantum will be the differentiators on top.
The full podcast interview is available on The Verge’s Decoder.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
