The Math Behind the Pivot
Rather than trying to compete directly with Amazon, Microsoft, and Google in public cloud infrastructure, Krishna ran the numbers and found the gap too wide to close. Catching up would have required burning five to ten billion dollars a year for half a decade, with little guarantee of cracking the top three. Instead, IBM bought Red Hat and repositioned itself as the neutral layer connecting enterprise clients to all the major cloud providers — a partner to the hyperscalers rather than a rival trying to displace them.
Cutting What Was Dragging Them Down
IBM's legacy IT infrastructure services division was shrinking by around five percent every year. Keeping it meant the rest of the business had to run twice as fast just to stay flat. In 2021, IBM spun it off into a separate public company called Kyndryl — a painful separation involving roughly a third of the entire workforce — freeing the remaining business to grow without that anchor. It's the kind of decision that looks obvious in hindsight and requires real nerve at the time.
Why Frontier AI Is a Trap for Most Companies
Krishna is skeptical of the economics underpinning the current AI infrastructure boom. By his estimate, the capital being poured into data centers globally implies a need for one to two trillion dollars in new annual revenue to justify the investment — revenue that doesn't yet exist. IBM's own AI strategy deliberately avoids the trillion-parameter frontier model race, focusing instead on smaller, domain-specific models embedded directly into the back-office systems of large enterprises: banks, insurers, and consumer goods companies that need AI that is controllable, auditable, and accurate within a narrow domain.
The Mainframe Nobody Talks About
While Wall Street periodically panics that AI coding tools will render enterprise software obsolete, Krishna draws a sharper line. The front-end interfaces and workflow tools — perhaps a quarter of the total software market — are genuinely vulnerable to disruption. But the underlying transaction engines that process every credit card swipe and airline booking in real time are a different story entirely. IBM's latest mainframe runs fraud detection and machine learning inference directly on the chip, completing 450 billion operations a day without sending data to an external cloud. That's not going away.
Quantum's Quiet Countdown
Krishna places quantum computing roughly where GPUs were in 2015 — a technology with clear theoretical power, an active research community, and no mass market yet. IBM expects that to change between 2028 and 2030, when quantum hardware should become commercially viable for molecular simulation, financial risk modelling, and logistics optimisation. The geopolitical stakes are high: quantum systems running the right algorithms can break the encryption that secures most of the world's digital communications, making it as much a national security issue as a business one.