AI Infrastructure Thesis: Reframing the Hyperscaler Spending Risk

Akhil received a direct brief from the Goldman Sachs India CEO: hyperscalers were spending close to 600 to 700 billion dollars annually on AI infrastructure. What if that spending went bust?

He started with secondary research, looking for historical analogies. The railroad boom emerged as the clearest parallel: most companies that invested in railroad infrastructure during the boom did go bust, but the infrastructure they built became the backbone of the industrial revolution and global trade. The same pattern held for the fiber cable buildout of the late 1990s, which was widely dismissed as overinvestment at the time but became the physical foundation for today's internet and AI infrastructure.

Akhil used those analogies to reframe the question. The risk was not whether individual companies would survive the investment cycle, but whether the infrastructure being built would enable future value creation. He then extended the analysis to identify sectors where AI infrastructure investment could produce non-productivity outcomes: medical research, scientific discovery, and long-horizon applications where the payoff timeline was measured in decades rather than quarters.

The presentation landed with nuance. The CEO and leadership team were already familiar with the railroad example, but the framing challenged their primary focus on near-term productivity returns. The research did not shift their position entirely, but it introduced a new lens: AI investment was already showing up as revenue share gains through reduced time on low-value work, even where direct productivity metrics were flat. That reframe gave leadership a more defensible way to evaluate the program's returns.