In 2022, Jason Eichenholz was standing at a podium at an automotive industry conference, talking about lidar sensors and the promise of self-driving cars, when an executive pulled him aside after his remarks. “Jason,” the executive told him, “you don’t have a front-row seat. You’re part of the show.” Eichenholz was then cofounder and chief technology officer at Luminar, a company whose optical sensing technology was helping push autonomous vehicles toward commercial reality. The executive’s words, he would later reflect, captured something far larger than any single industry’s transformation.
The insight was this: when a technology reshapes an entire sector, the companies that built the old infrastructure rarely define the new one. In the autonomous vehicle race, it was Waymo and Zoox, companies born in Silicon Valley rather than Detroit, that kept advancing while Ford’s Argo AI folded, General Motors’ Cruise stumbled, and Volkswagen’s Cariad struggled to find its footing despite tens of billions in collective investment. The incumbents weren’t incompetent. They simply approached the problem from within a framework that the technology itself was rendering obsolete.
Eichenholz, now founder and CEO of Relativity Networks, which develops optical networking for AI infrastructure, believes the same pattern is unfolding again, this time at a scale that dwarfs the autonomous vehicle moment.
The driving force is artificial intelligence, and specifically the enormous physical infrastructure required to build and run it. The hyperscalers, Google, Microsoft, Amazon, Meta, are constructing data centres at a pace and a power appetite that existing grids and telecommunications networks were simply never designed to accommodate. Rather than wait for utilities and carriers to catch up, these companies are increasingly doing what the best-funded technology companies do when institutions move too slowly: they are building the infrastructure themselves.
Microsoft signed a twenty-year power agreement that helped make possible the restart of the Three Mile Island nuclear facility. Elon Musk has pursued dedicated natural gas generation for xAI’s data centre campuses. Texas has begun requiring many large new data centres to bring their own generation capacity, a tacit acknowledgment by regulators that the public grid cannot keep pace with AI’s appetite. When a single campus can justify billions of dollars in dedicated energy investment, the economics of infrastructure shift in ways that blur the line between technology company and utility.
The same logic is reshaping telecommunications. The carrier networks that underpin today’s internet were designed for consumers, enterprises, and mobile devices. AI changes the traffic patterns entirely, routing vast flows of data between geographically distributed campuses in ways that consumer-oriented networks were never built to handle efficiently. Rather than wait for carriers to evolve their architecture, hyperscalers are purchasing networking equipment directly from manufacturers and building private optical networks between their own facilities. They are becoming builders of infrastructure rather than customers of it.
None of this reflects a failure on the part of utilities or telecommunications providers. Eichenholz is careful to make that distinction. These are institutions built around long investment cycles, regulated returns, environmental review processes, and rate-case proceedings that can stretch across years. AI infrastructure is being deployed in months. The mismatch is not one of competence. It is one of speed, and the companies generating the demand have decided they cannot afford to wait for the institutions designed to serve that demand to catch up.
The parallel to autonomous vehicles is instructive precisely because it does not end in triumph for the technology insurgents. Waymo operates robotaxi services in a handful of American cities. Zoox is still developing. The traditional automakers, after painful reorganizations, are still building most of the world’s cars. The incumbents adapted, eventually, but not before innovators had forced a fundamental rethinking of what the system could be.
What that history suggests for AI infrastructure is that the hyperscalers will not replace utilities or carriers entirely. Their role, Eichenholz argues, will become more important, not less, as AI scales. But the institutions built for a previous era of energy and connectivity will be compelled to rethink their architectures, their timelines, and their assumptions about who sets the pace of change. The technology companies are not waiting for permission. They are already building. The question, as Eichenholz frames it, is not whether the infrastructure will change. It is who will build it, and whether the institutions that governed infrastructure in the twentieth century will shape that process or simply respond to it.
The show, as that automotive executive might have put it, has already started.
