I saw two stories this week that, on the surface, had very little to do with each other. One was about an underwater AI data center off the coast of China. The other described SpaceX’s plans to place AI compute infrastructure into orbit.
Most readers would probably treat them as unrelated curiosities. I think the threads that bind them tell us something far more important.
Something is happening here that goes well beyond GPUs, beyond the latest large language model and beyond whichever AI application is making headlines this week. Whether it is Claude, ChatGPT, Codex or the next model that captures the industry’s attention, they are all dependent on something much larger that receives surprisingly little discussion.
The real story unfolding beneath artificial intelligence may not be the software at all. It may be the unprecedented infrastructure required to support it.
The scale of that buildout is difficult to overstate. It will demand new power generation, expanded transmission networks, enormous quantities of fiber, advanced cooling systems and data centers unlike anything the industry has built before. It will reshape energy policy, industrial planning and perhaps even geopolitics. Long after people stop arguing about which model performed best on a particular benchmark, the infrastructure built to support those models will still be standing.
That is why these two stories caught my attention. They suggest that the challenge of building AI infrastructure has become so significant that we are beginning to look for answers in places that would have sounded absurd only a few years ago.
The first story came from China, where engineers have begun operating what is being described as the world’s first wind-powered underwater data center. The concept is straightforward, even if the execution is anything but. Place computing infrastructure adjacent to offshore wind generation and use the surrounding seawater as a natural cooling system, reducing both energy consumption and the need for freshwater resources. It is an elegant idea that attempts to solve two of the largest operating challenges facing modern data centers: Electricity and heat.
The second story took the opposite approach. Rather than placing compute beneath the sea, SpaceX outlined a vision for placing AI infrastructure into low Earth orbit. Solar panels would provide abundant energy while the vacuum of space would serve as a giant heat sink through radiative cooling. The company envisions modular compute payloads that could evolve as semiconductor technology advances, potentially creating an entirely new class of off-world computing infrastructure.
At first glance, these projects could not appear more different. One heads toward the ocean floor while the other heads toward space. One relies on wind power and seawater, the other on solar energy and the vacuum beyond our atmosphere.
Yet they are trying to solve exactly the same problem.
Both seek access to abundant energy. Both seek more efficient cooling. Both attempt to escape the increasingly difficult process of building massive data centers on land. The engineering solutions differ dramatically, but the underlying challenge is identical.
Neither proposal is perfect, and neither should be viewed as inevitable. Underwater facilities present obvious maintenance and reliability concerns. Hardware fails. Components need replacement. Technology evolves rapidly. Servicing equipment hundreds of feet beneath the ocean surface is hardly routine. Orbital computing faces an equally daunting list of obstacles involving launch economics, networking, latency, hardware replacement and operational resilience. It remains entirely possible that neither approach proves commercially viable.
Ironically, that uncertainty makes the stories even more significant.
Companies do not invest billions exploring ideas like underwater and orbital computing because they are easier than conventional data centers. They pursue them because conventional data centers are becoming increasingly difficult to build at the scale artificial intelligence demands.
The AI conversation still tends to revolve around chips, yet chips are only one piece of a much larger equation. Every GPU requires electricity. Every rack requires cooling. Every cluster requires networking. Every campus requires land, substations, transmission capacity and permits. Utilities around the world are struggling to meet projected demand. Communities increasingly object to large-scale developments because of their impact on water supplies, power grids and local infrastructure. Environmental reviews add years to construction schedules. Interconnection queues continue to grow longer.
The constraints facing AI today increasingly resemble the constraints that shaped previous industrial revolutions. The challenge is no longer simply inventing better technology. It is building the physical systems that allow that technology to exist at scale.
History offers a useful perspective. The Industrial Revolution was not merely the story of steam engines. It required railroads, steel mills, ports and supply chains capable of moving raw materials and finished goods across continents. The automobile transformed society only after governments and private industry built highways, bridges, fueling stations and manufacturing networks that made mass transportation practical. The internet changed the world because decades of investment created fiber networks, satellites, cellular infrastructure and the data centers that quietly powered the digital economy.
Electrification may be the closest comparison. We celebrate the inventions electricity enabled, yet the real achievement was the construction of the grid itself. Power plants, transmission lines and distribution networks became the platform upon which thousands of future innovations could be built. The infrastructure ultimately proved more important than any individual application because everything else depended upon it.
Artificial intelligence may be following the same path. Today’s headlines belong to increasingly capable models and increasingly sophisticated applications. History, however, may remember this period for something else entirely: The enormous physical infrastructure buildout required to support them. The algorithms will capture our imagination, but the enduring legacy of this generation may ultimately be measured in gigawatts, transmission corridors, fiber networks and data centers stretching from the bottom of the ocean to the edge of space.
That realization should also broaden the policy conversation.
Infrastructure of this magnitude has historically involved more than private investment alone because its effects extend far beyond the companies that build it. Railroads, electrical grids, interstate highways and telecommunications networks all evolved through combinations of entrepreneurial vision and public policy because they became foundational to economic growth and national development.
AI infrastructure appears to be approaching that same level of significance.
This is not an argument against private enterprise. The speed of innovation over the past several years demonstrates the extraordinary capabilities of technology companies and entrepreneurs. Their responsibility, however, is to create value for shareholders and investors. Society’s responsibilities are necessarily broader. They include resilience, accessibility, energy policy, competition, environmental stewardship and the long-term public interest.
If the infrastructure supporting artificial intelligence is destined to become as fundamental as electricity or communications, should its development be guided primarily by a relatively small number of corporations? Is any single government capable of ensuring the right outcome when the supply chains, energy demands and strategic implications are global by nature? Or does infrastructure this consequential demand a level of international coordination that extends beyond national borders and commercial interests?
Those questions deserve more attention than they currently receive because the infrastructure decisions being made today will shape economic opportunity, national security and technological leadership for decades to come.
Perhaps that is the real lesson hidden within these two stories. They are not really about underwater servers or orbital data centers. They are about the extraordinary lengths humanity is already willing to consider in order to build the foundation of the AI age.
If we are prepared to reshape the oceans, the skies and perhaps even space itself to power artificial intelligence, then we should also be prepared to ask who that infrastructure ultimately serves. Projects this consequential should be built not merely for the benefit of a few men, but for the benefit of all mankind.
