Data Centers: From Mainframe Rooms to Orbit, and Why Land Still Wins Through 2040

How data centers are evolved, the underwater and orbital frontiers China and the US are both racing toward, and why the $700 billion build-out still points investors back to land.

Prepared by Richstorm.co

Key Takeaways

• Data centers have moved through three eras: centralized mainframes (1950s-1980s), distributed client-server and internet-era facilities (1990s-2000s), and hyperscale cloud campuses (2010s-present), each driven by a shift in what computing was being used for.

• The current era is defined by AI training and inference, workloads far more power- and cooling-intensive than prior generations, with the five largest US hyperscalers planning roughly $660 to 725 billion in 2026 infrastructure spending alone, nearly double 2025 levels.

• China currently leads in both alternative locations: an operational underwater data center since 2023 cutting cooling costs by 90 percent, and 24 satellites already running real AI models in orbit as part of a stated path to 2,800 satellites by 2030.

• The US holds an edge in orbital compute hardware, having flown the most powerful individual GPU yet operated in space, but China's deployed satellites already demonstrate working AI applications at smaller scale.

• Even under each program's own stated 2030 to 2040 targets, underwater and orbital capacity combined would represent well under one percent of total AI data center capacity, with orbital scale-up additionally constrained by a cooling problem the industry itself describes as having no proven solution.

• For investors, the clearest and largest exposure remains land-based capacity, hyperscalers, data center REITs, and the chip supply chain, where growth is measured in hundreds of thousands of additional megawatts through 2040, not the alternative locations, however technologically compelling they are.

The Past: From a Room to the Cloud

The earliest data centers were not built for businesses at all. The first facility built to house a computer, ENIAC, went up in 1945 at the University of Pennsylvania, with similar facilities following for defense and intelligence purposes through the 1950s. Even at this stage, the core physical challenge that defines data centers today was already present: these machines generated enormous heat and required dedicated cooling and airflow from the start.

The 1960s and 1970s established what most people picture as the classic data center: centralized mainframes from IBM and similar manufacturers, housed in expensive, climate-controlled rooms, accessible only to large corporations, universities, and governments. IBM's System/360 allowed organizations to scale up to more powerful machines while keeping the same software, establishing mainframes as the backbone of business computing for payroll, banking, and airline reservations.

The 1990s brought decentralization, the personal computer revolution and client-server networks meant computing power was no longer concentrated in a single room, though organizations still needed central facilities to house servers, giving rise to the modern term "data center." The dot-com boom drove construction of much larger facilities even as the bubble itself crashed by 2000-2002. The defining shift of the 2000s and 2010s was cloud computing: Amazon Web Services began offering compute and storage as a service in 2006, creating the economic logic for hyperscale facilities, often exceeding a million square feet, run by a small number of companies that rent capacity to everyone else.

The Present: AI Has Changed the Economics, and the Geography

The current era is defined by a single new workload, AI training and inference, which consumes power and generates heat at a scale prior cloud workloads did not.

Estimates for combined 2026 capital expenditure from the largest hyperscalers cluster in the $660 to 725 billion range, an increase of roughly 60 to 80 percent over 2025's already-record spending of around $400 billion. Company-level guidance includes Amazon at roughly $200 billion, Alphabet at $175 to 190 billion, Meta at $115 to 145 billion, Microsoft at $110 to 120 billion or more, and Oracle around $50 billion, with roughly 75 percent of this spending directed specifically at AI infrastructure.

Capital intensity for these companies has reportedly reached 45 to 57 percent of revenue, levels described by industry analysts as historically unprecedented, and US data center construction spending reached a monthly rate of over $45 billion by the end of 2025, up 85 percent from two years prior. This spending surge has sparked visible community pushback over electricity rates and water consumption, with more than two-thirds of planned US facilities sited in rural areas partly as a result. This local resistance, and the underlying power and water constraints behind it, is the direct motivation for every alternative-location proposal now in development.

The Race to Alternative Locations: Underwater and Orbital

Two alternative locations have moved from concept to real deployment in just the past few years, and in both cases China currently holds the operational lead, while the US holds an edge in specific technology components.

Underwater data centers are now commercially operational in China. A facility off Hainan, launched by China National Offshore Oil Corporation, has been operating since 2023 and was recently upgraded to handle AI workloads, processing data, running cloud services, and training AI systems using seawater cooling that reduces cooling energy costs by roughly 90 percent and achieves a server failure rate below 1 percent due to the stable, oxygen-free environment. A second, larger facility off Shanghai, powered by an adjacent offshore wind farm, achieves over 95 percent renewable power. Both facilities remain small relative to gigawatt-scale land facilities, the Shanghai site operates at 24 megawatts, but represent the only commercial-scale underwater computing in the world. Microsoft's earlier Project Natick demonstrated the same concept years ago but has not moved to commercial deployment, and no comparable Western facility currently operates.

Orbital data centers have also moved from concept to real, if small, deployment, on both sides. China's Three-Body Computing Constellation launched its first 12 satellites in May 2025, each delivering about 744 trillion operations per second, and has since run real AI models in orbit, including an 8-billion-parameter model that surveyed 189 square kilometers of terrain in northwest China and Alibaba's Qwen-3, one of the first general-purpose AI models operated in space. A second Chinese program, Adaspace's "AI Cloud" constellation, launched its first 12 of a planned 2,800 satellites in May 2026. On the US side, Nvidia-backed Starcloud launched a single satellite carrying an H100 GPU in November 2025, becoming the first to train a large language model in space, a chip described as 100 times more powerful than anything previously operated in orbit. Days before its IPO in June 2026, SpaceX unveiled its own AI1 concept, orbital server racks cooled by radiating heat into space, with prototype satellites planned for 2027.

Both countries have stated remarkably similar long-term ambitions: China's 15th Five-Year Plan (2026-2030) explicitly calls for "gigawatt-class space-based digital and intelligent infrastructure," while SpaceX's Elon Musk has said space will become "the lowest-cost place to put AI" within a few years, citing roughly five times more solar power availability in orbit due to the absence of clouds and night cycles.

The Thermal Wall

The gap between today's deployments and either country's stated targets comes down to a single, well-defined physical constraint. AI training clusters on Earth require 10 to 50 megawatts of power. China's operational Three-Body satellites each produce roughly 744 trillion operations per second with an estimated 500-watt thermal load, a gap representing a 100 to 1,000 times scaling challenge that the industry itself describes as having no proven solution. Radiative cooling in the vacuum of space, dumping heat via radiation alone, is over 1,000 times slower than the water cooling used in land-based AI data centers, since there is no air or water in orbit to carry heat away.

China's current approach is built around this constraint rather than against it: the Three-Body constellation uses passive radiators sized for sub-kilowatt heat rejection, and its path to 2,800 satellites by 2030 maintains this distributed, lower-power-per-satellite approach rather than concentrating megawatt-scale computing on individual satellites. This means the 24 operational satellites demonstrate that small, low-power AI workloads function in orbit, a real and useful result for applications like onboard image processing, but do not demonstrate a path to the gigawatt-scale training infrastructure both nations describe in their long-term plans. SpaceX's own framing of AI1 is candid on this point: whether orbital compute is actually cheaper than land-based infrastructure is, in the company's own words, the question the whole vision rests on, and the one it has not yet answered.

Underwater faces a different and more tractable challenge. Scaling from 24 megawatts to gigawatt scale is largely a manufacturing and deployment problem, building and submerging more modules of a design that already works, rather than an unsolved physics problem. This makes gigawatt-scale underwater capacity by the early-to-mid 2030s a plausible engineering trajectory, even if a specific date remains uncertain, in a way that orbital's 2030 target currently is not.

Where Things Stand: A Snapshot

 AI data center capacity by location, US vs. China, 2026-2040

What This Means for Investors

Even taking each program's own stated ambitions at face value, underwater and orbital capacity combined would represent on the order of 1,700 megawatts by 2040 against land-based capacity in the hundreds of thousands of megawatts, well under one percent of total AI data center capacity. The alternative-location lines are visually indistinguishable from zero across the entire period.

This points the clearest investment exposure back to land: the hyperscalers themselves, dedicated data center operators and REITs, and the chip and infrastructure supply chain feeding hundreds of billions of dollars in annual, already-contracted capital expenditure. This spending is real and already showing up in reported results, even if its eventual returns remain debated, hyperscaler free cash flow is falling sharply as a direct result, and analyst opinion ranges from confident to cautious.

The underwater and orbital stories are genuinely worth following as technology, China's operational facilities represent real, working firsts, and the pace of announcements from SpaceX, Google, Nvidia, and Chinese state-backed programs suggests this area will continue generating news. But the pattern is now a familiar one across this publication's recent coverage: a real, technically impressive proof-of-concept generates attention disproportionate to its current, or even fifteen-year-projected, share of the underlying economic activity. For a portfolio, the distinction matters: these are technologies to watch, not yet capacity to underwrite.

Figures reflect hyperscaler earnings guidance, Chinese state media and industry reporting, and company announcements current as of mid-2026. Capacity figures for 2030-2040 are illustrative, based on stated plans, not independently verified forecasts, and are subject to substantial change.