in short
The real bottleneck for AI isn't chips or models --- it's electricity. What follows is a map of the three places the industry is now racing to generate it: on land, under the ocean, and in orbit.
The Real AI Bottleneck Is Power, Not Chips
For three years the AI boom has been told as a chip story. When that got stale, it became a model story — whoever trains the smartest system wins. Both versions miss the same thing. They treat compute as the scarce resource, when the real bottleneck sits one layer beneath it.
Every cluster, however advanced its chips or elegant its architecture, runs on electricity that has to come from somewhere. That question — where the power comes from — is now shaping how the next decade of AI infrastructure actually gets built. And it’s being answered in three very different places: on land, under the ocean, and in orbit.
Follow the Money: Why Hyperscalers Are Betting on Nuclear Power
The best evidence for this isn’t a forecast or a think-tank projection. It’s where the biggest technology companies are actually spending their money.
As of mid-2026, hyperscalers have committed close to 9.8 gigawatts of nuclear power across roughly thirteen separate deals — multi-decade contracts, several of them reviving reactors that had already been switched off. These aren’t loose bets from companies with cash to burn. They’re some of the most disciplined capital allocators alive, locking in commitments that stretch twenty years and more. Nobody signs a nuclear contract on that timeline for a problem they expect a better chip to solve.
And every dollar spent restarting a reactor, financing offshore wind, or filing satellite plans with a regulator is a dollar not spent on the next generation of chips — the thing most coverage still treats as the binding constraint. If chips were the bottleneck, the money would follow the chips. Instead it’s following power: the plain physical ability to generate electricity at gigawatt scale. The industry is ready to generate intelligence. It is not yet ready to generate the power that intelligence needs. That gap — not any shortfall in model capability — may be the defining miss of this generation.
Earth: Powering Data Centers with Nuclear and Gas
The most obvious place to put a data center is still the ground under your feet, and for now that’s where most compute lives. But the terrestrial build-out is straining. Hyperscaler capital spending has climbed to a pace that increasingly leans on bond issuance and project-level financing rather than cash on hand — Meta structured roughly $27 billion of its Louisiana build-out through Blue Owl specifically to keep the debt off its balance sheet (Fortune). When the best-funded companies in the world start financing like that, it tells you how hard they’re stretching.
Nuclear Deals Reviving Shuttered Reactors
The nuclear deals are the clearest tell. Microsoft signed a twenty-year, $16 billion agreement with Constellation to restart a reactor at the former Three Mile Island — now rebranded the Crane Clean Energy Center — bringing a shuttered plant back online to serve data center demand, with power targeted for 2027 (NGA Nuclear Dispatch). Meta went further, signing its own twenty-year deal with Constellation for the full output of the Clinton Clean Energy Center in Illinois. These aren’t contingency plans. They’re primary strategy, and they mark a depth of vertical integration into energy markets that technology companies never used to need.
Natural Gas as a Bridge Fuel
Not all of it is nuclear. To power Hyperion, its enormous campus in Richland Parish, Louisiana, Meta struck a deal with the utility Entergy to build ten gas-fired plants — and the project has since ballooned to five gigawatts of compute and a total power footprint north of seven (Fortune). Gas is the fast, contentious option: it trades a cleaner long-term energy mix for near-term speed, run as a bridge while slower, firmer power comes online.
The Grid Bottleneck and "Power-Ready" Land
Speed is the whole game, because the scarcity shows up first as time. In many regions a new large power user waits years simply for permission to connect to the existing grid; in the hottest markets, utilities have quoted interconnection timelines stretching a decade. So the market has started pricing power access straight into the dirt. Sites with a confirmed, funded connection — what the industry now calls “power-ready” land — command a premium of ten to fifteen percent over otherwise identical sites still stuck in the queue (JLL, 2026 Global Data Center Outlook; Data Center Frontier). That premium is a clean measurement of how thoroughly power access now outweighs location, size, and every other traditional real estate variable.
Policy Responds: The Ratepayer Protection Pledge
Governments have noticed. In March 2026 the White House rolled out a Ratepayer Protection Pledge, with the largest hyperscalers committing that their data center demand wouldn’t push up electricity bills for ordinary households. The pledge itself is the story: the collision between AI infrastructure and the public power supply is no longer hypothetical. It’s here, and it’s big enough that the federal government wrote policy to manage it — a pledge that by mid-2026 had expanded to cover most of the power delivered to American homes and businesses.
Of the three frontiers, Earth is the most developed, the most capital-intensive, and the most politically visible — which is exactly why it’s the most constrained. And there’s a pattern here worth holding onto, because it repeats: density and limitation move together. The more established the ground beneath a technology, the harder that ground pushes back. This isn’t really a chip race, and it isn’t quite an energy race either. It’s a contest of elements — earth, water, and orbit — each striking a different bargain between how developed a resource already is and how much room it has left to grow.
Sea: Underwater Data Centers Cooled by Seawater
While American and European companies negotiate nuclear contracts, a quieter and frankly more elegant answer has been running commercially off the coast of China.
China's Undersea Data Center Off Shanghai
If Earth’s problem is density — too much infrastructure chasing too little power, and too little patience from regulators and neighbors — the ocean offers the opposite: vast, largely unclaimed thermal capacity with almost none of the political weight of land. About ten meters down, near Shanghai’s Lingang coast, a data center sits co-located with offshore wind turbines that feed it power directly. Seawater does the cooling — no chillers, no evaporative towers, no fresh water consumed at all. The facility reportedly cuts electricity use by close to twenty-three percent against an equivalent site on land, while eliminating water consumption entirely. In a world where data center water use has become its own local flashpoint, that last part matters more every year.
Microsoft's Project Natick: Proven, Then Shelved
The idea isn’t new. Microsoft ran a version of it years ago as Project Natick, sinking a sealed data center off the coast of Scotland to test a simple premise: cold ocean water is an excellent, free coolant, and the seabed is a stable, tamper-resistant place to keep hardware. Natick worked. Then it was shelved. The honest question is why one version of this became a research footnote and the other became a commercial facility running live workloads — and the answer isn’t the technology. Natick was built as an experiment, meant to prove a point and then be pulled back up. China’s facility was built from day one as commercial infrastructure, wired to a power source and meant to run indefinitely. The differentiator wasn’t the engineering. It was the commitment.
Wave Power, Wind, and the Economics of Free Cooling
Others are chasing adjacent versions of the same logic. Panthalassa, backed by a $140 million round, is building floating, wave-powered nodes that generate electricity and run compute inside a single ocean-going unit. Aikido has unveiled a floating structure that fuses an offshore wind turbine directly with AI-grade hardware — power source and data center as one object. None of this yet approaches the scale of the land build-out, and that’s precisely the point: it isn’t a novelty. It’s a proven, operating alternative that barely registers in the Western conversation about where compute goes next.
Cooling earns this much attention for one reason: arithmetic. In a conventional data center, cooling can eat a large share of total energy — power that runs not a single calculation, spent purely to keep the machines from cooking themselves. Solve cooling for free with ambient seawater and you don’t get a minor efficiency bump. You get a structural cost advantage that compounds over the twenty- or thirty-year life of the facility. Which is exactly why it deserves more attention from allocators than it currently gets.
Space: The Race to Build Orbital Data Centers
If the sea is the quiet, proven alternative, orbit is the loud, unproven one — and it deserves to be treated as genuinely open, neither inevitable nor absurd. By the logic of elements, it’s the least dense and least constrained of the three: no grid, no seabed, no neighbors to object. That absence of friction is exactly what should make a careful reader wary rather than sold. The environment that’s easiest to build in on paper is also the hardest to reach and the least tested in practice.
What SpaceX's S-1 Reveals About Orbital Compute
The strongest sign that this is serious, and not a marketing exercise, comes from an unusually credible place: a regulatory filing. SpaceX’s S-1 states a goal of launching one hundred gigawatts of compute into orbit every year (SemiAnalysis). An S-1 is a disclosure document, written under legal obligation for investors and regulators — not a press release built for headlines. Stating a number there carries a different kind of weight than announcing it at a product event.
Who's Building Data Centers in Space
The momentum is real, and it’s broad. Google is exploring TPU clusters in orbit through a research effort it calls Project Suncatcher — and is reportedly in talks to help SpaceX with its space-based plans as well. Starcloud has filed for a constellation of as many as 88,000 satellites (Via Satellite). Orbital Compute, a months-old startup, has asked the FCC for permission to deploy up to 100,000 satellites of its own (SpaceNews). NVIDIA has begun actively promoting hardware — its Space-1 module — built specifically for orbital compute (NVIDIA Newsroom). None of this is podcast musing. These are filings, product launches, and infrastructure plans from companies with the balance sheets to pursue them.
The Unsolved Physics and Economics of Orbital Compute
And yet the case for orbital compute is genuinely unsettled, and it should be shown that way. On Earth, heat leaves hardware through air and water — convection and conduction, both efficient, both free. In the vacuum of orbit, neither is available. Heat can only escape by radiation, a far slower process to engineer at the densities modern accelerators demand. No one has yet demonstrated this working commercially, at scale, across years of continuous operation. Launch cost is the second open question. It has fallen dramatically over the past decade, and reusable rockets are the only reason the conversation exists at all — but “falling” is not “solved,” and a recurring launch bill is a cost structure no terrestrial or undersea data center has to carry.
There’s a subtler trap underneath the technical ones. The fact that communities on Earth increasingly resist new data centers is not, on its own, evidence that space is the rational economic alternative. Those are two different arguments, and sliding from one to the other is the weakest move in this debate: an emotional case for leaving Earth is not a financial case for the unit economics working in orbit (CNBC). A careful reader should reach the end of this section without a settled view on whether orbital data centers become real infrastructure or a well-funded dead end. That uncertainty isn’t a hole in the reporting. It’s an honest picture of where the evidence sits.
Who's Building AI Data Centers --- and Who's Fighting Them
The builders are mostly the companies already named — the hyperscalers supplying both the capital and the demand — alongside vertically integrated outliers like SpaceX. Having absorbed xAI in a $1.25 trillion merger in February 2026 (CNBC), SpaceX is now, all at once, a launch provider, a satellite operator, and the operator of some of the largest AI compute clusters on Earth, with orbital ambitions stacked on top. Behind the builders sit the enablers: nuclear and small-modular-reactor developers, offshore wind operators, and — for floating and submerged sites — marine classification societies now being asked to certify a class of asset they’ve never had to evaluate.
Governments are no longer spectators. The Ratepayer Protection Pledge is one piece of policy written for this exact moment, and it won’t be the last. Expect more targeted intervention as the collision between AI demand and public utilities becomes a permanent fixture of energy policy rather than a passing spike.
Capital is moving the same way. Infrastructure funds and, increasingly, sovereign wealth vehicles have started treating power-and-compute as a single bundled asset class rather than two separate sectors — buying land, energy contracts, and data center capacity together instead of underwriting them one at a time. That bundling is its own signal. When long-horizon institutional money stops separating “energy” from “technology,” it usually means the market has already decided they’re the same trade.
It would be dishonest to call any of this frictionless. Community resistance has hardened into a genuine, recurring force: by one count, at least seventy-five projects were blocked or delayed by organized local opposition in the first quarter of 2026 alone (NRDC) — roughly as many as in all of the previous year — over water use, noise, air quality, and strain on the grid. That resistance is a real and serious thread in this story, even if it isn’t the focus here. And it’s the reason all three of these paths exist at all. The easiest option of them all — quietly expanding on Earth, without consequence — simply isn’t available anywhere anymore.
Earth, Sea, or Space: Where Should AI Capital Go?
None of this resolves cleanly. Earth is proven, but dense, strained, and increasingly contested. Sea is proven, quiet, and — outside China — barely in the conversation. Space is well-funded and seriously pursued, and still an open question on the physics and the economics alike. Each element trades development for room to grow, and none has yet found a way to have both.
If the real constraint on the next decade of artificial intelligence was never intelligence at all, but the far older problem of earth, water, and air — which of the three would you put ten years of capital behind?
source
covers
~9.8GW/~13 deal aggregate
Microsoft/Crane Clean Energy Center restart
Meta/Entergy Hyperion gas-plant buildout
Meta/Blue Owl $27B off-balance-sheet financing
“Power-ready” land premium (10–15%)
China’s Lingang underwater data center
Panthalassa, Aikido floating platforms
SpaceX S-1’s 100GW/yr orbital compute goal
SpaceX–xAI merger ($1.25T, Feb 2026)
Google/Starcloud/Anthropic orbital momentum
Orbital Compute’s FCC filing
NVIDIA’s Space-1 module
Skeptical case on space-based data center economics
Community opposition count (Q1 2026)
