Starcloud Is Moving the Data Center Off the Planet
For the last two years, the story of AI has been told in models. GPT this, Gemini that, a new benchmark every week.
But underneath the models is a far less glamorous story, and it's the one that's starting to matter more. Every frontier model is a claim on physical infrastructure: a building full of GPUs, drawing tens or hundreds of megawatts, cooled by rivers of water, wired into a power grid that was never designed for this kind of load. The models are software. What makes them possible is concrete, copper, and electricity.
And that infrastructure is hitting a wall.
Utilities are quoting multi-year waits for interconnection. Communities are fighting data center permits over water use and noise. Grid operators are warning that the pace of AI buildout is outrunning the pace of new generation. The constraint on artificial intelligence, increasingly, isn't the intelligence. It's the watts—and the land, water, and permits that turn watts into working compute.
That is the constraint Starcloud is attacking, and it's doing it from the most unexpected direction imaginable.
Starcloud is building data centers in space. Not as a thought experiment or a decade-out research program, but as flying hardware—GPUs in orbit, powered by the sun, cooled by the vacuum, running real workloads for real customers. The company's premise is deceptively simple: if energy and cooling are the true bottlenecks of the AI era, then the best place to put a data center may not be on Earth at all.
The Bottleneck Was Never the Chip
To understand why an orbital data center is anything other than science fiction, you have to reframe what a data center actually is.
We think of a data center as a computer. It's more accurate to think of it as a machine for converting electricity into computation and then getting rid of the heat. The GPUs are almost incidental to the hard part. Nvidia will happily sell you all the silicon you can afford. What's scarce is everything around the silicon: reliable power at gigawatt scale, water and infrastructure to carry away the heat, and permission from a grid and a community to plug it all in.
Those constraints are physical and they are getting worse. A single 40-megawatt cluster can consume more than a million tons of water a year for cooling. New AI campuses are being sized in the hundreds of megawatts and, increasingly, gigawatts—loads that rival small cities. On Earth, adding that capacity means new substations, new transmission lines, new generation, and years of environmental review, all in an environment of rising local resistance to data center sprawl.
Starcloud's thesis is that the industry has been optimizing the wrong variable. Everyone is racing to buy chips, when the thing that actually gates deployment is power and thermal capacity. And on those two variables specifically, space has an unfair advantage.
Why Space Is a Better Place to Put a Data Center
The case for orbit rests on two physical facts that terrestrial operators can't change no matter how much they spend.
The sun never sets in orbit. In the right orbit, solar panels receive near-continuous sunlight at roughly 1,361 watts per square meter—the raw solar flux at the top of the atmosphere, before clouds, weather, night, and atmospheric scattering take their cut. That's five to ten times the usable energy of even the best terrestrial solar sites. There's no night to store power through, no weather to hedge against, and no grid interconnection queue to wait in. Power generation stops being a negotiation with a utility and becomes a function of how many square meters of panel you can deploy.
Cooling is free, and it doesn't cost a drop of water. On Earth, getting heat away from dense GPU racks is a brutal, water-hungry engineering problem. In space, there's no water bill and no cooling towers—heat can be radiated directly into the near-absolute-zero background of space. A terrestrial data center fights thermodynamics with chillers, evaporation, and municipal water. An orbital one sheds heat the way the Earth itself does: by radiating it into the dark.
Put those two advantages together and you get a facility that generates its own power for free, cools itself for free, and answers to no zoning board. As Starcloud frames it, space offers continuous solar energy and radiative cooling that can scale toward gigawatts while sidestepping the permitting and resource constraints that increasingly bound growth on the ground.
That's the pitch. The reason it's a company and not just a white paper is that Starcloud has started to prove the hard parts are tractable.
From White Paper to Orbit in Under Two Years
Plenty of people have written that data centers in space would be nice. Starcloud is notable because it moved from idea to orbit with startup speed.
The company was founded in early 2024—originally as Lumen Orbit—by a team that reads like a deliberate assembly of the disciplines this problem requires: Philip Johnston, who came from McKinsey; Adi Oltean, a veteran of SpaceX and Microsoft Azure; and Ezra Feilden, formerly of Airbus Defence and Space. It's a combination of business, hyperscale cloud, and spaceflight engineering—the three worlds you'd have to fuse to build a data center that flies.
After publishing a white paper laying out the physics and economics, the team went through Y Combinator, where they raised one of the largest seed rounds in demo-day history. Then, remarkably fast, they flew.
In November 2025—just 21 months after founding—Starcloud launched Starcloud-1, carrying an Nvidia H100 into orbit. That single fact is a bigger deal than it sounds. The H100 is roughly 100 times more powerful than any GPU previously operated in space, where hardware has historically been generations behind the ground because of the harsh radiation and thermal environment. With that satellite, Starcloud became the first company to train a large language model in space, and the first to run a version of Google's Gemini in orbit.
The point of Starcloud-1 wasn't commercial revenue. It was proof. It demonstrated that modern, power-hungry AI silicon can be powered, cooled, and operated in orbit—that the physics survives contact with reality. In a category where skeptics reasonably ask "but does it actually work up there?", Starcloud answered by doing it.
The Hard Part Is Heat—and That's the Moat
It would be a mistake to pretend orbital data centers are easy. They aren't, and the hardest problem is exactly the one space is supposed to solve: cooling.
Radiating heat into space is free, but it is not fast. A two-sided radiator held near room temperature emits only a few hundred watts per square meter—orders of magnitude slower than the liquid cooling that pulls heat off AI chips on Earth. Scale that up and the numbers get demanding: a meaningful orbital cluster needs radiator surfaces measured in thousands of square meters, alongside equally large solar arrays. The engineering challenge of space data centers isn't generating power or finding a place to dump heat in principle—it's building deployable structures big enough, light enough, and reliable enough to do both at scale.
This is precisely why Starcloud is a serious company rather than a pitch deck. The thermodynamics are unforgiving, which means the team that actually masters large deployable radiators, high-efficiency solar, optical networking, and radiation-tolerant system design builds something very hard to copy. The difficulty is the defensibility. If putting an H100 in orbit were trivial, it wouldn't be worth a company. The moat is the pile of unglamorous spacecraft engineering that stands between a compelling white paper and a satellite that reliably runs Gemini overhead.
Starcloud's approach is to integrate all of it—solar generation, liquid-cooled compute, optical networking, and large deployable radiators—into a single orbital platform, and to iterate on that platform in flight rather than in simulation. Each satellite is both a product and an experiment that de-risks the next, larger one.
Starcloud-2 and the Customers Already On Board
The clearest signal that this has moved beyond demonstration is who is flying on the next satellite.
Starcloud-2 is slated to launch in October 2026, and it's a substantial step up: it will carry Nvidia's Blackwell B200 chips—current-generation, top-of-the-line AI silicon—and roughly 100 times the power generation of the first satellite. Crucially, it's not flying alone. Some of the largest names in computing are putting hardware and workloads on board.
Nvidia is supplying the Blackwell GPUs directly, and gains something valuable in return: real telemetry on how its most advanced silicon behaves under the radiation and thermal stresses of orbit—data that feeds its own space-grade roadmap. AWS is flying a production server blade to test parts of its custom-accelerator and Graviton roadmap in orbital conditions. Google Cloud is flying workloads. And Crusoe, the AI cloud operator, is using Starcloud-2 as a development platform for edge inference.
Read that list again. Nvidia, AWS, Google Cloud, and Crusoe are not companies that fly experiments on unproven infrastructure for fun. Their participation is the market's way of saying that orbital compute is worth taking seriously—that the hyperscalers themselves want a look at whether the next frontier of data center capacity is above the atmosphere. For an early-stage company, having that roster of partners on your second satellite is an extraordinary validation.
The Economics That Make It More Than a Moonshot
None of this would matter if the unit economics were hopeless. Starcloud's argument is that they're not—and that the crossover is closer than intuition suggests.
The comparison the company draws is between the lifetime cost of electricity on Earth and the up-front cost of launch and solar deployment in space. Over a typical five-year GPU lifecycle, a 40-megawatt data center's electricity bill alone runs on the order of $175 million at current power prices—and that's before land, water, cooling infrastructure, and the cost of simply waiting years for grid capacity. Starcloud's estimate is that launching and deploying an equivalent orbital system could cost on the order of tens of millions, with essentially no ongoing fuel or grid costs once it's up. The sun is free, and it keeps shining.
The variable that makes or breaks this math is launch cost, and it's moving in exactly the direction Starcloud needs. The per-kilogram cost of reaching orbit has fallen dramatically over the last decade and is set to fall further as heavy-lift vehicles mature. Orbital data centers are a bet not just on the physics of space power and cooling, but on the continuation of the launch cost curve—a curve that has been one of the most reliable deflationary trends in modern engineering. As launch keeps getting cheaper, the economics of putting compute in orbit keep getting better, while terrestrial power and water only get scarcer and more contested.
The Fastest Unicorn in YC History
The market has responded to that thesis with unusual speed. In March 2026, Starcloud raised a $170 million Series A led by Benchmark and EQT Ventures, at a valuation of $1.1 billion. That made it the fastest company in Y Combinator's history to reach unicorn status—just 17 months after demo day.
Early backing from Orange Collective, the YC-alumni fund that invests exclusively in Y Combinator companies, put Starcloud in a portfolio increasingly defined by infrastructure bets on the AI era—from identity verification to AI-native professional services to, now, the physical substrate of compute itself. It's a fitting addition. If the previous wave of portfolio companies asked how AI changes software, Starcloud asks a more elemental question: where the machines that run AI should physically live.
The speed of the fundraise reflects a shift in how investors are thinking about AI's constraints. A year ago, the scarce resource was models and talent. Today, it's power and land. Capital is flowing toward whoever can credibly expand the supply of usable compute—and a company proposing to tap the largest power source in the solar system, with no water bill and no permitting queue, is a direct answer to the question keeping data center operators up at night.
What Starcloud Is Really Building
It's tempting to file Starcloud under "audacious space startup" and move on. That undersells what's actually happening.
Starcloud isn't really a space company. It's an infrastructure company that happens to have concluded, correctly, that the binding constraints of its industry—power and cooling—are better solved off the planet than on it. The satellites are a means. The end is a new geography for compute: one where capacity is limited by how fast you can deploy solar panels and radiators, not by how long a utility keeps you in the interconnection queue.
If that vision holds, the implications ripple outward. The AI buildout stops being a zero-sum fight over terrestrial power and water. Data centers stop competing with homes and farms for the grid. And the ceiling on how much intelligence humanity can afford to run gets set by the output of the sun rather than the capacity of a substation in Virginia.
There's a long road between a satellite running Gemini overhead and gigawatts of commercial compute orbiting the Earth. The radiators have to get bigger, the launches cheaper, the reliability proven over years. But Starcloud has already done the thing that separates the credible from the imaginary: it put the hardware in orbit, turned it on, and ran the workload. The rest is engineering and time—and both are on its side.
The last decade moved computing from the basement to the cloud. Starcloud is betting the next one moves it to the sky.