The data center is leaving the planet.

A refrigerator-sized box, 350 miles up, doing math

Somewhere over your head right now, a satellite about the size of a small fridge is running an Nvidia H100. It is not taking photos of weather systems or relaying phone calls. It is training artificial intelligence. In November 2025, that box - Starcloud-1 - became the first machine to train a large language model entirely in space. Then it sent a note home: "Greetings, Earthlings. Or, as I prefer to think of you - a fascinating collection of blue and green."

Starcloud is the company that put it there. Three years old, around twenty-six people, headquartered in Redmond, Washington, and recently worth $1.1 billion. They are not building a better satellite. They are arguing that the data center - the windowless, water-hungry, power-starved warehouse that quietly runs the modern internet - belongs in orbit. It is a strange thing to believe. They have a satellite that suggests they might be right.

AI is hungry. Earth is running low on outlets.

Here is the inconvenient arithmetic. Every new frontier AI model needs more compute, which needs more chips, which need more power and more cooling. On Earth, all three are getting harder to find. Data centers fight for grid connections that take years to permit. They drink water by the millions of gallons to stay cool. They sit on land that towns increasingly do not want them on.

The energy is the real wall. A single hyperscale AI campus can demand as much electricity as a mid-sized city, and the grid was not built with that in mind. You can build the chips faster than you can build the power plants to feed them. That is the tension Starcloud exists inside: the appetite is exponential, the infrastructure is not.

Three founders, one contrarian premise

Starcloud began in early 2024 in El Segundo, California, under the name Lumen Orbit. (A trademark scuffle with Lumen Technologies later forced the rename - the new one is better anyway.) The founders are an unusual trio. Philip Johnston, the CEO, came from McKinsey, where he worked on satellite projects for national space agencies, and arrived with degrees stacked from Harvard, Wharton, Columbia, and a CFA charter. Ezra Feilden, the CTO, came from Airbus Defence and Space. Adi Oltean, chief engineer, came from SpaceX and Microsoft Azure - which is to say, he has built both rockets and clouds.

Their bet rests on three lines on a chart that all point the same way: launch costs are falling (thank you, reusable rockets), GPUs are getting more efficient, and demand for AI compute is climbing nearly vertically. Cross those trends and a once-absurd idea becomes a spreadsheet. In orbit, the sun shines almost continuously - no night, no clouds, no weather - so solar power is abundant and free of fuel. And space is already cold, so you can dump heat by radiating it away, no cooling towers, no rivers. The pitch is less science fiction than power arbitrage.

Two satellites, one very long ladder

Starcloud-1 is deliberately humble: roughly 130 pounds, one H100 GPU, solar panels, radiation shielding, and a cooling system borrowed from the heat-rejection hardware that keeps the International Space Station livable. It was never meant to be a business. It was meant to prove that an off-the-shelf data center chip - one a hundred times more powerful than any GPU previously flown - could survive radiation, run, and do useful work in orbit. It did. It trained an open-source model written by Andrej Karpathy and held a conversation with the ground.

Starcloud-2 is where ambition shows. Slated to fly before the end of 2026 and funded largely by the Series A, it is designed to generate roughly a hundred times the power of its predecessor and to carry Nvidia's Blackwell B200, one of the most capable AI chips money can buy. After that, the company talks about Starcloud-3 and -4, then gigawatt-class clusters - the kind of scale that, on Earth, would need its own power station. In orbit, it needs a bigger solar array. The constraint that strangles terrestrial data centers simply does not apply.

The money is forward-looking. The hardware is not.

Skeptics will note - correctly - that Starcloud has no commercial customers yet, and that a $1.1 billion valuation for a company with a refrigerator in orbit is a bet on the future, not a reward for the present. Fair. But the present is not nothing. The H100 worked. The Series A, led by Benchmark and EQT Ventures, was heavily oversubscribed and pulled in Macquarie Capital, NFX, Y Combinator and a cast of angels that includes former Boeing CEO Dennis Muilenburg and former Starbucks CEO Kevin Johnson. Earlier backers included In-Q-Tel, the CIA's venture arm - which tends to fund things governments find useful.

Move the warehouse, save the watershed

Strip away the rockets and the mission is oddly practical: stop straining Earth's grids and rivers to feed the AI boom. If the heaviest compute lives in orbit - drawing on solar power that would otherwise go nowhere and shedding heat into a vacuum that costs nothing - then the data center stops competing with cities for electricity and with farms for water. Starcloud frames this not as a moonshot but as the obvious destination once the trend lines cross. Sustainability, here, is a side effect of economics rather than a marketing department.

A new space race, with cooling fans

Starcloud is not alone in noticing this. Google has floated its own orbital-compute concept, and the giants of cloud will not cede the sky quietly. That is the clearest sign the idea has graduated from fringe to plausible: the incumbents are looking up. The open questions are real - radiation hardening at scale, servicing hardware you cannot drive a truck to, the economics of launching gigawatts of silicon. None of them are settled. All of them are now worth arguing about, which they were not three years ago.

Back to the box over your head

So look up again. The refrigerator-sized satellite is still there, still running its H100, still doing math no machine had done in orbit before it. A year ago that box was a science experiment. Now it is the first rung of a ladder a billion dollars says leads somewhere. Starcloud has not yet moved the data center off the planet. It has done something narrower and harder to dismiss: it has made the question serious. The next time the grid groans under another AI campus, someone in the room will say "what if we built it up there" - and they will no longer be joking.