The cloud is moving to orbit. Here is what is real so far.

#ai #cloud #space

On 1 October, Google put four of its AI chips into orbit on a satellite about the size of a refrigerator. Two weeks later, Nvidia's CEO said his engineers are designing a new kind of server with SpaceX, meant to run in space. For software developers, the cloud they deploy to may one day include machines that are not on the ground at all. This post looks at what is real today, who gains from it, and what the next step looks like.

What is real comes first. In November 2025, a startup called Starcloud launched Starcloud-1, a satellite of about 60 kilograms carrying one Nvidia H100 GPU. It was the first data center class GPU to run in orbit, and the team used it to train a small language model in space [1][2]. On 1 October 2026, Google's Project Suncatcher prototype followed with four TPUs. Because of heat limits, those chips run Google's open Gemma model for 15 minutes at a time, and the mission is planned to operate for a year [3].

That is two satellites and five data center chips.

Then there is what has been proposed. In February, the FCC accepted SpaceX's application for up to one million orbital data center satellites [4]. Starcloud filed for up to 88,000 on 3 February [2], and a young startup called Orbital filed for up to 100,000, with 10 gigawatts of computing power [5]. Together, those three filings ask for about 1.19 million satellites, by my own count.

Between the two sit the designs. In June, SpaceX showed AI1, its first data center satellite design, at about 150 kW peak and 120 kW sustained, with a 70 metre wingspan [5]. A Google research paper published in October says each data center satellite may need 50 to 100 kW, roughly what one rack draws on Earth [6]. On 15 October, Jensen Huang said Nvidia and SpaceX engineers are designing a new server together, one that could run in space and on Earth [7].

So the honest picture is this. A few chips are flying, the designs are serious, and the filings are requests for permission, not proof that anything works at scale. The companies say so themselves. SpaceX's IPO prospectus calls orbital AI compute a very hard technical problem [8], and the lead of Project Suncatcher does not expect space compute to beat ground data centers on cost within five years [9].

Every new place to run software creates new work. The people closest to the hardest constraint usually gain first, and in orbit that constraint is power.

Developers who write lean software come first. In orbit, every watt has to be collected by a solar panel and every bit of waste heat has to leave through a radiator. Code that does the same work with less compute is worth more there than anywhere else. The model running on Google's prototype is Gemma, a small open model [3]. Engineers who know how to make models smaller and faster gain a place where that skill is the whole job.

Electronics, thermal and power engineers gain too. Orbital is hiring engineers for thermal, power, mechanical and electrical work [10]. Starcloud's next satellite is planned to carry the largest commercial deployable radiator flown so far [11]. Hardware that survives radiation and sheds heat without air is becoming its own product category.

For infrastructure and reliability engineers, the job changes shape. A technician cannot walk up to a server in orbit and replace a part [3]. Systems that report clearly, recover on their own and move work away from a failing node become the core of the job. Teams that already run large fleets through automation start with an advantage.

In large scale engineering, the cost of every compute cycle is visible. In my own work on systems at that scale, I saw decisions that were perfectly acceptable on a small system become a clear no once scale came in. Precomputing a new data point for every customer, back through all their history, is fine when the data is a few gigabytes. When the data runs to exabytes, the better move is to first ask which customers actually need it.

Teams whose data is born in space gain in a very direct way. Earth observation satellites can produce hundreds of terabytes of raw data a day, more than the links to the ground can easily carry [12]. Processing that data in orbit means sending down answers instead of raw images. Starcloud planned to use its first satellite to analyse radar data from Capella Space for exactly this reason [13]. Farmers, shipping companies and city planners who rely on satellite data stand to get their answers sooner.

Satellite operators gain as well. Starlink satellites perform a collision avoidance maneuver about every two minutes on average, and most of the processing behind them still happens on the ground [12]. More compute on board lets those decisions happen where the satellite is.

Founders and small teams have room here too. Starcloud designed, built and launched its first satellite in 21 months on 3 million dollars of pre-seed money, and in August it raised again at a 2.3 billion dollar valuation [14]. Big companies will build the satellites. Small teams can build the layer around them: scheduling, monitoring and testing for compute that nobody can touch.

If I had to name who gains most, it is two groups: people whose data already lives in orbit, and developers who treat compute as something scarce.

In 2018, Microsoft sealed 864 servers inside a steel tube and lowered it to the seafloor off the Orkney Islands. When the tube came back up in July 2020, those servers had failed at one eighth the rate of a matching group on land [15]. Microsoft did not go on to build undersea data centers, but it kept the lessons for its research on reliability [16].

I am not trying to predict what orbital data centers will become. What I take from Project Natick is simpler. Putting compute somewhere new teaches engineers things they could not learn at home, even when the place itself does not become the new normal.

The next step is not a million satellites. It is small and specific. Google plans to launch two satellites in 2027 to test whether they can share work over laser links between them [17]. Starcloud plans a second satellite with its large radiator and paying customers on board [11]. These are plans, not achievements, and plans in space often move.

There are real hurdles. Heat has to leave through radiators, launch still costs a great deal, and nobody can drive over to fix a broken server. People have cleared hurdles like these before, one experiment at a time. That is the question this series keeps asking: what changes for engineers when orbit becomes an ordinary place to run software? The answer starts with satellites like the one Google launched four weeks ago.

For now, running compute in orbit costs far more than running it on the ground. One analyst's rough model puts the cost of launching and running a GPU in space for a year at ten times or more the cost on Earth [12]. The clearest early customers are the jobs that need compute next to the sensor: processing Earth observation data in orbit and helping satellites avoid each other. For a small team, the opening is the software that makes remote, power-limited compute easy to use.

Companies are looking at orbit because compute on Earth is scarce and power hungry. The same scarcity shows up on every developer's bill as tokens and GPU hours. An AI coding tool that reads the whole repository for every question, or retries a failing call without limits, spends compute nobody planned for. The tools that last will treat every call as having a cost and will show that cost before they spend it.

The picture changes again once humans live and work beyond Earth. Communication between planets already depends on relays. On Mars, the rovers send most of their science data up to orbiters, which pass it on to Earth, because each rover loses its line of sight to Earth for about 12 hours a day. A one-way signal between the two planets takes between 3 and about 22 minutes [18]. In September, NASA chose Blue Origin, under a contract worth up to 700 million dollars, to build a dedicated Mars telecommunications orbiter, due for delivery by the end of 2028 and planned to operate at Mars by 2030 [19]. Around the Moon, ESA's Moonlight programme plans five satellites for communication and navigation, with first services in late 2028 [20].

A relay that can also compute is a natural next step from there. With minutes of delay on every message, a question asked on Mars cannot wait for an answer from Earth. A data center sitting in orbit next to the relay could process data where it arrives, answer local questions on the spot, and send only the results across the long link home. Seen that way, the cost question also looks different. On Earth, an orbital data center competes with a cheaper building on the ground. For a crew on Mars or a base on the Moon, the comparison is with having no nearby compute and no fast link at all. When communication becomes something a community depends on, the price of building that infrastructure is weighed against the value of staying connected.

Sources

The 1.19 million total and the five chip count are my own sums of the figures below.

DEV Community

The cloud is moving to orbit. Here is what is real so far.

Sources

Top comments (0)