AI in orbit: why Elon Musk and Sam Altman want to build data centers in space
One of the hottest technology topics in December was data centers - data centers - in space. Suddenly, just about everyone is talking about it: Nvidia, Google, Amazon, a number of startups, and the ubiquitous Elon Musk. Let's see what makes the idea of orbital data centers so attractive and whether it will be possible to make money on it in 2026.
Dyson's sphere and Kardashev's scale
"And the entire Galaxy... began to exploit the energy of the stars with unimaginable activity. Ma not only that each solar system was surrounded by a dense network of light traps, accumulating "escaping" solar energy in order to use it wisely (the galaxy now shone less brightly) - many stars that were not suitable for the role of suns, were destroyed, and their vast reserves of intra-atomic energy were put to work" - wrote in 1937 Olaf Stapledon in his fantastic novel "The Star Maker". Nearly a quarter of a century later in 1960, the prominent Anglo-American physicist Freeman Dyson proposed the idea of detecting extraterrestrial civilizations. In his opinion, a highly advanced civilization will sooner or later build a sphere (or, rather, a swarm of objects) around its star to more fully capture its light energy. As a by-product of this activity, the sphere, unrelated to the nature and construction of these objects, will emit excessive heat in the infrared range, and this radiation can be detected by human means of space observation.
Later he supplemented his reasoning with the assumption that the matter for the creation of such a sphere could be obtained by "disassembling" a large planet about the size of Jupiter. And he admitted that he got the idea from The Star Maker. Such a sphere was called a Dyson sphere.
If you think it's all pure theory, it's not - in 2024, 7 objects have been detected within 1,000 light-years of earth glowing in a certain infrared spectrum. "These could be Dyson spheres because they behave as our models predict, but it could also be something else," CNN quoted the study's authors as cautiously commenting.
In 1964, Soviet astrophysicist Nikolai Kardashev proposed a scale of development, according to which civilizations of the first type use the energy of their planet, the second - a star, and the third - an entire galaxy.
According to some sources, humanity has not even reached the first type yet: our energy consumption according to Kardashev's advanced scale is estimated at about 0.73. However, we already need energy, and a lot of it. It looks like it's time to aim for level two.
When AI is stymied by an outlet
The frenzied race to build more and more data centers, spurred by artificial intelligence, is fueling not only the shares of chipmaker Nvidia, but also generating an increasing demand for the electricity needed to power as well as cool huge new server farms, according to a McKinsey study. They estimate that current data center demand of 82 gigawatts of energy will more than 2.5 times increase in 2030 - and that's under a baseline, not aggressive, development scenario.
Analysts write that existing power grids can no longer cope. Additional solutions include gas turbines and nuclear power, but these come with their own challenges, including regulatory hurdles and the need for significant capital investment.
Meanwhile, we are literally swimming in a vast ocean of energy, and it is free. This is the energy of thermonuclear fusion of our star, the Sun. The power needed by data centers in five years is equivalent to 219 billion watts. At the same time, the Sun radiates about 386 trillion trillion trillion watts into space (twice "trillion" is not a typo).
The sun is a huge free fusion reactor in the sky. Creating tiny fusion reactors on Earth is utter stupidity... Stop wasting money on pathetic little reactors, unless of course you recognize that you only need them for your pocket science project, for crying out loud
One of the main advocates and practitioners of building huge data centers, OpenAI CEO Sam Altman, has promised to invest "trillions of dollars" in them. However, he understands the attendant problems - it requires taking up large tracts of land, providing huge electricity and water consumption (for cooling), and doing something about the noise that disturbs local residents, Wired wrote in September. Altman believes that data centers could be placed in space.
The advantages are obvious - a sea of free solar energy, besides there is no cloud cover, no disgruntled neighbors, no need to buy land and obey endless bureaucratic rules of governments. The latter is especially important, says "the godmother of space law," lawyer and political scientist Maison Hanlon in an interview with Wired:
If you are a U.S. company looking to put data centers in space, the sooner the better. Before Congress says, "Oh, we need to regulate that."
A number of problems prevent the idea from being put into practice: the high cost of delivering components to orbit, protection from radiation and space debris, maintenance, heat dissipation - The Information lists just a few of them.
However, it seems that very many people from the world of technology do not consider these problems unsolvable, and on the contrary, they see the idea of space data centers as very promising.
Hello, earthlings
In November 2025, Nvidia-backedUS startup Starcloud (formerly known as Lumen Cloud) launched a satellite with an Nvidia H100 AI chip into orbit, CNBC writes, the first graphics chip of such power in orbit. In early December, Starcloud said it had trained the NanoGPT AI to speak archaic Shakespearean English in space for the first time in history, and also launched a compact model of Google's AI called Gemma on the chip. When asked to make the first AI statement from space, Gemma responded:
Greetings, Earthlings! Or, as I prefer to call you, a marvelous combination of blue and green.
Of course, one chip with a consumption of about 700 watts is far from a multi-gigawatt data center with thousands of such chips, but the first step has certainly been taken. Starcloud claims to be aiming for a 5 gigawatt data center, which, in addition to chips, will require assembling a huge field of solar panels in space with a side of 4 kilometers - a total of 16 square kilometers). By the way, co-founder and chief engineer of the Starcloud project Adi Oltean calls himself on the X network a person who aims to reach the second level of civilization according to Kardashev.
Not only startups are interested in the idea. In December, The Wall Street Journal wrote, citing sources, that the team of the Blue Origin space company of Amazon founder Jeff Bezos has been working for more than a year on the technologies needed for orbital artificial intelligence data centers.
Google CEO Sundar Pichai talked about Project Suncatcher, which is exploring the possibility of launching its own TPU AI chips into orbit. The first launches are scheduled for early 2027, and within a decade, he believes orbital data centers will become the norm. And, unlike Starcloud, Google sees such a data center not as a monolithic structure, but as a constellation of 81 satellites, which are located within a radius of 1 km from each other and connected by high-speed laser optical communication. The atmosphere and weather in space do not interfere with such communication.
Hmm, the future is uncertain, but who is it that we already have thousands of satellites in orbit, linked together by laser communications? That's right, it's SpaceX, the world's richest technopreneur Elon Musk's company.
Just increase the number of Starlink V3 satellites with high-speed laser links. SpaceX will be doing this.
A month later, commenting on Salesforce CEO Marc Benioff's remarks, Musk was more specific. In his opinion, the project created by SpaceX super-ship Starship will be able to put into orbit from 300 to 500 gigawatts of power AI-chips per year, which in a couple of years will exceed the entire energy consumption of the U.S. economy. Being a visionary, Musk, of course, does not stop at this "trifle" and goes much further in his insights.
These figures are colossal by earthly standards, but have no significant value on the Kardashev type two scale. For it it is necessary that on the Moon satellites of artificial intelligence working on solar energy with the capacity of more than 100 thousand gigawatts per year were produced
How to capitalize on AI in space
Data centers in space are better by all parameters than on Earth if they could be easily placed there, says Askar Akhmedov, investment director of Atlas Capital.
While such an idea might have seemed absurd before, now that a one gigawatt data center costs about $50 billion, all options are being considered. For comparison, the ISS costs $150 billion, which means that the costs are already of the same order of magnitude
According to Akhmedov, a data center in space will consist of several components. Of course, computing chips - here, in general, everything is the same as on Earth. Network equipment will most likely not be cable (they are heavy, even fiber optics), but laser-optical. Manufacturers of such equipment include, for example, Tesat or Mynaric. The latter was bought out by RocketLab in September.
Cooling systems will also be needed, but the companies that are doing this are either non-public or already part of SpaceX for now. For example, many parts need to be created using 3D metal printing and SpaceX is doing this on its own, Akhmedov says.
And finally, electricity - it is logical to assume that U.S. data centers will buy solar panels from U.S. or, at worst, European companies, he believes. These include: Spectrolab (part of Boeing, USA), SolAero (part of RocketLab, USA) and Azur Space (part of 5N Plus, Canada). The latter two companies already supply parts for Starlink satellites.
The bottom line is that only SpaceX and RocketLab are fairly "clean" bets for data center development in space
RocketLab is publicly traded and has performed well this year, with shares more than doubling at a market capitalization of $28.6 billion.
As for SpaceX, it's not public, but according to several media outlets at once, it could go public as early as 2026 at a $1.5 - $2 trillion valuation and be "the craziest IPO in stock market history."
Profitable or not, it certainly won't be boring.
This article was AI-translated and verified by a human editor
