Space Data Centre Explained: The Economics, the Physics, and the Starship Problem
Why Companies Are Racing to Put Data Centres in Space — And Why One Physics Problem Could Stop Them All
AI is consuming electricity at a rate that is straining power grids across the United States, Europe, and Asia. A growing number of companies — from venture-backed startups to SpaceX itself — believe the answer is to move computing infrastructure into orbit. The logic is compelling. The physics is brutal. Here is what is actually happening.
The world’s AI boom has a power problem — and it is getting worse faster than anyone planned. The US Department of Energy projects that data centres will account for up to 12 percent of all American electricity demand by 2028, driven almost entirely by the explosion of AI training and inference workloads. That is up from roughly 4 percent today. Meta, Amazon, Microsoft, and Alphabet have collectively committed approximately $725 billion in capital expenditure for 2026 — almost entirely for data centres, chips, and GPU clusters.
Communities from Virginia to Frankfurt to Singapore are pushing back. Land is running out near the power grids that serve major data centre campuses. Water for cooling is increasingly regulated. Environmental reviews are slowing construction. The US alone needs to build the equivalent of a new large power plant every few weeks just to keep pace with AI compute demand.
Against this backdrop, a question that would have sounded ridiculous five years ago is now being asked seriously by engineers, investors, and regulators: what if you moved the computers to space?
This is not a theoretical discussion anymore. Starcloud, a Washington-based startup backed by Nvidia and Andreessen Horowitz, placed the first Nvidia H100 GPU in orbit on November 2, 2025, and trained an AI model in space in December 2025 — the first time a large language model has run on commercial GPU hardware beyond Earth’s atmosphere. SpaceX filed with the FCC in January 2026 for authorisation to launch up to one million data centre satellites. Google announced Project Suncatcher — a plan to deploy solar-powered satellites running tensor processing units. Blue Origin announced orbital data centre plans using its New Glenn rocket. And a wave of venture-backed startups, including Orbital, Starcloud, and Cowboy Space Company, are racing to build the infrastructure before launch costs fall enough to make the economics work. The race is real. Whether the physics cooperates is the question.
The race to build data centres in space — background context. All rights respective owner.
Why Space Makes Sense — In Theory
The case for orbital data centres is built on three genuine advantages that terrestrial infrastructure cannot match.
Power. In low Earth orbit, a solar panel generates five to ten times more energy per square metre than the same panel on Earth’s surface. There are no cloudy days. There is no night cycle in a sun-synchronous orbit. The sun is effectively always on. For an industry that is straining national power grids, the appeal of essentially unlimited solar electricity is obvious.
Land. There are no planning permissions in orbit. No zoning boards. No community opposition. No environmental review delays. No competing land uses. Space is not running out of space. Elon Musk has stated publicly that SpaceX’s goal is to launch 100 gigawatts of AI compute to orbit annually — for comparison, that is roughly one fifth of the entire annual electricity production of the United States, deployed not on Earth but above it.
Latency for space-generated data. Observation satellites, weather monitoring systems, and communications constellations generate enormous volumes of data that currently must be transmitted to Earth for processing. If the processing hardware is in orbit alongside the sensors, the data never needs to come down at all — only the results do. For applications like wildfire detection, maritime search and rescue, and military intelligence, that latency reduction is operationally significant.
Existing satellites are essentially very expensive radios. They receive a signal, relay it, and transmit something back. The computing they do is minimal — just enough to manage their own operation. An orbital data centre is something different: a satellite whose primary purpose is computation, not communication. Instead of relaying signals, it runs AI models, processes data, and returns only the outputs. Starcloud’s Starcloud-1 satellite, launched in November 2025, demonstrated exactly this — it ran Google’s Gemma language model in orbit, processing queries and returning responses. The satellite was not a relay. It was a computer in space.
The Economics: Why It Doesn’t Work Yet
The honest reason orbital data centres remain mostly theoretical — despite genuine investment and genuine hardware in orbit — is that the launch economics are currently prohibitive for large-scale deployment. This is the central tension in the entire industry and it deserves careful explanation.
Getting one kilogram of payload to low Earth orbit costs approximately $2,720 on a Falcon 9 rocket, based on the current list price of $74 million per launch divided by a reusable payload capacity of roughly 17,500 kilograms. An Nvidia H100 GPU — the chip currently at the heart of most AI inference workloads — weighs approximately 2.4 kilograms. At $2,720 per kilogram, that’s around $6,500 just to launch the chip into orbit, before you’ve added the satellite chassis, solar panels, thermal management systems, radiation shielding, communications hardware, or any of the dozens of other components required to keep it operational.
| Rocket | Cost per kg to LEO | Era / Status | What changes at this price |
|---|---|---|---|
| Space Shuttle | ~$54,000/kg | 1981–2011 | Only governments could afford orbit |
| Falcon 9 (expendable) | ~$13,200/kg | Early SpaceX | Commercial satellites became viable |
| Falcon 9 (reusable, today) | ~$2,720/kg | 2026 current | Starlink became economically feasible |
| Starship (target) | <$100/kg | When operational | Orbital data centres become economically viable |
The table above illustrates why Orbital’s founder Euwyn Poon — who raised $5 million from Andreessen Horowitz’s Speedrun accelerator in May 2026 to build a network of 10,000 inference satellites — stated plainly that the current Falcon 9 price “makes this not economically feasible.” Poon’s entire business model, like most in this space, is contingent on SpaceX delivering Starship at commercial scale.
Nearly every orbital data centre company is betting on the same thing: SpaceX getting Starship to commercial readiness at a price that drops launch costs by roughly 27 times versus current Falcon 9 rates. Starship has completed 11 test flights as of May 2026. It has not yet reached commercial operational status. Until it does — and until SpaceX demonstrates the rapid reusability required to hit sub-$100/kg pricing — the business case for large-scale orbital compute cannot close. This is not pessimism. It is arithmetic. The entire industry knows this and is building toward it anyway, running small-scale proof-of-concept missions now to be ready when the economics shift. The question is when, not if.
The Physics Problem Nobody Is Advertising
Launch economics get most of the attention in coverage of this industry. The harder problem — the one that Jensen Huang, CEO of Nvidia, specifically called out in February 2026 — is cooling.
On Earth, data centres cool their chips using air and water. Both are abundant. Both work through convection — heat moves from the chip into the surrounding medium and is carried away. In space, there is no air. There is no water. There is no convection at all. The only mechanism available to remove heat from electronics in orbit is thermal radiation — the chip must literally glow its waste heat away as infrared light into the void of space.
“It’s counterintuitive, but it’s hard to actually cool things in space because there’s no medium to transmit hot to cold. All heat dissipation has to happen via radiation, which means you need to have a radiator pointing away from the Sun to do that.”
The physics here is governed by the Stefan-Boltzmann law, which determines how much heat a surface can radiate based on its area, temperature, and material properties. The numbers are striking. According to analysis published by EE Times in March 2026, dissipating just one megawatt of waste heat from orbital compute requires approximately 1,200 square metres of radiator surface — a structure roughly 35 by 35 metres, deployed in space. For context, that is larger than the floor area of most houses, attached to a satellite.
The World Economic Forum’s June 2026 analysis identified a further complication: the design tension between keeping GPUs close together — which helps with the high-bandwidth interconnects that AI inference requires — and spreading them apart to make cooling easier. Put chips close together and heat concentrates, requiring long runs to large radiators. Spread them out and cooling becomes easier, but the latency and bandwidth between chips degrades, which matters for AI workloads.
Satellites in low Earth orbit cycle between direct sunlight and shadow approximately every 90 minutes. In sunlight, solar panels generate power and temperatures rise. In shadow, generation stops and temperatures drop sharply. This repeated expansion and contraction stresses every component — solder joints, connectors, memory cells, and logic gates — degrading reliability over time in ways that simply do not occur in a temperature-controlled terrestrial data centre. Radiation from the Van Allen belts and cosmic rays cause additional damage by flipping memory bits and degrading semiconductor materials over time. These are not unsolvable problems — satellite engineers have managed them for decades — but they represent real constraints on the density and reliability of compute hardware that can operate in orbit. A server rack designed for a terrestrial data centre cannot simply be launched into orbit and expected to function reliably.
Who Is Actually Flying Hardware Right Now
The Orbital Founder Angle — Why It Actually Matters
Orbital’s Euwyn Poon built and sold Spin, an e-scooter company, to Ford in 2018 after deploying 250,000 scooters across 100 cities in roughly a year. That operational track record — managing hardware at massive scale, navigating regulatory environments across hundreds of jurisdictions, and executing rapid physical deployment — is precisely why Andreessen Horowitz’s Andrew Chen backed him despite no space background.
After leaving Ford, Poon bought a single Nvidia A100 GPU, co-located it in a Santa Clara data centre, and began serving open-weight AI models. That hands-on experience with the economics of AI compute — understanding the cost per inference, the power consumption, the revenue potential — gave him the ground-level perspective to identify the orbital compute opportunity from the demand side rather than the supply side. Most space infrastructure companies are built by aerospace engineers. Poon approached it as a compute economist who happened to end up in space.
Whether that distinction matters in the long run is an open question. Building 10,000 radiation-hardened inference satellites is a fundamentally different engineering challenge from deploying scooters at scale. But the a16z view — that the orbital compute opportunity is large enough to accommodate many different approaches and many different companies — is hard to argue with given the scale of what SpaceX, Google, and Blue Origin are planning.
My Take — Mr Wangdoo
The thing that strikes me most about this industry is how clearly everyone involved understands the problem. Euwyn Poon said openly that current launch costs make his business infeasible. Jensen Huang said publicly that orbital data centre economics are poor today. The World Economic Forum published a detailed analysis of why cooling is a physics constraint that cannot be wished away. This is not an industry built on hype and obfuscation — it is an industry built on a very clear bet about when one specific technology, Starship, will work at commercial scale.
That clarity is both reassuring and concerning. Reassuring because the people building these companies are not pretending the problems don’t exist. Concerning because every company in the space is exposed to the same single risk: if Starship does not achieve commercial readiness at sub-$200/kg pricing within the next three to five years, the entire industry’s economics fail simultaneously. That is an unusual level of dependency for a sector attracting billions in investment.
What Starcloud’s December 2025 demonstration proved — training an AI model on an Nvidia H100 in orbit — is that the compute hardware works in space. That was the most important question and it has now been answered. The remaining questions are engineering ones about thermal management and radiation tolerance at scale, and economic ones about launch costs. Both are tractable. Neither is guaranteed.
The GAO’s April 2026 report noted that the FCC has received three applications from US companies for large satellite data centre constellations since January 2026. Regulators are taking this seriously. Capital is flowing. Hardware is flying. The space data centre is no longer science fiction — it is an engineering and economics problem with a known solution timeline. The timeline just depends on a rocket that has not yet flown commercially.
Frequently Asked Questions
Why would anyone put a data centre in space?
Three reasons. First, unlimited solar power — in orbit, solar panels generate five to ten times more energy per square metre than on Earth with no downtime for night or cloud cover. Second, no land use constraints — there are no planning permissions, zoning disputes, or community opposition in orbit. Third, for data generated in space, processing it there rather than transmitting raw data to Earth reduces latency significantly for time-sensitive applications like wildfire detection or maritime rescue. The fundamental driver is AI’s insatiable demand for compute power, which is straining terrestrial power grids and running out of suitable land near major electricity infrastructure.
Has anyone actually put AI compute in space yet?
Yes. Starcloud, a Washington-based startup backed by Nvidia and Andreessen Horowitz, launched the first Nvidia H100 GPU into orbit on November 2, 2025. In December 2025, the company trained a language model in space — the first time a large language model has run on commercial GPU hardware in orbit. Axiom Space deployed a data processing unit on the International Space Station in autumn 2025. The US Government Accountability Office confirmed in April 2026 that the FCC had received three formal applications from US companies for large satellite data centre constellations. The hardware works. The scale economics do not yet.
Why is cooling such a problem in space if space is cold?
This is the most counterintuitive aspect of orbital compute. Space is cold — roughly minus 270 Celsius in the void — but that temperature is irrelevant because heat transfer requires a medium. On Earth, air and water carry heat away from chips through convection. In the vacuum of space, there is no medium. The only way to remove heat from electronics in orbit is thermal radiation — the chip must radiate its waste heat as infrared light. This mechanism is governed by the Stefan-Boltzmann law and requires large surface areas. Analysis published in EE Times estimates that dissipating one megawatt of compute heat in orbit requires approximately 1,200 square metres of radiator surface. Modern AI inference facilities consume hundreds of megawatts. The radiator engineering challenge scales directly with compute power.
Why do all these companies depend on Starship?
Because the current launch cost makes large-scale orbital compute economically unviable. Falcon 9, the current state of the art, costs approximately $2,720 per kilogram to low Earth orbit. Starship — SpaceX’s next-generation fully reusable rocket — targets below $100 per kilogram when operating at commercial scale. That is a 27-fold reduction. At Falcon 9 prices, the cost of launching the hardware required for a commercially useful orbital data centre constellation is prohibitive. At Starship prices, it becomes competitive with terrestrial data centre construction in high-demand markets. Orbital’s founder stated this directly: Falcon 9 pricing “makes this not economically feasible.” As of May 2026, Starship has completed 11 test flights but has not yet reached commercial operational status.
What is SpaceX’s actual plan for orbital data centres?
SpaceX filed with the US Federal Communications Commission on January 30, 2026, for authorisation to launch up to one million orbital data centre satellites at altitudes between 500 and 2,000 kilometres. The filing projected that launching one million tonnes of satellites annually could generate 100 gigawatts of AI compute capacity. SpaceX and xAI merged in February 2026, creating a combined entity valued at $1.25 trillion, and orbital compute is explicitly listed as a strategic objective in SpaceX’s IPO S-1 filing. Elon Musk stated in February 2026 that within approximately five years he expects several hundred gigawatts of AI compute to be operating in space annually.
Who is Euwyn Poon and why did a16z back him for a space company?
Euwyn Poon founded Spin, an e-scooter company, in 2017 and sold it to Ford in 2018 after scaling to 250,000 scooters across 100 cities. After leaving Ford, he bought an Nvidia A100 GPU and co-located it in a Santa Clara data centre, building first-hand experience with AI compute economics. Andreessen Horowitz partner Andrew Chen backed Poon through a16z’s Speedrun accelerator program specifically because his operational track record — managing regulated hardware deployment at city scale — is relevant to the challenge of building an orbital compute constellation. Poon’s company Orbital emerged in May 2026 with a $5 million seed round and plans for 10,000 inference satellites. The company plans a 2027 demo flight to test radiation shielding and thermal management before committing to full constellation deployment.
- US GAO — Science & Tech Spotlight: Data Centers in Space — energy projections, FCC applications, technical challenges (April 28, 2026)
- World Economic Forum — Why cooling is the real obstacle to space-based data centres (June 2026)
- IEEE Spectrum — Orbital Inference Data Center Bets On Space GPUs — technical architecture analysis (May 2026)
- CNBC — Nvidia-backed Starcloud trains first AI model in space (December 2025)
- EE Times — The Hidden Physics of Running Data Centers in Orbit — Stefan-Boltzmann radiator analysis (March 2026)
- Orbital Radar — Launch cost per kilogram — Falcon 9 and Starship target figures (updated May 2026)
- Northeastern University — The next great space race: Building data centers in orbit — expert analysis on power, cooling, and radiation (January 2026)
- Axiom Space — Orbital Data Center programme — ISS deployment and January 2026 LEO node launches
