Orbital Data Centers: The Engineering Case

Overview

This report provides a comprehensive assessment of the technical, economic, environmental, and strategic advisability of moving large-scale data center functions into Earth orbit over the next 10 to 30 years. Currently, terrestrial hyperscale data centers face serious physical limitations: increasing power consumption, cooling water depletion, land use constraints, and grid instability. In response, the orbital data center presents a paradigm of continuous 24/7 solar energy utilization, passive radiative cooling leveraging the vacuum environment of space, and new digital sovereignty unconstrained by terrestrial jurisdiction.

This report clarifies under what conditions orbital data centers will be competitive with ground facilities. The analysis shows that in the early 2030s, when launch costs fall below the threshold of $100 to $500 per kilogram and a modular robotic maintenance ecosystem is established, orbital data centers will reach cost parity with ground facilities for certain workloads — such as AI training, defense, and batch processing — that do not require ultra-low latency. As of 2026, this is no longer a speculative thesis. Hardware is in orbit, capital has been committed, and the two largest economies on Earth are competing for position.

Ground infrastructure limitations

The computing infrastructure that powers the global digital economy is currently at a historic turning point. As AI models grow larger, the demand for computing resources is growing exponentially, with global data center capacity expected to reach 100 gigawatts by 2030 — roughly double the current installed base. The US Department of Energy projects that data centers will account for up to 12 percent of American electrical demand by 2028, driven almost entirely by AI training and inference workloads.

The first barrier is a constrained power grid. Major data center hubs such as Ireland and the Netherlands have introduced moratoria on the construction of new hyperscale facilities due to the load on existing power grids. Waiting times to connect to the terrestrial power grid can range from three to seven years in developed countries, creating the risk that billions of dollars in AI investments stagnate as idle silicon. A single large AI training cluster can consume 100 to 500 megawatts continuously.

The second barrier is a lack of cooling resources. Ground-based data centers convert almost all of their electricity to heat and consume vast amounts of freshwater to cool down. In areas facing water shortages, data center operations are driving resource conflicts with local residents and industry, and social operating licenses are rapidly being lost.

In contrast, moving into orbit means switching the source of resources from Earth's biosphere to outer space. Solar radiation is about 40 percent higher than on the ground, with no atmospheric attenuation or night-time blockage, so the capacity factor for solar power generation reaches over 95 percent compared to 10 to 24 percent on the ground. The FCC has received three applications since January 2026 from US companies seeking to operate large satellite constellations as data centers. That number did not exist six months earlier.

Thermodynamic challenges

The biggest and most counterintuitive challenge in operating a data center in orbit is thermal management. There is a common perception that the universe is cold, but a vacuum acts as a perfect insulator. While ground-based facilities use air and water to dissipate heat, in space only infrared radiation provides a continuous means of heat dissipation.

According to the Stefan-Boltzmann law, the heat dissipation capacity is proportional to the fourth power of the radiating surface temperature. NVIDIA GB300-class GPUs used for AI training consume approximately 1.4 to 2 kilowatts of power per chip, all of which converts to heat. To operate at a stable temperature of around 20 degrees Celsius, a large radiator surface area is required. Calculations show that cooling a 1 megawatt compute cluster requires thousands of square meters of radiating surface — a significant fraction of the system's overall mass and cost.

One of the major advantages of orbital data centers is that auxiliary power consumption for cooling can be reduced to the extreme. In a ground-based data center, cooling fans, compressors, and circulation pumps consume 30 to 50 percent of total power. Because radiative cooling in space is passive, PUE can theoretically achieve figures of 1.01 to 1.05. The latest research is developing photonic structures and special coatings with high emissivity in the mid-infrared spectrum to efficiently radiate heat toward the 3 Kelvin cosmic microwave background. Inorganic coatings using silica aerogels and phosphate geopolymers have been demonstrated to maintain high emissivity without degradation even in harsh radiation environments.

Radiation-hardened hardware

Outside Earth's atmosphere, electronics are constantly exposed to high-energy particles such as galactic cosmic rays, solar particle events, and protons captured in the Van Allen belts. These radiations have two major adverse effects on semiconductors.

The first is the total ionizing dose effect: charge accumulates in semiconductor structures over time, gradually degrading device performance and ultimately leading to failure. The second is the single event effect, which causes bit flips — where the contents of memory are rewritten by a single particle collision — or latch-up, where the circuit is destroyed by an overcurrent.

Modern AI accelerators are manufactured using ultra-fine process nodes such as 5nm and 3nm, making these smaller transistors more vulnerable to single event effects than older, larger transistors. Conventional radiation-hardened chips are more than a decade behind commercial chips in performance and are not suitable for cutting-edge AI training.

To resolve this paradox, three strategies are being adopted. First, triple modular redundancy: three identical processors perform the same calculations and decide the outcome by majority vote. Second, physical shielding with lead, tungsten, or hydrogen-rich polymers — though shielding adds approximately 1 kilogram of mass per kilowatt of computational power, increasing launch costs. Third, enhanced error correction code applied across memory and registers to detect and correct bit flips in real time.

Proton beam tests using Google's Trillium TPU suggest that even a radiation dose equivalent to a five-year mission in low Earth orbit would not cause fatal physical damage, and that the chip could be operated with proper bit-flip correction in place. In March 2026, NVIDIA released the Vera Rubin Space-1, a space-qualified module delivering up to 25 times more AI compute than the H100, engineered for the radiation, thermal cycling, and vacuum conditions of orbital operation. GPU manufacturers do not engineer space-qualified hardware until the market is large enough to justify the development cost. The announcement is a signal about where the industry expects demand to flow.

What is operating

Starcloud's first satellite, launched November 2025 aboard a SpaceX Falcon 9, carried an NVIDIA H100 GPU and demonstrated sustained AI inference in low Earth orbit — the most powerful computing hardware ever operated in space. Within weeks it had trained a small language model entirely in orbit and run a version of Google Gemini on the satellite. By March 2026, Starcloud had raised $170 million at a $1.1 billion valuation. Starcloud-2, scheduled for October 2026, will carry multiple H100s alongside NVIDIA's Blackwell chip, an AWS server blade, and a bitcoin mining module — the first multi-tenant orbital compute platform.

Axiom Space launched two orbital data center nodes on January 11, 2026. NVIDIA named six initial partners for its space computing platform: Aetherflux, Axiom Space, Kepler Communications, Planet Labs, Sophia Space, and Starcloud. The breadth of that list spans communications infrastructure, Earth observation, commercial space stations, and pure orbital compute — a platform play, not a single-use-case product.

Google and SpaceX are in active talks on Project Suncatcher, Google's orbital data center initiative, with discussions involving SpaceX launching satellites carrying Google's TPU hardware into orbit. Google's internal timeline targets first test satellites by 2027. The longer vision involves gigawatt-scale orbital compute capacity. Meta has separately reserved one gigawatt of orbital solar energy and 100 gigawatt-hours of long-duration storage, with the intent to beam power to ground-based data centers — an orbital-energy-for-terrestrial-compute model that sidesteps the harder engineering problems of cooling and data transmission latency while accessing the solar capacity factor advantage of orbital operation.

Chinese enterprises have been operating AI compute in orbit for considerably longer, with less public attention. Zhongke Tiansuan had already been operating a space computer aboard a Jilin-1 satellite for over 1,000 days by early 2026. The company's Aurora 5000 system uses a domestically developed GPU, reducing dependence on US chip supply chains for orbital compute. The China National Space Administration deployed three test satellites in 2024 equipped with domestic AI accelerators, with plans for a 50-satellite AI constellation operational by 2028. China's 15th Five-Year Plan explicitly includes gigawatt-scale space-based digital infrastructure as a national priority, framed explicitly around digital sovereignty.

Launch economics

The biggest factor determining the feasibility of an orbital data center is the cost of orbital insertion. The $50,000 per kilogram cost of the Space Shuttle era was a barrier to all commercial activity. SpaceX's Falcon 9 brought that down to around $2,700 per kilogram. Once the fully reusable Starship is operational, launch costs are projected to fall to a range of $100 to $200 per kilogram. Starcloud estimates the equivalent energy cost of orbital solar at approximately $0.005 per kilowatt-hour, roughly 15 times lower than current wholesale electricity prices.

Cost estimates for building a 1 gigawatt orbital facility vary widely depending on launch cost assumptions. Currently, the cost of building a 1 gigawatt facility on the ground is estimated at around $15 to $20 billion. In orbit, launch costs account for more than 40 percent of total investment.

At a launch cost of $500 per kilogram, an orbital facility costs about three times as much as a ground-based facility. However, once $100 per kilogram is reached, a cost reversal is predicted in the early 2030s, making building in orbit an economically reasonable option when accounting for rising ground energy prices, land acquisition costs, and regulatory compliance costs. If terrestrial electricity costs continue rising and orbital solar costs fall as solar panel mass and launch costs both decrease, the crossover point arrives earlier than the capital cost comparison alone suggests.

One of the critical risks of orbital infrastructure is that physical repairs are not easily possible. Ground-based data centers can replace faulty GPUs within hours. In orbit, ROI depends on managing annual failure rates estimated at 5 to 9 percent. Full module replacement through frequent launches, or the establishment of an in-orbit robotic maintenance ecosystem, is key to ensuring long-term economics.

Communication and latency

When placing a data center in orbit, networking is as important as computing power. AI training workloads require massive data synchronization — such as the All-Reduce operation — and require low-latency, high-bandwidth interconnections. Within a ground-based data center, fiber channels provide petabit-level communication with microsecond delays. To reproduce this in orbit, free-space optical communication (laser communication) is essential.

According to Google's analysis, by forming constellation clusters in which satellites fly closely together over a range of hundreds of meters to several kilometers, it is possible to increase received power and secure terabit-level communication bandwidth within the cluster.

Round-trip latency to ground users from low Earth orbit is approximately 5 to 20 milliseconds, which is acceptable for about 75 percent of real-time inference workloads. AI training is relatively tolerant of ground delays on the order of milliseconds and is expected to be the first major application for orbital data centers. For backhaul, existing constellations such as Starlink are being used to distribute orbital compute results to ground stations around the world.

Robotic maintenance

For an orbital data center to function as infrastructure rather than a disposable satellite, it must be upgradeable and repairable in orbit. DARPA's Robotic Servicing of Geosynchronous Satellites program and Northrop Grumman's Mission Extension Vehicle demonstrate pioneering technologies for this purpose.

The next generation of orbital infrastructure is expected to adopt a modular architecture similar to Lego blocks. As a compute module breaks down or becomes obsolete, a robotic arm pulls out the old module and plugs in a new one. Power, communication, and cooling fluid circulate via standardized docking ports. Orbital tugs that refuel satellites to maintain orbit or re-enter the atmosphere at the end of their life function as part of the broader maintenance ecosystem.

DARPA's robotic servicing vehicles are scheduled to operate in geostationary orbit after 2025. Application of these technologies to low Earth orbit data center clusters will enable significant extensions of operational lifetimes — from 10 years to over 15 years.

Geopolitical sovereignty

The physical location of a data center has always been linked to legal and political jurisdiction. Ground-based facilities must comply with the laws of their home country — GDPR, the US CLOUD Act, or local export control regulations. Orbital data centers do not exist on the territory of any particular state, necessitating a new legal interpretation based on the 1967 Outer Space Treaty.

Under the proposed digital flag state framework, a satellite is considered to be under the sovereignty of its country of registration. The interior of the satellite enclosure, regardless of its orbital position, is treated as the territory of the registering country and is immune from physical or legal intervention by other states. The point at which data is transmitted and received through ground stations acts as a border, while any processing or storage in orbit is treated in the same way as activities on the high seas under international law.

For smaller nations or corporations concerned about power shortages or the physical seizure of a ground data center by an adversary, the orbital data center acts as an ultimate safe haven. Security must be embedded at the silicon level, as physical intervention is difficult in the space environment. Companies such as SEALSQ are developing semiconductors with post-quantum cryptography for orbital data centers, providing a root of trust to protect against future quantum computer attacks and to secure satellite-to-satellite communications and AI model intellectual property.

The regulatory machinery has not kept pace. The FCC evaluates satellite constellation applications on technical and interference grounds and has no framework for evaluating constellations whose primary purpose is AI compute rather than communications. The ITU coordinates frequency spectrum and orbital positions but has no mandate over compute infrastructure. No international body has jurisdiction over the strategic or national security implications of orbital AI processing capacity. The most strategically sensitive computing infrastructure being built today is operating under regulatory frameworks designed for communications satellites.

Environmental impact

Orbital data center proponents claim it is the ultimate green AI platform, capable of reducing water consumption and carbon emissions on the ground. However, a detailed life cycle analysis reveals complex challenges.

Giant rockets like Starship burn approximately 1 kiloton of liquid methane per launch. Methane is considered a relatively clean fuel, but soot (black carbon), water vapor, and alumina particles released in the stratosphere and mesosphere can have a warming effect estimated at 500 times more powerful than the same emissions released at ground level. The ASCEND study found that for orbital computing to be as environmentally responsible as a ground-based data center powered by renewable energy, rocket emissions would need to be kept below 1.9 kilograms of CO₂ equivalent per kilogram of payload. Current rockets exceed this threshold by a factor of 10 to 25.

When satellites burn up in the atmosphere at the end of their lifespan, they release metal vapors such as aluminum oxide, which have been linked to ozone depletion and risks to Earth's radiation budget. Plans to launch tens of thousands of satellites also increase the risk of Kessler syndrome — a chain reaction of space debris. Orbital data center operators must bear the costs of autonomous collision avoidance systems, forced deorbit propulsion to ensure atmospheric entry within five years of mission completion, and potential debris removal costs estimated at $100 million to $500 million per object in the event of a collision.

Three waves

The industrialization of orbital infrastructure is expected to proceed not as a single event but as a wave of technological and economic developments.

The first wave, from 2026 to 2030, covers edge reasoning and national security: real-time processing of satellite data, secure computation for national defense, and low-latency hubs for financial transactions. This wave consists of hundreds of small satellite nodes, with prototype clusters ranging from 100 kilowatts to 500 kilowatts. Starcloud, Axiom, and China's Zhongke Tiansuan are all first-wave operators. The reliability of off-grid systems that do not depend on terrestrial power grids is evaluated during this period.

The second wave, from 2030 to 2040, is hyperscale AI training: large-scale language model training, artificial general intelligence development, and global cloud backbone infrastructure. This wave involves gigawatt-class clusters with thousands to tens of thousands of satellites. Starship launch costs fall below $200 per kilogram, and stricter energy regulations on the ground make orbital data centers the mainstream compute infrastructure. This is where Google's Project Suncatcher, NVIDIA's space computing platform, and equivalent Chinese programs are targeting.

The third wave, from 2040 to 2055, integrates with the lunar and deep space economies: control of lunar bases, data relay for deep space exploration, and support for resource extraction in space. Compute bases at Earth-Moon Lagrange points and in lunar orbit become viable. Manufacturing of orbital structures using lunar resources begins, creating a true space economy that does not rely on launches from Earth.

Conclusions and recommendations

Orbital data centers are no longer the domain of science fiction. They have emerged as a logical response to the energy and physical constraints facing ground-based infrastructure, and as of 2026 the transition from thesis to hardware is underway.

On economic validity: if launch costs fall within the range of $100 to $500 per kilogram, orbital data centers will compete with ground facilities on capital costs and surpass them on power costs and cooling efficiency. On technical priorities: development will focus on advanced radiator designs for waste heat management and system-level radiation-resistant designs for utilizing commercial chips in space. On geopolitical value: by establishing digital flag state standards, orbital data centers will provide strategic value as a sovereign cloud for nations and corporations. On environmental obligations: developing international standards and investing in technologies to ensure the sustainability of the space environment — debris management, re-entry pollution — are essential for the long-term survival of the industry.

The satellite industry has priced orbital positions for decades as communications infrastructure. An orbital position that can support a megawatt-scale AI inference cluster has different value characteristics than one that supports a communications relay. The thermal envelope available at a given altitude, the radiation environment at different orbital inclinations, the access to continuous solar illumination, and the proximity to terrestrial population centers all take on new pricing dimensions when the tenant is a GPU cluster rather than a transponder. As the orbital economy shifts from communications to communications plus compute, the gap between what existing registry systems capture and what the market needs to know grows wider.

Orbital data centers will be an essential bridge for 21st-century digital civilization to expand beyond the limits of Earth's biosphere. Whoever controls this frontier will have sovereignty over the next generation of computing resources and intelligence.