Orbital data centers are coming. The question is what gets built first.

April 2026

The AI industry and the space industry are converging on the same conclusion: compute needs to move to orbit. But the path to getting there is more demanding, and more interesting, than the headlines suggest.

The energy wall

The International Energy Agency projects that global data center electricity consumption will reach 1,100 TWh in 2026, equivalent to Japan's entire national consumption [1]. By 2030, demand from AI optimized data centers is expected to more than quadruple [2].

The infrastructure buildout is enormous. Amazon, Microsoft, Alphabet, and Meta collectively spent roughly $410 billion on data center infrastructure in 2025. In 2026, that figure is expected to reach $650 to $700 billion [3]. Despite this, the bottleneck has shifted from the server rack to the electrical substation. In Northern Virginia, the world's largest data center market, grid operators have issued capacity warnings through 2028 and effectively halted new permits [1]. Gartner predicts power shortages will operationally constrain 40% of AI data centers by 2027 [4].

Most of the developed world's electrical grids were built between the 1950s and 1970s. Upgrading them takes years, sometimes decades. AI infrastructure expands in months [5]. This mismatch is what moved orbital data centers from science fiction to strategic roadmaps.

Two tracks, two timelines

A solar panel in orbit produces roughly five times the electricity of its ground equivalent. No atmosphere, no weather, no day/night cycle for most orbits, no interconnection queues, no building permits [3]. The physics makes the case. But two very different approaches are emerging.

The first is the hyperscaler vision: solving AI's terrestrial energy crisis by moving massive compute to orbit. SpaceX filed FCC plans for millions of data center satellites. Starcloud raised $170M at a $1.1B valuation and proposed a constellation of 88,000 satellites. Blue Origin announced TeraWave, roughly 5,400 satellites for high throughput connectivity [6]. Google's feasibility study puts the viability threshold at $200/kg launch cost, down from roughly $1,000/kg today, projected around 2035 [7].

The second is edge computing for space itself. Axiom Space launched the first orbital data center nodes in January 2026 [8]. NVIDIA released its Space-1 Vera Rubin and Jetson Orin platforms for constrained orbital environments [9]. This track addresses a need that already exists: roughly 14,000 smallsats will launch between 2024 and 2033, generating data that today must wait hours for ground station visibility before it can be processed.

These tracks are not in competition. They address different problems on different timelines. But they share a common dependency: nobody has yet built the foundational compute, storage, and networking hardware qualified to operate reliably in orbit at the performance levels required.

Real challenges, real solution paths

None of these obstacles are theoretical. But none of them are dead ends either. Each has a credible engineering trajectory.

Power at scale. The ISS generates about 100 kW, roughly what a single modern AI server rack consumes [7]. That is a massive gap. But the solar array industry is closing it. Next generation deployable arrays like Redwire's iROSA already exceed 100 W/kg of specific power and deliver over 20 kW per unit, using rollable flexible blankets that stow at 40x the volume density of legacy rigid panels. New materials such as aerospace grade ultra thin glass (30 to 50 micrometers) demonstrated at CES 2026 enable even larger deployable surfaces that resist atomic oxygen and UV degradation in LEO. The trajectory is clear: larger, lighter, more efficient foldable arrays will progressively unlock kilowatt then multi kilowatt class power budgets per satellite. The challenge is real, but the supply chain is moving.

Thermal dissipation. In a constantly illuminated orbit, equipment temperatures stay above 80°C [10]. Heat can only leave through radiation, not convection. But a new generation of thermal management systems is emerging, drawing from the same engineering principles as terrestrial data center cooling: pumped fluid loops, direct to chip cooling, and dedicated radiator panels. Companies in the thermal control space are progressively scaling the amount of power they can dissipate in orbit. This is not a solved problem, but it is an active engineering frontier where each satellite generation pushes the envelope further. The industry is learning to dissipate more and more heat per unit, mission after mission.

Launch costs and sustainability. Over 70% of an orbital data center's carbon footprint comes from the launch itself [11]. This is significant, but launch vehicles are getting more capable and less expensive with every iteration. SpaceX's Starship, if it reaches projected cadence, could drive costs below $200/kg. In Europe, the development of next generation reusable launchers is a strategic priority that would both reduce the cost and the carbon intensity of orbital deployment. The path to sustainable orbital infrastructure runs through cheaper, cleaner, more frequent launches, and the trajectory is pointing in the right direction.

Hardware durability. Components must survive radiation, microgravity, and debris without physical maintenance. And by the time a space grade chip reaches orbit, two generations of commercial components have already surpassed it. This is where our work fits in directly. Rather than waiting for the semiconductor industry to produce space specific silicon, we believe the right approach is to qualify high performance commercial electronics for the orbital environment, building radiation tolerance, fault containment, and system level resilience into the architecture itself. This is the technology brick that makes everything else possible, and it is the core of what we do at Coros Space.

Our approach: qualifying commercial hardware for space

At Coros Space, we think the dominant conversation misses an operational reality. The orbital data center will not be built all at once. It will be built incrementally, starting with the most immediate need: processing space data where it is generated.

A step by step vision. First, process space generated data in orbit: Earth observation, telemetry fusion, onboard AI inference, autonomous spacecraft decision making. Once the infrastructure proves itself, open it to terrestrial data. This sequence is dictated by technical and economic reality, not aspiration.

Our model: a constellation of server satellites. We see the space data center as a network of server satellites interconnected through high bandwidth optical links, both between themselves and with client satellites. When an observation satellite captures an image, it sends data to the nearest server node. Our system processes it in orbit, can trigger new acquisitions based on what it finds, and relays results through the constellation to the ground. The processing chain moves from centralized on the ground to distributed in orbit.

Full data center payloads. Inside each server satellite, we integrate all three pillars of a data center node: high-performance compute, mass storage, and high-bandwidth networking. Our unit pairs a modular compute architecture with a high-reliability safety barrier that enforces strict segregation between payload and spacecraft avionics. It supports expansion for storage and future accelerators through COTS high-throughput interfaces, and the architecture scales through interconnection of multiple modules for either redundancy or aggregated performance.

Why COTS, not space grade silicon. This is perhaps our most consequential technical decision. Orbital data centers will operate on thinner margins than terrestrial ones. Designing exclusive space grade systems is counterproductive: it lengthens development cycles, inflates costs, and creates a permanent performance gap with the commercial state of the art.

NVIDIA's Space-1 Vera Rubin Module is a meaningful signal: the market is large enough to justify a dedicated space-grade GPGPU. But it also illustrates the structural constraint that shapes our strategy. Development and adoption cycles in space are fundamentally longer than on the ground. By the time a space-grade component completes qualification, clears regulatory review, and reaches operational deployment, the commercial state of the art has moved two generations forward.

The economics of in-orbit computing require a different answer: a COTS-to-Space strategy, grounded in the engineering discipline to qualify high-performance commercial components for the orbital environment within six months. This cadence keeps performance competitive, costs absorbable, and positions us to integrate each successive generation of commercial hardware as it matures, rather than locking operators into yesterday's silicon.

Looking ahead

The in orbit data center market is projected to reach $39.1 billion by 2035. Two timelines coexist: resolving AI's terrestrial energy crisis (2040 to 2045, dependent on launch costs reaching $200/kg) and processing space data in orbit (addressable now).

The second path does not require a fivefold reduction in launch costs. It requires compute payloads that are reliable, performant, and qualified for the orbital environment at costs constellation operators can absorb. Those who master these foundational bricks will be best positioned when the market scales.

‍

References

[1] AI Data Centers Now Use More Power Than 30 Countries, Tech Insider, April 2026
[2] AI and Data Centre Electricity, IEA
[3] The Race to Build Orbital Data Centers, SpaceNews, April 2026
[4] Data Center Power Crisis 2026, Enki AI, February 2026
[5] 2026 Predictions: AI Sparks Data Center Power Revolution, Data Center Knowledge, February 2026
[6] Space Based Data Center, Wikipedia
[7] Big Tech's Next Move: Data Centers in Space, NPR, April 2026
[8] Orbital Data Centers, Axiom Space, January 2026
[9] NVIDIA Launches Space Computing, NVIDIA Newsroom, March 2026
[10] Four Things We'd Need to Put Data Centers in Space, MIT Technology Review, April 2026
[11] Space Data Centers and Decarbonization, Carbone 4