Orbital Data Centers & the Energy Arbitrage Reshaping AI Infrastructure

1   The Terrestrial Demand Crisis

The coming demand for AI compute infrastructure represents one of the largest capital deployment challenges in modern history. Multiple institutional forecasts converge on a central conclusion: terrestrial infrastructure cannot scale fast enough to meet projected demand.

Goldman Sachs' April 2024 "Generationai" report projected US data center power demand reaching 35 GW by 2030 — a roughly 160% increase from 2023 levels.1 The International Energy Agency's January 2025 Electricity Market Report projects global data center electricity consumption could exceed 945 TWh by 2030, approximately doubling from 2024 levels.4 Semianalysis and Boston Consulting Group have published more aggressive estimates exceeding 60 GW for the US alone by 2028.5 McKinsey's June 2024 analysis aligns with the Goldman lower bound at 35 GW by 2030.6

The range of 50–100 GW of additional US capacity represents an aggressive but empirically grounded scenario if frontier model training continues scaling at current rates. Epoch AI's longitudinal analysis shows training compute has grown at approximately 4.2x per year since 2010, with acceleration to ~5x per year since 2020 for frontier models.7 At this trajectory, power demand projections at the upper end of institutional forecasts are not outliers — they are extrapolations of observed trends.

To contextualize: 50–100 GW of new capacity is equivalent to 50–100 new nuclear power stations. US nuclear plant construction timelines currently average 10–15 years from licensing to operation. The permitting bottleneck alone makes this buildout physically impossible within the required timeframe on terrestrial grids.

Beyond grid limitations, community opposition is emerging as a binding constraint on terrestrial data center expansion. OED tracks a growing pattern of municipal and county-level resistance driven primarily by energy and water consumption concerns. Tucson, Arizona became one of the first US cities to unanimously reject a gigawatt-scale data center proposal, citing generational impacts to energy and water resources.8 This is not an isolated event — documented pushback has occurred in Prince William County, Virginia; Chandler, Arizona; multiple counties in rural Georgia and Texas; and several jurisdictions in Ireland and the Netherlands.9

A 2024 CBRE report found that permitting timelines for hyperscale data centers in the US have extended by an average of 8–14 months compared to 2021 baseline levels, attributable in significant part to community opposition and environmental review requirements.10 For capital allocators, this represents a measurable increase in deployment risk and time-to-revenue for terrestrial data center investments.

2   The Launch Economics Inflection

The orbital data center thesis rests on a single enabling technology: fully reusable heavy-lift launch vehicles. Without a step-change in launch costs, space-based compute remains economically unviable regardless of its physics advantages. OED assesses that this step-change is now underway.

The distinction between expendable and reusable launch vehicles is not merely a cost story — it is a capacity story. With an expendable upper stage like the current Falcon 9, building one vehicle per day for a year yields only one operational upper stage at year-end because each is destroyed on use. With a fully reusable vehicle like Starship, building one per day for a year yields 365 vehicles that accumulate in the operational fleet. This mathematical reality means that annual payload-to-orbit capacity could increase by 100–1,000x even with modest production rates.11

SpaceX's Boca Chica facility includes two dedicated Starship production buildings designed to achieve one vehicle per day at full production tempo.12 Current Falcon 9 payload-to-LEO capacity is approximately 22.8 tonnes per flight; Starship targets 100–150 tonnes. Total global mass-to-orbit in 2023 was approximately 720 tonnes, dominated by Starlink deployments.13 Even at 100 Starship launches per year at 100 tonnes each — a conservative fraction of stated targets — annual capacity would reach 10,000 tonnes, representing an order-of-magnitude increase over current global throughput.

SpaceX is not operating alone. Blue Origin's New Glenn completed its first launch in February 2025; Rocket Lab's Neutron targets 2026; Relativity Space's Terran R is in development. Each targets partial or full reusability.14 Even with a healthy dose of skepticism applied to individual program timelines, the aggregate trajectory is clear: the era of abundant, low-cost mass-to-orbit is approaching.

OED's analysis, corroborated by the European Commission's Ascend study conducted by Thales Alenia Space, ArianeGroup, and university partners, identifies approximately $500 per kilogram to LEO as the economic breakeven point at which orbital data centers become cost-competitive with terrestrial equivalents on a total-cost-of-ownership basis.15

The comparison framework centers on three primary cost categories for terrestrial solar-powered data centers: permitted land (often the largest single cost, particularly in North America, where land costs are linked to energy pricing); battery storage (required to maintain operations through nighttime hours, representing a significant capital and replacement expense); and solar cells themselves. In orbit, permitted land and battery storage costs are eliminated entirely — orbital real estate is effectively free, and sun-synchronous orbit provides continuous solar exposure — while the solar cell requirement is reduced by approximately 6x because one square meter of solar panel in space produces six times the energy of one square meter on Earth due to the absence of atmospheric attenuation, weather interference, and day-night cycling.3

The offsetting cost is launch: transporting chips, solar panels, and radiators to orbit. The breakeven occurs when launch cost falls below the combined terrestrial costs of permitted land, battery storage, and 6x the solar capacity.

Current Falcon 9 marginal launch cost is approximately $2,700–$3,000/kg to LEO.18 SpaceX's aspirational Starship target is $10–$100/kg at full operational tempo. Independent analysts estimate $200–$500/kg as achievable within 3–5 years of Starship entering commercial service.19 The Ascend study's independent breakeven estimate of approximately €400–600/kg closely aligns with OED's $500/kg threshold.15

3   The Physics Case — Energy, Cooling, and Scale

The orbital data center thesis ultimately rests on three physics advantages that are not subject to debate — only the engineering required to exploit them at scale remains uncertain.

Solar irradiance in low Earth orbit is approximately 1,361 W/m², roughly 36% higher than peak terrestrial irradiance at the surface and available approximately 95% of the time in a dawn-dusk sun-synchronous polar orbit.3 Terrestrial solar installations achieve an average capacity factor of approximately 25% in the US due to atmospheric attenuation, weather, and day-night cycling.22 The net effect is that one square meter of solar panel in space produces approximately six times the energy of one square meter on Earth. This is not an engineering claim — it is measured physics, confirmed by NASA SORCE mission data and decades of satellite solar array performance.

The concept of space-based solar power was first articulated by Isaac Asimov in the 1940s and later formalized by Peter Glaser in 1968. The historical limitation was the energy loss in beaming power from orbit to Earth's surface. The orbital data center thesis sidesteps this entirely: rather than transmitting energy to Earth, you move the compute workload to the energy source and transmit only the results — data, not power. Data transmission via optical inter-satellite links and laser ground stations is orders of magnitude more efficient than power transmission via microwave or laser energy beaming.23

Terrestrial data centers consume approximately 40% of total energy input for cooling alone, using massive HVAC systems and an average of roughly 5 million gallons of water annually per facility.16 In the vacuum of space, heat can only be dissipated through thermal radiation — convection and conduction require a medium that does not exist in orbit. While this imposes engineering constraints (large radiator surfaces are required), the "ambient" temperature of deep space is approximately 3 Kelvin (−270°C), providing an essentially infinite heat sink at zero marginal energy cost.

At scale, radiative cooling requires substantial infrastructure. Current orbital data center concepts envision radiator arrays measured in square kilometers for gigawatt-class facilities — structures far larger than anything previously constructed in space. However, the fundamental thermodynamic principle is well-established: the Stefan-Boltzmann law governs radiative heat transfer, and the engineering challenge is one of scale and deployment, not physics validation.24

To ground this analysis in physical reality: current concepts for a five-gigawatt orbital data center cluster envision Starship-class spacecraft delivering 100-ton modules of compute hardware, connecting to a solar array approximately 4 kilometers across and a radiator surface of approximately 1 kilometer — expanding to 16 km² at full deployment. These are modular, additively constructed systems, not monolithic structures. Each module is independently functional, with the cluster scaling over years of successive launches.25

For context, this solar array would be approximately 40x larger in linear dimension than the International Space Station's solar arrays, but would be deployed as thin-film panels rather than rigid structures. It would not be visible to the naked eye from Earth's surface, nor would it measurably affect sunlight reaching the ground — the area is negligible relative to Earth's cross-section.

4   On-Orbit Proof Points — What Has Been Demonstrated

The orbital data center concept has moved from whitepaper to demonstrated hardware. OED tracks the following milestones as material evidence for institutional evaluation.

5   Competitive Landscape — OED Space-Based Computing Sector Map

OED tracks 51 entities within the Space-Based Computing sector (Sector 17). The landscape spans pure-play orbital data center companies, hyperscaler research programs, defense-oriented edge compute providers, and sovereign initiatives. Below is the top-tier entity map as of this assessment.

6   Sector Risk Assessment

OED identifies seven primary risk factors for the Space-Based Computing sector. These risks apply sector-wide rather than to any individual entity.

7   The Geopolitical Dimension — Resource Competition and Strategic Compute

At its root, the largest driver of large-scale interstate conflict is competition for finite resources. Over the coming decades, the most consequential resource competition between nations will center on energy and water for compute infrastructure — the physical substrate of AI capability, economic competitiveness, and military advantage.

The US Department of Defense's 2024 National Defense Strategy explicitly identifies compute infrastructure as a strategic asset critical to maintaining technological advantage.34 The intelligence community's investment arm, In-Q-Tel, has made direct investments in orbital computing companies, confirming that the national security establishment views space-based compute as strategically relevant — not merely commercially interesting.35

OED assesses that the geopolitical calculus strengthens the orbital data center thesis in two ways. First, it provides an additional demand driver beyond commercial AI workloads: sovereign compute requirements that are insulated from terrestrial infrastructure vulnerabilities (grid attacks, natural disasters, land-use constraints). Second, it reframes the competitive dynamic: nations that develop orbital compute capability gain access to a theoretically limitless energy source — the sun — without competing for the fundamentally finite energy, water, and land resources on Earth's surface.

8   Investment Positioning & Monitoring Framework

For allocators seeking historical analogs, the orbital data center thesis shares structural characteristics with three precedent infrastructure transitions. Subsea fiber optic cables (1990s): initially dismissed as impractical, required massive upfront capital, became critical infrastructure generating recurring revenue at high margins — today carrying >95% of intercontinental data traffic. Cloud computing (2006–2012): AWS was treated as an Amazon side project for six years before reaching $1B revenue; it now generates >$100B annually and restructured the entire enterprise computing industry. Commercial launch reusability (2015–present): SpaceX's Falcon 9 booster landing was widely mocked until it worked, then restructured the entire launch industry within five years. In each case, the core physics and economics were sound; execution risk was the gating variable, and early movers captured disproportionate value.

The Space-Based Computing sector is currently at the "first hardware on orbit" stage — analogous to the first transatlantic fiber cable (TAT-8, 1988) or the first AWS EC2 instances (2006). The transition from demonstration to commercial revenue is the next critical inflection, and OED expects this to occur with Starcloud-2 in late 2026 if the unit economics thesis is validated.