The Race to Orbit: Compute Constellations, Strategic Dependency, and the New High Ground

Executive Summary

The driver is straightforward: terrestrial data centres are hitting hard limits on power, water, and planning consent. AI compute demand is growing roughly 50 percent annually. Space offers uninterrupted solar power, passive radiative cooling, and no land-use politics. The economics are not yet proven at gigawatt scale, but multiple well-resourced actors have concluded that the trajectory justifies the investment.

For defence and security audiences, the implications extend well beyond energy efficiency. Orbital compute constellations are dual-use infrastructure by design, already attracting Pentagon interest for missile tracking, ISR data processing, and autonomous space operations. The organisations that establish operational dominance in this layer over the next 24 to 36 months will hold structurally advantaged positions in both commercial AI competition and any future contested scenario. Europe and the UK currently have no programme at Sunrise or ADA scale with a dedicated orbital compute mandate, and are materially lagging in both industrial base and deployment timeline.

Strategic Context: Why Now

The terrestrial bottleneck

AI-related electricity consumption is projected to grow 50 percent annually through 2030. Microsoft, Google, and Amazon have collectively committed over $500 billion to new data centre construction, yet power purchase agreements, water rights, and planning consents are already binding constraints. UK and European grid connection queues stretch to 2030 and beyond; some US states have imposed moratoriums on new hyperscale facilities. The orbital solution addresses these constraints structurally: satellites in sun-synchronous orbits receive near-continuous solar illumination, radiate waste heat directly into space without cooling infrastructure, with proponents projecting long-run marginal compute costs well below terrestrial grid rates of $0.04 to $0.10 per kilowatt-hour, though no independent techno-economic analysis yet substantiates specific figures at scale.

The NVIDIA validation moment

At GTC 2026 on 16 March, NVIDIA CEO Jensen Huang announced the Vera Rubin Space-1 Module: a radiation-hardened AI compute platform that NVIDIA claims delivers up to 25 times the inference performance of the H100 GPU in orbital conditions — a vendor benchmark figure for space-based inferencing in constrained scenarios, not a general-purpose replacement specification. Six partners immediately confirmed deployment plans: Aetherflux, Axiom Space, Kepler Communications, Planet Labs, Sophia Space, and Starcloud. NVIDIA's entry legitimises the market and positions it as the dominant early-entrant vendor. The ecosystem lock-in dynamics mirror those it established in terrestrial AI, though it is too early to describe this as a universal or exclusive standard. If that dominance consolidates, a chip-level compromise would affect multiple operators simultaneously — a concentration risk scenario with no direct analogue in terrestrial data centre security.

The Blue Origin entry and the Bezos ecosystem

Project Sunrise, filed with the FCC on 19 March 2026, proposes up to 51,600 data-processing satellites paired with TeraWave, a 5,408-satellite high-throughput connectivity constellation announced in January 2026. The two programmes function as a system: TeraWave provides the backbone, Sunrise provides the compute layer.

The strategic logic is distinctive. Blue Origin's ownership structure gives it a captive hyperscale cloud customer in AWS — something no other space company possesses. AWS generated over $107 billion in 2024 and was on a strong growth trajectory through 2025. If orbital infrastructure could handle even five percent of AWS compute workload, the avoided capex on land, utilities, water rights, and cooling infrastructure runs to billions annually. No other actor combines rocket capability, broadband constellation, compute constellation, and captive hyperscale cloud in a single ownership structure. That combination is the source of Blue Origin's structural advantage in this race.

The Competitive Landscape: All Active Players

The table below covers all twelve organisations with active orbital compute programmes as of March 2026, ranging from operational hardware in orbit to regulatory filings and funded development. This landscape did not exist in any commercial form two years ago.

Operational hardware: what is already in orbit

Three organisations are already running compute hardware in LEO. Starcloud (formerly Lumen Orbit) placed the first NVIDIA H100 GPU in space in November 2025 and trained the first large language model in orbit the following month; Starcloud-2 (October 2026) will carry Blackwell hardware and deploy the first AWS Outposts in space. Kepler Communications launched ten optical compute nodes on 11 January 2026, operating as an IP-based mesh network to Space Development Agency standards. Axiom Space deployed two orbital data centre nodes on the same flight, explicitly targeting national security workloads including multi-sensor fusion and space threat tracking.

The SpaceX / xAI full-stack ambition

SpaceX merged with xAI on 2 February 2026 at a $1.25 trillion combined valuation and filed an FCC application for up to one million orbital data centre satellites. SpaceX's own filing projects 100 gigawatts of AI compute capacity, a figure derived from its stated 100 kilowatt per satellite design target rather than from an independently reviewed engineering model. On 21 March 2026, Musk revealed three further developments that materially escalate the programme.

Terafab is a joint SpaceX/Tesla/xAI chip fabrication facility in Austin. Musk described its target as one terawatt of annual processor output, which he claimed is 50 times the combined output of all current advanced chip manufacturers, at an estimated $20–25 billion. These are Musk's own projections from a company event rather than independently verified engineering targets. He framed Terafab as the explicit prerequisite for the orbital programme: without it, there are no chips and no constellation. If delivered at anything near the claimed scale, it would represent a strategically significant fraction of global advanced semiconductor output and constitutes a direct challenge to TSMC's dominance at a moment of acute geopolitical sensitivity around Taiwan.

AI Sat Mini is the initial satellite design: at scale with Starship V3 (124 metres tall), it is over 170 metres long, carries 100 kilowatts of power for onboard AI processors, and features a 100 square metre radiator. The Mini designation signals intent — future satellites are planned at one megawatt. SpaceX is also developing the D3, a custom processor optimised for orbital temperatures and radiation, completing a full proprietary stack: chip design, fabrication (Terafab), launch (Starship), satellite (AI Sat Mini), connectivity (Starlink V3 at 1 Tbps), and AI model layer (xAI/Grok). At the same event, Musk stated Starlink laser links will exceed Blue Origin's TeraWave 6 Tbps specification.

The IPO dimension is inseparable from the programme. SpaceX targets a 2026 listing at a valuation above $1.5 trillion, with proceeds earmarked for the orbital build-out. Wall Street has not yet priced this in; successful FCC authorisation and early deployment would trigger a significant re-rating.

China, the energy-compute players, and European response

ADA Space's Three Body Computing Constellation launched its first twelve satellites in May 2025, each carrying 100 Gbps optical links and AI accelerators delivering 744 trillion operations per second, targeting a 2,800-satellite distributed supercomputer in LEO. Unlike US commercial programmes, this is explicitly state-directed: a national asset where AI, ISR, and communications converge under sovereign control, with no commercial return requirement and no FCC regulatory exposure.

Two US startups are pursuing a distinct energy-compute convergence model that deserves separate attention. Aetherflux (Galactic Brain, Q1 2027 target, $60 million raised including Andreessen Horowitz and Breakthrough Energy) combines space solar power beaming with on-orbit compute on the same satellite infrastructure and carries early military interest in its power-beaming capability for contested environments. Sophia Space ($10 million seed, February 2026) has developed a passive radiative cooling approach for its modular TILE compute platform that is technically differentiated from every other player; it has attracted Pentagon interest specifically for missile warning and tracking. Both are NVIDIA Vera Rubin Space-1 partners.

Europe's ASCEND programme (Thales Alenia Space, demonstration mission 2026) is government-backed, focused on data sovereignty and net-zero framing, and at least five years behind the commercial trajectory. The UK has no programme at comparable scale with a dedicated orbital compute mandate, though smaller ISR and edge compute work exists within ESA and UKSA-linked programmes. Lonestar Data Holdings is pursuing sovereign data storage rather than AI compute, targeting the disaster recovery market with a long-term lunar deployment concept.

Security and Defence Implications

Dual-use by design

Orbital compute constellations are not neutral infrastructure. Axiom Space frames its nodes explicitly around national security workloads. Sophia Space has attracted Pentagon interest for missile tracking. Kepler operates to SDA communication standards. Aetherflux is developing energy-beaming capability with stated early military interest. The distinction between commercial and military orbital compute is already becoming as difficult to maintain as it proved for communications constellations. Ukraine demonstrated how quickly that boundary collapses under operational pressure.

Amplified attack surface

The proposed constellation sizes, ranging from 2,800 in the Chinese programme to one million in SpaceX's FCC filing, represent a qualitative shift in cyber-physical attack surface. Vulnerabilities documented across mega-constellation security research apply with amplified force to compute constellations:

• →Compromise of a compute satellite affects not only communications but active data processing, model inference, and potentially autonomous decision-making pipelines. The consequences of compromise are an order of magnitude more severe than for a communications node.
• →Inter-satellite laser links, the primary backbone for all major programmes, represent high-value targets for jamming, spoofing, or kinetic interdiction. They are also harder to protect than RF links given their directionality requirements.
• →Supply chain integrity is more critical for compute satellites than connectivity satellites. A backdoor in a GPU or AI accelerator deployed in orbit is both harder to detect and harder to remediate. The NVIDIA ecosystem's emergence as the dominant hardware platform concentrates this risk: a compromise at the chip design or firmware level would affect multiple operators simultaneously.
• →The harvest-now, decrypt-later threat applies to orbital compute traffic with particular force. Adversaries intercepting encrypted AI inference traffic today may extract sensitive model outputs or classified inputs once quantum capability matures. No current orbital compute programme has published a post-quantum cryptography transition plan.

Sovereignty and structural dependency

The AWS and Blue Origin relationship is the clearest example of a strategic dependency that has not been seriously examined by European or UK defence planners. If AWS becomes the dominant provider for orbital compute, and Blue Origin controls the launch and operational infrastructure, European governments will have limited leverage over a critical node in their AI and ISR architecture.

The window to establish sovereign or allied alternatives is open now and will narrow significantly as US commercial programmes reach operational scale in 2027 to 2028. Once defence procurement decisions are made around a specific orbital compute provider, switching costs become prohibitive. The ASCEND programme is at least five years behind the commercial trajectory. The UK has no programme at comparable scale with a dedicated orbital compute mandate, though smaller ISR and edge compute experiments exist within ESA and UKSA-linked work.

The Chinese programme adds a competitive dimension that is structurally different from the US commercial race. ADA Space's Three Body Computing Constellation is not subject to commercial pressures or FCC regulatory oversight. It is a state-directed capability being developed in parallel with China's Guowang and Qianfan communications constellations. The convergence of communications, compute, and ISR in a single state-controlled orbital architecture represents a significant asymmetric capability.

Regulatory and governance vacuum

The FCC has created what it describes as the friendliest regulatory environment in the world for the space industry, introducing new modular license types for variable trajectory and multi-orbit systems. No tailored licensing category yet exists specifically for orbital data centres — functionally, these systems are licensed under existing fixed-satellite or experimental categories. The gap is conceptual and environmental: there is no framework calibrated to data-centre-scale constellations in terms of debris mitigation standards, environmental review, or spectrum allocation for inter-satellite compute links.

Combined proposals currently before the FCC total over 1.3 million satellites from US programmes alone. Adding Chinese, European, and other national programmes, the cumulative orbital population proposed for the next decade exceeds anything in existing space traffic management frameworks. The Kessler cascade risk analysis that applies to communications constellations applies with equal force to compute constellations operating in the same orbital shells.

Questions for Defence and Policy Planners

• 1. What doctrine governs the use of commercially operated orbital compute infrastructure during a conflict or crisis? The Starlink precedent established that private operators retain unilateral authority over service availability, geographical restrictions, and use-case limitations. Orbital compute providers will face the same pressures with higher stakes.
• 2. How are sovereign AI model training and inference workloads protected in an orbital environment? If classified defence AI migrates to orbital infrastructure, the security architecture must operate in a contested and degraded environment. No current certification framework addresses this requirement.
• 2a. What are the implications of Terafab for allied chip supply security? If SpaceX achieves anything approaching the claimed one terawatt of annual processor production, it would represent a strategically significant fraction of global advanced semiconductor output under Musk control. Combined with the D3 custom space chip, this creates a chip dependency risk for any allied nation whose orbital compute infrastructure runs on Musk-ecosystem hardware.
• 3. What is the UK and European response to the structural advantage the Bezos ecosystem holds through AWS and Blue Origin? This is not a market question. It is a strategic dependency question that requires explicit government policy.
• 4. At what constellation size does orbital debris risk become a compounding strategic threat? Combined proposals before the FCC already exceed 1.3 million satellites. A debris cascade in a critical orbital shell could simultaneously deny access to communications, compute, and ISR infrastructure.
• 5. Who holds targeting authority over adversary orbital compute nodes processing autonomous weapons guidance or time-critical ISR data? The legal and doctrinal framework for kinetic or cyber response to orbital compute infrastructure is entirely undefined.
• 6. How should NVIDIA's position as dominant early-entrant vendor for orbital compute factor into supply chain security assessments? If its current six-partner ecosystem consolidates into a de facto standard — which is not yet certain — a compromise at the chip or firmware level would affect multiple operators simultaneously, a concentration risk scenario with no direct analogue in terrestrial data centre security.

Assessment

The orbital compute race is real, accelerating, and under-examined by defence and security institutions in the UK and Europe. The week of 19–21 March 2026 alone produced three major developments: Blue Origin's Project Sunrise FCC filing, NVIDIA's Vera Rubin Space-1 announcement at GTC, and Musk's Terafab and AI Sat Mini reveal in Austin. The pace of disclosure has outrun the pace of strategic analysis.

Blue Origin's Project Sunrise retains a structural advantage through AWS vertical integration. But the Terafab announcement materially changes the SpaceX risk picture. If SpaceX achieves sovereign chip fabrication at the scale Musk describes, the Musk ecosystem would be the only actor controlling the full stack from silicon to orbit to AI model, at a production scale that no other commercial or national actor could match within a decade. The D3 custom chip compounds this: a proprietary orbital processor would give SpaceX a performance and cost advantage in the orbital compute environment that mirrors its launch cost advantage in the ground segment.

NVIDIA's Vera Rubin Space-1 announcement at GTC 2026 is equally significant. It establishes the hardware ecosystem that will underpin the entire industry, mirrors NVIDIA's role in the terrestrial AI build-out, and creates concentration risk at the compute layer that has no current governance response.

The security vulnerabilities of orbital compute infrastructure are an amplified version of those already documented for communications constellations. The difference is consequence: compromising a compute node in orbit is not merely a denial of service. It is potential access to the AI systems, decision pipelines, and intelligence products that depend on that infrastructure.