16.3.2 — Orbital AI & Compute Infrastructure — maturity: speculative

Federated Learning Across Constellations

Training shared AI models across multiple satellite constellations without raw data ever leaving the spacecraft or crossing a foreign ground station.

Auto-drafted reference page — content reviewed for shape, individual references not yet link-checked.

Every Earth-observation, signals-intelligence and weather constellation accumulates vastly more sensor data than it can downlink. The bottleneck is not compute or storage — it is the radio link. Federated learning inverts the classical approach: instead of shipping petabytes of raw imagery or RF captures to a ground data centre, each satellite trains a local model increment on-board and transmits only gradient updates — kilobytes, not gigabytes. Aggregation happens either at a designated orbital relay node or, in the most sovereign-friendly architecture, at a nationally controlled ground segment that never hands custody of raw data to a foreign cloud provider.

The national security implication is acute. A state that relies on a commercial constellation operator for AI model training is, in practice, handing that operator's jurisdiction visibility over what the model is learning to detect — enemy ship classes, missile plume signatures, illegal deforestation patterns, refugee movements. Federated learning severs that link. Model weights and gradients are mathematically uninvertible to raw imagery under standard differential-privacy guarantees, meaning the aggregation point can be operated by an ally, a neutral commercial prime, or even a rival, without exposing the underlying intelligence collection.

The architecture remains firmly speculative but is tractable within a five-to-eight year horizon. Inter-satellite links carrying compressed gradient tensors at 10–100 Mbps are already being demonstrated by commercial LEO broadband primes. The missing piece is radiation-hardened, energy-efficient AI accelerators with enough TOPS to complete a meaningful training epoch during a 15-minute orbital pass. Once that compute floor is reached, a federated constellation becomes a continuously self-improving sensor network — one that gets smarter with every orbit without exposing a single raw frame to an adversary's subpoena or export-control regime.

What matters

Raw sensor data never leaves sovereign hardware: federated gradient updates are the only artefact that crosses a network boundary.
Differential privacy budgets (ε < 1.0 under Rényi DP) can be enforced on-board before any gradient is transmitted, bounding inference risk mathematically.
Inter-satellite link bandwidth of 10–100 Mbps is sufficient for gradient payloads; raw EO imagery requires 1–10 Gbps, making classical centralised training an RF impossibility at scale.
Model aggregation schedules double as a classification firewall: a state can federate with allied constellations for shared threat detection while withholding gradients from sensors covering denied areas.

Sovereignty score: 9/10 — A nation that cannot control where its constellation's training data is processed and whose hands touch its model weights has, in practice, outsourced the intelligence value of its entire space programme.

Foreign cloud aggregation creates a legal exposure: US CLOUD Act, EU GDPR and equivalent instruments allow host-jurisdiction authorities to compel access to training data held by commercial aggregators, regardless of the customer's nationality.
Commercial federated-learning platform providers are concentrated in two or three US and Chinese technology stacks; dependency on either creates a single point of geopolitical leverage over a nation's entire AI-enabled sensing capability.
Model poisoning — injecting malicious gradient updates through a compromised node — is a supply-chain attack vector that can silently degrade a constellation's detection performance; sovereign aggregation infrastructure is the only environment where the full gradient audit trail can be held under national cryptographic custody.
Allied constellation federation agreements require reciprocal gradient-sharing protocols; a nation without its own federated-learning infrastructure has no standing to negotiate terms or withhold gradients covering its most sensitive collection areas.

Reference architecture

Payload: Radiation-tolerant AI accelerator module, 4–16 TOPS sustained throughput, 15–40W power envelope; supports on-board stochastic gradient descent with Rényi differential privacy noise injection (ε configurable 0.1–2.0); inter-satellite link transceiver at V-band (60 GHz) or optical (1550 nm), 10–100 Mbps duplex for gradient tensor exchange
Bus class: 12U to 16U cubesat or ESPA-class microsat (80–150 kg), 400–600W total power via deployable solar arrays; thermal management for sustained AI accelerator duty cycle is the primary design driver
Orbit: Sun-synchronous LEO at 500–550 km for EO-federated variants; 53° inclined Walker constellation for broad-coverage RF or multi-domain fusion variants; 6-plane, 36-satellite baseline provides sub-90-minute inter-node contact windows sufficient for gradient aggregation epochs
Ground segment: Sovereign gradient aggregation server cluster (air-gapped from commercial internet); 2–3 national TT&C stations (X-band uplink/downlink, S-band housekeeping); encrypted gradient relay via national fibre backhaul to sovereign AI compute facility; no raw data ever touches ground segment
Software & data
Latency
Cost band
Lead time

References

ESA Phi-Lab: Federated Learning for Earth Observation — ESA's Phi-Lab has actively explored federated learning as a mechanism to train EO AI models without centralising sensitive raw imagery, publishing experimental results on gradient compression and privacy budgets.
ITU-R: Spectrum and Data-Rate Constraints for Non-Geostationary Satellite Systems — ITU-R Report S.2196 documents the spectrum constraints facing LEO constellations, establishing why downlink bandwidth is the primary bottleneck for any ground-centric data processing architecture.
NASA Technical Reports Server: On-Orbit Machine Learning for Earth Science — NASA's survey of on-orbit ML identifies federated and split-learning architectures as the most promising path to continuous model improvement across heterogeneous satellite fleets without full data downlink.
European Union Agency for Cybersecurity (ENISA): Artificial Intelligence Cybersecurity Challenges — ENISA's threat landscape for AI systems explicitly addresses gradient inversion attacks and recommends differential privacy as the primary countermeasure for federated deployments in critical infrastructure.
World Economic Forum: Data Free Flow with Trust — Pathways to Cross-Border Data Sharing — The WEF report identifies federated learning as a key enabler of cross-border AI collaboration that satisfies data-localisation regulations, directly applicable to multi-national constellation programmes.