Orbital AI Compute Pod
Think of this as the "rack" inside the orbital data center.
Physical specs, AI compute configurations, storage, networking, and pod-level power/thermal
Physical & Mechanical
Form Factor
Unpressurized modular pod
Nominal external envelope: ~1.5 m (W) × 1.5 m (H) × 2.5 m (L)
Mounting
Slides onto spine rails with kinematic locating features
Single robotic latch for structural attachment and release
Interfaces (Rear Bulkhead)
- High-voltage DC bus connectors
- Coolant quick-disconnects (supply/return)
- High-speed optical or copper data backplane
- Low-voltage control + telemetry
Compute & Storage (Configurable)
AI Compute
50–150 kW thermal design power per pod (configuration-dependent)
Accommodates GPU/AI-ASIC blades arranged in ruggedized racks. On-pod high-speed fabric (NVLink-class or custom interconnect) enables low-latency communication between accelerators within the pod.
Storage
Multi-petabyte class per pod using radiation-tolerant SSD arrays
Local erasure-coded object store for resilience. Storage capacity scales with mission requirements, supporting everything from real-time data processing to long-term archival.
Networking
Spine-level optical backplane connecting pods within a node. Node-to-node crosslinks via external optical or RF terminals enable distributed computing across multiple orbital nodes, creating a mesh network for data sharing and compute offloading.
Power & Thermal Inside the Pod
Power
• Local DC/DC for rack and board rails
• Hot-swappable power modules behind access panels (for servicing)
Thermal
• Cold plates on accelerators and dense logic
• Pumped liquid loop from pod to node-level radiator manifold
• Internal health monitoring for leak detection and temperature excursion
Radiation Environment & Chip Reliability
Orbit & Shielding Strategy
Nodes operate in Sun-synchronous LEO (~600–700 km) where trapped radiation is manageable but non-trivial.
Each compute pod includes:
- A structural 'vault' around the AI blades (Al / composite + localized high-Z / hydrogen-rich materials where it buys the most)
- Layout and materials tuned to reduce total ionizing dose (TID) and single event effects (SEE) on GPUs, memory and power electronics
COTS AI Hardware with System-Level Mitigation
The concept assumes H100-class, largely commercial GPUs/ASICs, not fully rad-hard bespoke chips. Instead, we:
- Derate operating voltages and temps to give margin against upsets
- Use ECC everywhere (HBM, SSDs, inter-pod fabric) and periodic memory scrubbing
- Add watchdog logic and checkpoint/rollback so a transient fault at the chip level doesn't corrupt long-running training or inference workloads
Redundancy and Graceful Degradation
AI workloads are deployed across multiple pods with redundant model shards and data copies. If a pod experiences persistent radiation-induced faults, it can be logically isolated and down-binned (e.g., used for lower-criticality tasks) without taking down the node. Pods are designed for robotic replacement, so persistent high-dose hardware can be swapped out on-orbit.
Control Electronics & Safety
Critical control paths (power control, thermal control, attitude, comms) use radiation-tolerant components and fault-tolerant architectures (e.g., TMR for key controllers, hardened FPGAs), separating them from the high-density AI compute where occasional soft errors are acceptable and correctable.
Reliability & Servicing
Fault Isolation
Pod-level power and fabric isolation; a failed pod can be powered down without impacting others. This modular approach ensures that hardware failures are contained and don't cascade through the entire node.
Upgrades
Pods can be swapped to upgrade GPUs/ASICs or change workload (AI, compression, storage-heavy, etc.). Standardized interfaces keep spine and power/thermal structure unchanged across generations, enabling technology refresh without redesigning the entire node.