Move the workload only when the complete system improves
The relevant comparison is a common task delivered to a defined service level. An orbital design must justify launch mass, radiation mitigation, thermal rejection, intermittent links, maintenance difficulty and replacement risk. The benefit may come from processing near an orbital data source, reducing downlink volume, creating a physically distinct custody domain or improving resilience for a specific mission.
The correct question is not “Can this run in space?” It is “What measurable value appears only when this task runs there?”
Evaluate value, constraints and recoverability together
Orbit becomes more plausible when large source data already exists on a spacecraft and local processing materially reduces transfer.
Interactive terrestrial users may be poorly served; time-sensitive orbital decisions may benefit from local execution.
Burst, archival and scheduled workloads can differ sharply from continuous high-density compute.
Workloads need explicit recovery points, deadlines, integrity requirements and degraded modes.
Sovereignty may justify a distinct infrastructure boundary, but jurisdiction and ground access still apply.
Rapidly obsolescent workloads face a stronger penalty when replacement requires launch.
Start with bounded, delay-tolerant tasks
- Sensor-data preprocessing: filter, compress or extract features before downlink, with the original retained or recoverable.
- Integrity-checked archival storage: controlled retention or custody where access frequency, recovery time and physical fault domains are explicit.
- Event-driven edge inference: run a frozen, validated model near an orbital data source and transmit a compact result plus provenance.
- Coordination and provenance: record authorised machine events, state transitions or custody receipts across a distributed mission.
- Resilience replicas: maintain a carefully governed recovery copy outside a terrestrial failure region.
Each class needs its own hardware, link, thermal, security, regulatory and economic proof. They should not be combined into one generic “AI in space” claim.
Some workloads should remain on Earth
Tasks are weaker candidates when they require continuous low-latency interaction with terrestrial users, frequent physical upgrades, unrestricted access to enormous terrestrial datasets, dense sustained power without adequate radiator area, or maintenance that cannot tolerate mission-level delay.
A negative decision is valuable. It narrows the orbital architecture to tasks that can survive the engineering and economics review.
Benchmark orbit against strong terrestrial alternatives
- Freeze inputs, outputs, quality thresholds and deadline.
- Measure useful work completed, not nominal processor capability.
- Include transfer, retry, throttling, failure and recovery.
- Compare grid-connected, dedicated-generation and renewable-linked terrestrial systems.
- Include complete orbital lifecycle cost and uncertainty.
- Publish the cases where terrestrial infrastructure wins.
Proof of Pulse is developing this workload-first approach so future economics and sustainability claims can be traced to reproducible service results.