‘Resilience comes from designing for disconnection, not assuming more connectivity’: The future of battlefield AI systems lies in both coordination and local capability

The US military's integration of AI systems across its military branches is facing challenges on the battlefield in the Middle East. Drones, sensors, AI-targeting and decision-making all rely on continuous connectivity and compute resources. A single AI data center destruction can severely impact an army's functionality. The issues extend beyond just hardware.

Drone adaptability becomes critical when communication is disrupted or hardware is compromised. How can a drone adjust its mission parameters without connectivity? How can AI models continue processing without access to crucial battlefield data? The design of battlefield AI must enable continued operation even when communications fail or hardware is disabled.

Iran's use of traditional data centers in the Middle East has exposed their vulnerabilities. These data centers are easy targets for missiles and can bring entire systems down with a single hit. Palantir is addressing this by deploying shipping containers filled with Nvidia hardware as decentralized compute solutions. These containers can be quickly deployed on the frontlines, connecting directly into operational workflows for local processing.

Scaleout, a company specializing in edge AI and federated learning, is building infrastructure tailored for battlefield conditions. Their platform allows for distributed data processing without centralization, enabling local inference using approved models even when connectivity is lost. Redundancy is key, with multiple aggregation points ensuring uninterrupted operation if one point fails. The AI workload is designed to not assume constant link availability.

Centralized AI networks offer consistency, substantial compute, and centralized security, while distributed systems reduce latency and allow local functions to continue during disconnections. However, distributed systems risk fragmentation, with partial or stale information potentially leading to risky targeting decisions and suboptimal logistics.

A hybrid approach, combining central coordination when available with clearly bounded local capability when disconnected, is recommended. Human decision processes and command-and-control authority remain paramount.

During disconnections, AI drones and sensors adapt by transitioning to a pre-approved failsafe mode based on the mission. Connected platforms share detections, receive tasking, and contribute to the broader operational picture. When disconnected, a drone continues local inference using its last approved model, caching detections and telemetry until connectivity returns. Onboard processing keeps perception and navigation functional, but without external context, new instructions, or fleet coordination.