Cloud AI is incredible — until the network goes down. For DOD, "the network goes down" is not an edge case; it is a design requirement.
DIL is the norm, not the exception
For DOD operational environments, intermittent connectivity is expected. Disconnected, Intermittent, and Limited-bandwidth (DIL) environments are the norm for:
- Tactical units and deployed forces
- Submarine operations and remote sensing stations
- Classified networks that are air-gapped by design
Joint communications doctrine (JP 6-0, Joint Communications System) is the keystone reference for how the joint force plans, executes, and assesses communications system support — the substrate every “AI at the edge” story still rides on. The Department’s Data, Analytics, and AI Adoption Strategy, meanwhile, explicitly calls out interoperable, federated infrastructure — a polite way of saying your model has to run where the mission is, not only where the fiber is.
| Dimension | Typical cloud AI | Edge / DIL reality |
|---|---|---|
| Connectivity | Persistent, high-bandwidth | Intermittent, limited, or none |
| Model choice | Frontier-scale when online | Sized to hardware and power |
| Data residency | May traverse public paths | Stays inside the enclave |
A practical four-layer framework
After building and deploying a local AI computing cluster using Apple Silicon and mobile device processors, I use this framework for edge AI under disconnected constraints.
Layer 1 — Model selection for constraints
Not every model belongs at the edge. Selection criteria for disconnected deployment differ from cloud deployment. You optimize for:
- Inference speed on limited hardware
- Model size relative to available storage
- Accuracy-per-parameter efficiency
- Offline operation — including periods without internet for model updates
Quantized, distilled, and small language models (SLMs) become workhorses — not because they beat frontier models on benchmarks, but because they run in the environment. Practical introductions appear in industry guides such as Hugging Face on quantization and the ONNX Runtime documentation for portable inference — useful companions to any edge bake-off.
Layer 2 — Runtime architecture
The agent runtime must be lightweight, self-contained, and free of fragile external dependencies. Language choice matters enormously:
- A Python runtime with a deep dependency tree is a liability at the edge
- A compiled Rust or C++ binary with statically linked dependencies is an asset
The runtime should include local model inference, a task queue for asynchronous work, a local data store for context persistence, and a sync protocol for when connectivity returns.
Layer 3 — Data sovereignty
In disconnected classified environments, data cannot leave the enclave. The AI system must process, store, and act entirely within the local security boundary:
- Inference happens locally
- Training data stays local
- Model updates arrive through secure, out-of-band mechanisms — not over-the-air from a public cloud
NIST’s AI Risk Management Framework and the companion Generative AI Profile stress data lineage, access control, and monitoring — all harder in DIL settings, which is why this layer is non-negotiable.
Layer 4 — Graceful reconnection
When connectivity returns, the system needs a disciplined protocol for:
- Syncing local state with centralized systems
- Uploading logs and metrics for monitoring
- Receiving model updates or configuration changes
- Resolving conflicts between local and central data
This is the layer many edge architectures neglect — and where operational resilience is won or lost.
Demonstrated, not hypothetical
This framework is not theoretical. I have run it on a cluster of Apple Silicon devices and mobile processors, showing that useful AI inference fits in a rucksack. The models are not GPT-4 — and they do not need to be. They need to parse reports, classify data, flag anomalies, and support decision-making where the alternative is no AI at all.
Takeaway
The DOD's AI ambitions will succeed or fail based on whether AI works where warfighters operate — not only where the network is fast.
Further reading
- JP 6-0 — Joint Communications System (JCS) — doctrinal foundation for joint comms in contested and constrained environments.
- DoD Data, Analytics, and AI Adoption Strategy (PDF) — federated infrastructure and enterprise alignment for analytics and AI.
- NIST AI RMF Generative AI Profile — structured risk considerations for generative models, including data and monitoring — portable to edge deployments.
- ONNX Runtime documentation — cross-platform inference and deployment patterns teams use when cloud APIs are not an option.
- CISA Secure by Design — product-level security expectations that still apply when the “product” is a fielded inference stack.