The electromagnetic threat is here, not coming
China has constructed overlapping EW installations across Fiery Cross, Mischief, and Subi Reefs in the South China Sea. CSIS satellite analysis shows at least six paved antenna sites at Mischief Reef, two new radomes at Subi Reef, and five vehicle-mounted jammers connected to fixed antenna arrays. A roofed shelter constructed in 2025 at Subi Reef houses mobile EW units. These systems can jam communications, disrupt radar including AN/SPY-1 on Aegis ships, and geolocate foreign forces operating in the area. The International Federation of Air Line Pilots has issued warnings about Chinese warships conducting GPS jamming across the South China Sea, Philippine Sea, eastern Indian Ocean, and northwest of Australia.
Russia's EW arsenal in Ukraine has reduced effectiveness of GPS-guided weapons by up to 90% and cut drone hit rates in half. The R-330Zh Zhitel jams GPS, SATCOM, and cellular out to 25km. The Krasukha-4 neutralizes LEO surveillance satellites and ground-based radars at 150-300km range and has been used to jam Bayraktar TB2 drone communications. Ukrainian forces have destroyed at least 23 Zhitel systems since February 2022, but Russia continues to deploy them because they work.
Iran claims a 10x increase in drone production since 2025. The Shahed-238 represents a generational leap: turbojet-powered at 500-600 km/h with terrain-contour-matching algorithms and electro-optical terminal seekers capable of autonomous target recognition. Iran is expanding its national network of autonomous drone bases and has developed EW systems for counter-UAV operations.
What EW denial does to cloud-dependent AI
A War on the Rocks analysis from April 2025 on spectrum supremacy concluded that military forces have become 'entirely dependent on the electromagnetic spectrum for intelligence gathering, communications, positioning/navigation/timing, surveillance, and weapons guidance.' Legacy doctrine 'assumes complete EMS dominance,' which is untenable against peer competitors.
When a near-peer adversary jams or degrades the communication links between a tactical node and its cloud AI backend, every capability that depends on that link fails. Object detection fails because frames cannot reach the cloud vision API. Threat analysis fails because LLM queries cannot reach the cloud. Transcription fails because audio cannot be uploaded. The node becomes a camera with a blank screen.
This is not a degraded mode. It is a total loss of the AI capability that justified the system's procurement. The operator reverts to Mark I Eyeball and whatever training they received. The investment in AI infrastructure, cloud subscriptions, and integration work produces zero return the moment the adversary decides to contest the spectrum.
Architectural requirements for EW-resilient AI
The January 2026 DoD AI Strategy memo states AI must operate 'on-board, in real time, and often without any sort of connectivity or centralized compute resources.' Converting that mandate into system requirements produces a specific set of architectural constraints.
All inference must execute on the node itself. Not on a FOB server connected by tactical radio. Not on a cloud endpoint connected by SATCOM. On the device the operator is holding. If the device has power and the software is running, every AI capability must be available regardless of link state.
The compute policy must be classification-aware. In a multi-classification environment, sensitive data cannot leave the device for cloud processing even when links are available. The routing policy must enforce data boundaries at the MDM layer, not rely on operator discipline.
The mesh must survive partial and total link denial. When an adversary jams one frequency band, traffic must shift to alternative link types automatically. When all links fail, the node must continue operating independently and sync when any link recovers. The mesh routing must be power-aware because EW-contested environments often correlate with austere logistics where battery life determines operational endurance.
Evidence must be captured and retained locally. If a detection occurs during a period of link denial, the evidence, the AI assessment, and the operator's decision must all be preserved on the node for later upload and review. A detection that cannot be audited is a detection that cannot support a targeting decision or withstand legal review.
How EdgeLance meets these requirements
EdgeLance was architected for the disconnected case first. Every node runs a full AI inference stack locally: object detection, language model threat analysis, speech-to-text, and segmentation. The platform supports a range of open and approved models selected per mission loadout, from Gemma and Llama to Whisper and customer-provided alternatives.
The compute policy engine routes inference across local hardware, base GPU resources, and approved cloud endpoints based on classification level, link availability, and device capability. When links are denied, the policy defaults to local-only processing with no capability loss. When links recover, cloud resources can augment local inference, but the local stack never stops running.
EdgeLance Mesh routes across seven link types: WiFi, LoRa, Starlink, LTE, satellite phone, BLE, and Iridium SBD. Jamming one band shifts traffic to the others automatically. The routing algorithms co-optimize for throughput and power consumption, extending operational endurance in the austere logistics environments that typically accompany EW-contested operations.
Classification-aware MDM enforces data boundaries at the device level. Stealth mode suppresses all RF emissions when the threat environment demands electromagnetic silence. Every detection, assessment, and operator decision is logged locally with timestamps and hash chains for later audit, even if the evidence cannot be transmitted in real time.
Implications for force design
The electromagnetic threat environment in the Pacific, European, and Middle Eastern theaters is not going to get friendlier. Chinese EW capability is expanding. Russian EW doctrine is being validated daily in Ukraine. Iranian autonomous systems are becoming more sophisticated.
Any AI capability the joint force fields in these theaters must assume that communication links will be contested, degraded, or denied during the periods when that capability is most needed: at the start of hostilities, during maneuver, and at the point of engagement. Systems that lose their AI capability under EW pressure are not combat systems. They are peacetime tools.
The investment in edge AI, local inference, mesh resilience, and disconnected operation is not a hedge against a possible future threat. It is the minimum architectural requirement for any AI system that will be used in the environments where the joint force is most likely to fight.