Anduril Industries has been constructing what it refers to as a lattice of interconnected autonomous systems—drones, sensors, and ground vehicles—in a warehouse complex in Costa Mesa, California. These systems are intended to exchange information and make tactical decisions without waiting for a signal to travel to a server and back. Based in part on its architecture, the business was awarded a $250 million AFWERX contract in 2024. It’s not solving an exotic problem. Since the conflict in Ukraine painfully revealed how reliant contemporary military technology has become on connectivity that an adversary with strong electronic warfare capabilities can destroy in a matter of minutes, every US military planner has been gazing at this fundamental issue.
For this reason, edge AI—artificial intelligence that analyzes data locally on hardware instead of transferring it to a cloud or satellite link for analysis—has emerged as one of the fastest-growing investment areas in the defense sector. Within ten years, the military edge AI industry is expected to reach approximately $39 billion from its estimated $6.8 billion in 2024. These figures show a shift in procurement that began before to Ukraine but picked up speed after the unavoidable reality of GPS denial and Starlink jamming on the battlefield became apparent. The weapons system, drone, or autonomous car must be able to operate outside the network if it cannot be trusted.
Over the last three years, significant advancements have been made in the hardware that enables this. Ruggedized AI inference chips, such as Qualcomm’s AI 100 and NVIDIA’s Jetson Orin, can run complex computer vision models on a power budget measured in watts rather than kilowatts, allowing them to be embedded in systems tiny enough to fit inside a drone that weighs a few kilos. The difference between the capabilities of a local system and those of a cloud-connected system has significantly decreased. An incoming danger must be detected, classified, and engaged by a counter-drone system in less than 200 milliseconds, which is faster than a human can consciously understand the situation. This processing may now be done on-device. That would not have been feasible two years ago due to the latency caused by a round-trip to a cloud server.
For practical reasons, the counter-drone application has emerged as the most obvious short-term driver of investment: low-cost commercial drones converted for military use have become a recurring tactical issue that traditional air defense systems were not built to address at scale. The FPV drone tactics that both sides have developed and the Shahed-series drones that are widely used in Ukraine represent a type of threat that is inexpensive to manufacture, challenging to track with legacy radar, and numerous enough to overwhelm response capacity if each engagement requires human approval for each individual intercept. When the drone is three kilometers away and closing in, edge-AI targeting systems that can automatically detect and queue threats for response—even if a human still approves individual engagements—compress the decision timeline in ways that matter.
The formal architecture attempting to make this consistent across services is the Pentagon’s CJADC2 initiative, or Combined Joint All-Domain Command and Control. Procurement criteria have been shaped in a way that favors companies developing hardware-software stacks intended for disconnected environments due to its stated requirement for AI-enabled processing at the tactical edge, independent of cloud dependency. Several US Army units have implemented Palantir’s AI Platform for Defense. In order to integrate edge AI sensor processing into their current weaponry platforms, Northrop Grumman and Raytheon have both undertaken large acquisitions. For embedded autonomy in aircraft systems, Lockheed Martin and Shield AI entered into a cooperation.
Instead of being addressed as a footnote, the ethical restraint that permeates all of this should be emphasized clearly. Formalized in 2020, the DoD’s AI Ethics Principles mandate that humans retain “appropriate levels of human judgment over the use of force.” Given how quickly edge-AI systems are expected to process and respond, this demand presents a genuine design tension. The distinction between a system acting on its own assessment in 200 milliseconds and a system making a recommendation in 200 milliseconds is functionally different, but in a communications environment that is rapidly deteriorating, it can become hazy in ways that procurement documents don’t fully address. The deployment over the next few years may yield enough operational data to make such instructions more specific. They might also create situations that compel the query in uncomfortable ways.
