Rhino poachers in South Africa's Sabi Sand Nature Reserve have a new adversary: a drone that does not wait for a human to watch the feed. The Matrice 4T, DJI's industrial quadrotor, is running AI-based threat detection directly on board, combining thermal imaging, long-range zoom, and an onboard model that flags suspicious movement and alerts rangers before poachers disappear into the bush. No footage is streamed back to a server. No analyst in a control room is watching frames. The drone decides, and the Reaction Team responds.
This deployment is not a concept demo. It is running right now in the Sabi Sand, according to DJI's developer contest page, and it is the most concrete illustration of a problem that the FCC's December 2025 ban on new DJI drones completely misses.
The FCC added DJI to its Covered List on December 22, 2025, blocking equipment authorization for new models of foreign-made drones and critical components. The agency's stated concern: unauthorized surveillance, sensitive data exfiltration, supply chain vulnerabilities. DJI sells more than half of all U.S. commercial drones, and more than 80 percent of the nation's 1,800 state and local law enforcement and emergency response agencies that operate drone programs rely on DJI technology. The ban is framed as a national security response to Chinese-made hardware that could send data back to Beijing.
The problem with that framing is the Sabi Sand. DJI has been quietly building out onboard AI capabilities, edge inference running directly on the drone without a cloud connection, that make the data transmission concern largely beside the point. The vulnerability the FCC is trying to close is one where the drone is a passive sensor that sends footage home. The threat model that actually matters now is a drone that is a thinking machine, making autonomous decisions in the air.
The technical substrate is DJI's Manifold 3, a computing module roughly the size of a deck of cards that delivers up to 100 TOPS of AI compute, the kind of processing power found in a high-end laptop, while drawing 25 watts and weighing 120 grams, according to DJI's developer documentation. It measures roughly 4 by 2 by 1.5 inches, operates in temperatures from minus 4 to 122 degrees Fahrenheit, and carries an IP55 weather resistance rating, per DroneXL. The module snaps onto the Matrice 4 Series and older Matrice 400 platforms. Submissions to the Enterprise Onboard AI Challenge must target the Matrice 4 Series, Matrice 4D Series, Dock 3, Matrice 400, or Manifold 3 platform, and the program accepts entries through May 10, with finalists in May and winners in June.
The Sabi Sand deployment uses the Matrice 4T variant, which combines a thermal imaging sensor with a long-range optical zoom. The AI detection model identifies threat signatures in real time and pushes alerts to the Reaction Team, the reserve's anti-poaching rapid response unit. The system works because it does not need the cloud. In a wildlife reserve with intermittent connectivity, depending on a remote link is not reliable.
The same edge inference logic applies to the other use cases in the DJI program: power line inspection, where a drone needs to flag anomalies without waiting for a remote analyst, and infrastructure monitoring, where a ground control station may be out of range. The common thread is speed and independence from connectivity.
What the FCC ban actually blocks is new hardware entering the U.S. market. It does not address the DJI drones already deployed across American law enforcement agencies, and it does not address the software layer that determines what those drones can do autonomously. A drone running onboard AI at the edge is a fundamentally different product than a drone that streams video to a server, and the regulatory framework treating them the same is the gap.
The FCC's underlying assumption is that Chinese-made drones pose a surveillance risk because they could exfiltrate data. That assumption was reasonable when drones were flying cameras. It is less coherent when the drone's AI is running on proprietary hardware that never transmits raw footage in the first place.
This is not a defense of DJI. The Sabi Sand deployment raises legitimate questions: who trained the detection model, who validated its accuracy, and who can audit what it records about animals versus humans? Edge AI deployment in wildlife protection is a genuinely interesting use case. Edge AI deployment by a Chinese-manufactured drone in U.S. critical infrastructure is a genuinely interesting policy problem. These are not the same question, and conflating them does not answer either one.
The prize for the challenge includes Manifold 3 units, DJI camera gear, and catalog placement in DJI's global enterprise sales channels. Some of those solutions will eventually run on drones already in U.S. law enforcement hands, on hardware that the FCC ban did not touch.
The FCC banned new DJI hardware. The real action is in the software already running on what is already there.
† Add footnote: "DJI market share claims — more than half of U.S. commercial drones; more than 80 percent of 1,800 law enforcement and emergency response agencies — are source-reported and have not been independently verified."
