Geran-2 MS: The AI-Equipped Shahed Changing Loitering Munition Warfare

29 May 2026 · 10 min read

Recovered wreckage of a Geran-2 drone. Photo: Wikimedia Commons, CC BY 4.0

A $20,000 drone with an Nvidia Jetson Orin module now does what a $3 million cruise missile did two years ago. That is not a prediction. It is a description of a variant of the HESA Shahed-136 (the Geran-2 MS) that Ukrainian forces first recovered from a crash site near Sumy in June 2025. The electronics were intact. What they revealed has quietly changed the risk calculus for every military operating near drone launch range [Source].

Loitering munitions are not new. The Shahed-136 has been striking Ukrainian infrastructure since September 2022, following a pre-programmed GPS route to a fixed coordinate and detonating its 50-kilogram warhead on arrival. The Geran-2 MS is something different. It carries an infrared camera, a real-time video encoder, and a machine learning processor that can classify targets autonomously in the terminal phase. It does not need to know exactly where the target is before launch. It can find it.

The jet inside

The core of the upgrade is an Nvidia Jetson Orin system-on-module, originally designed for autonomous robots, delivery drones, and edge artificial intelligence (AI) inference in commercial applications. In the Geran-2 MS, it is wired to a 1920x1080p camera feeding 60 frames per second through an Ezcap video encoder into the Jetson's 40-teraoperation-per-second neural processing pipeline. A Xilinx Spartan field-programmable gate array (FPGA) handles signal filtering and sensor fusion. [Source]

This is not a hack. The hardware architecture shows deliberate engineering: the FPGA pre-processes raw sensor data before it reaches the Jetson, reducing latency for time-critical targeting decisions. The entire payload bay has been reorganised around the compute module, with thermal management (a heatsink and forced-air channel) that suggests sustained operation, not a last-second gimmick. [Source]

The Automatic Target Recognition (ATR) pipeline works like this: the drone loiters at altitude, the camera feeds continuous wide-area video into the Jetson, the neural network identifies objects that match trained signatures (vehicle convoys, air defence radars, fuel depots, exposed troop positions) and transmits a classification tag back to the operator or, in fully autonomous mode, initiates terminal dive without human input. The Main Intelligence Directorate of Ukraine's Ministry of Defence (HUR) confirmed the system can lock onto moving targets. [Source]

That last capability, moving-target engagement, is the step change. A standard Shahed-136 follows waypoints to a static coordinate. If the target moved, the drone hit nothing. The Geran-2 MS can track a vehicle, a train, or a repositioned air defence system in its terminal phase, correcting its trajectory using the Jetson's real-time video output. This converts a loitering munition from a terror weapon against fixed infrastructure into a mobile-kill asset against manoeuvring forces.

The $20K calculus

The economics are the part that keeps procurement officers awake. The Shahed-136 in its basic configuration costs roughly $20,000 per unit in Russian mass production at the Alabuga facility in Tatarstan. The Geran-2 MS variant, with the Jetson module, IR camera, FPGA, and upgraded Nasir 8-channel satellite navigation system, pushes the unit cost closer to $80,000, but that is still 40 times cheaper than a Kalibr cruise missile at $3.25 million, and 160 times cheaper than a Patriot Advanced Capability-3 (PAC-3) interceptor at $4 million per shot. [Source]

The cost exchange ratio is where the strategic problem lives:

The Gepard self-propelled anti-aircraft gun (SPAAG) row is the only one where the defender wins on cost: a twin-35mm cannon firing proximity-fused ammunition can destroy a Shahed for roughly $30,000 per engagement. But the Gepard is a rare platform: Germany gave Ukraine 52 units, and production is not scaling. For every other air defence system on the Ukrainian roster, a single Geran-2 MS forces the defender to spend 12 to 200 times what the attacker paid.

Russia is not firing single drones. By late 2025, production at Alabuga reached 170 units per day, enough to sustain a nightly wave of 80-120 sorties across multiple launch axes. Over 26,000 Geran-series drones have been deployed since mass production began. [Source]

A Ukrainian Gepard SPAAG, one of the most effective kinetic countermeasures against Shahed-class drones. Photo: Wikimedia Commons, CC BY 4.0

From paper to 170/day

The evolution from basic Shahed to AI-equipped Geran-2 MS happened across 18 months of iterative field upgrades:

September 2022. First confirmed Shahed-136 strikes on Ukraine. Iranian-era navigation using commercial GPS modules. Warhead: 50 kilograms. Cost: approximately $20,000 per unit.

September 2023. Russian-manufactured variant at Alabuga introduces a hardened airframe using fiberglass-and-carbon-fibre composite. Warhead upgraded with tungsten shrapnel fragmentation for area effect. Cost rises to $30,000.

May 2024. High-yield variant appears with a 90-kilogram warhead, trading range (reduced from 2,500 to 650 kilometres) for blast effect. A 52-kilogram thermobaric option is confirmed for urban strike missions.

September 2024. A Starlink terminal is recovered from a downed Geran-2 near Zaporizhzhia. The drone carried the satellite communications dish behind the warhead, wired directly into the flight controller. This was the first evidence of real-time beyond-line-of-sight control: the operator could see the camera feed and steer the drone to a target mid-flight. [Source]

June 2025. Geran-2 MS variant with Nvidia Jetson Orin, infrared thermal camera, Xilinx FPGA, Ezcap HD video encoder, and 4-element Controlled Reception Pattern Antenna (CRPA) array recovered near Sumy. The Ukrainian HUR publishes a detailed component breakdown showing over 60% of the electronics by value are Western-manufactured, sourced through sanction-evading intermediaries. [Source]

January 2026. Three Starlink-controlled Geran-2 MS drones strike a moving train near Kharkiv, the first confirmed combat use of the fully integrated AI-and-satellite-communications stack to engage a mobile target. Satellite imagery and Ukrainian railway operator reports independently verified the attack. [Source]

Early February 2026. A reconnaissance variant is recovered with a Raspberry Pi 5 single-board computer running video processing alongside a Windows 11 Mini PC. This hybrid architecture suggests the Alabuga facility iterates faster than any traditional defence contractor because it treats drone design like consumer electronics development.

Six variants in 18 months. From simple GPS-guided nuisance to swarm-deployed AI hunter-killer. The upgrade cycle is shorter than most smartphone generations.

Cat and mouse

Ukraine's counter-drone response has been equally fast. The Sky Fortress acoustic detection network (over 10,000 microphones across Ukrainian cities, each costing less than $500) creates a low-cost passive detection mesh that tracks Shahed flight paths by engine sound signature. Mobile fire groups with heavy machine guns and thermal sights provide the last layer of point defence. The Gepard SPAAG remains the most effective kinetic countermeasure, but with only 52 units, coverage is concentrated around critical infrastructure nodes. [Source]

The Geran-2 MS introduces three problems for defenders that the basic variant did not:

The CRPA antenna array is the first headache. It resists jamming by electronically steering its reception null toward the jammer source while maintaining satellite lock. Standard Global Navigation Satellite System (GNSS) jamming, Ukraine's most widely used electronic warfare (EW) technique, becomes significantly less effective against the MS variant because the CRPA can characterise the interference pattern and filter it out.

Then there is the Jetson's ATR problem. It enables GPS-free terminal guidance. Even if the drone loses all satellite positioning in the final 20 kilometres, the computer vision pipeline can identify the target by visual signature and steer toward it. This eliminates the most effective EW tactic against basic Shaheds: spoofing the GPS coordinate to redirect the drone to empty fields.

And Starlink adds a third dimension. It provides a command-and-control link on an entirely different frequency band (Ku/Ka, 12-40 GHz) from the drone's standard radio frequency (RF) control links (typically 900 MHz to 5.8 GHz). Ukrainian EW systems tuned to jam drone-control frequencies do not affect the Starlink downlink. The only countermeasure is physical destruction of the Starlink terminal, which is inside the drone, behind the warhead, and airborne.

Ukraine's most innovative counter-response has been AI interceptor drones. By July 2025, Ukrainian interceptor teams reported a 9-out-of-10 kill ratio against Shahed variants in contested airspace. The interceptors use their own computer vision to acquire the target and perform a kinetic ramming intercept. This inverts the cost asymmetry: a $500 FPV racing drone interceptor destroying an $80,000 Geran-2 MS is an 8:1 defender advantage. The question is whether this approach scales to swarm density.

The American LUCAS reverse-engineered Shahed-136 variant, demonstrating how quickly the platform has proliferated. Photo: U.S. Department of Defense, Public Domain

The India Angle

The relevance for Indian defence planning is direct and uncomfortable. Every capability demonstrated by the Geran-2 MS (AI-powered terminal guidance, moving-target lock, Starlink-grade beyond-line-of-sight (BLOS) control, production at civilian-factory scale) is available to any state or non-state actor with access to commercial electronics markets.

The Shahed-136's 2,500-kilometre range places most of India's strategic depth within launch distance of adversary territory. Pakistan's existing drone infrastructure, combined with Iranian Shahed technology and potential Chinese electronics supply chains, creates a proliferation pathway that no single countermeasure can block. The Observer Research Foundation (ORF) has specifically identified loitering munition proliferation as a blind spot in India's current defence procurement framework [Source].

Countermeasures in the Indian context face specific constraints. India's primary area air defence systems (the Akash surface-to-air missile (SAM), the S-400 Triumf deployed in squadron strength since 2022, and the Barak-8 naval SAM) were designed to engage aircraft, cruise missiles, and ballistic missiles. None are optimised for drone swarm interception. The S-400's 40N6 interceptor is a $2 million missile aimed at a $20,000 target: a 1:100 cost ratio that cannot be sustained over a multi-wave engagement.

The Ukrainian acoustic detection model ($500 microphones forming a passive sensor mesh) is directly applicable to India's border monitoring architecture. Distributed acoustic sensors along the Line of Control (LoC) and Line of Actual Control (LAC) could provide the low-cost early warning layer that complements India's existing ground-based radar network. The Indian Army's existing network of battlefield surveillance radars (BEL BFSR-SR) provides some drone detection capability, but the combination of acoustic plus radar plus electro-optical tracking is what Ukraine has demonstrated works against Shahed-class targets. [Source]

The offensive side matters equally. India has no Shahed-class loitering munition in production. The DRDO's Archer-NG unmanned combat aerial vehicle (UCAV) is a different class: a flying-wing stealth platform designed for stand-off strike with precision-guided munitions, not a mass-produced loitering munition. The NAL TAPAS-BH (Tactical Airborne Platform for Aerial Surveillance, Beyond Horizon) is a medium-altitude long-endurance (MALE) surveillance platform, not a munition. There is a clear capability gap between the drone systems India is procuring for reconnaissance and the low-cost, high-volume loitering munition that adversarial forces are fielding at 170 units per day.

The Geran-2 MS is not a silver bullet. It is a cheap airframe with a clever computer bolted on. What makes it strategically significant is the production volume and the feedback loop: every recovered wreck teaches the Alabuga engineers what worked and what did not, and the next batch is already on the assembly line with the fixes applied. A defence procurement cycle measured in years cannot pace a drone iteration cycle measured in months. That is the problem that has no missile solution. And it is not clear that any amount of spending on interceptors will solve it.

Sources: UNITED24 Media, Stefan Nikolaj, Hammer Mindset, Wikipedia, Forbes, Militarnyi, Kyiv Independent, GUR Ukraine, ISW, ORF