The drone war in Ukraine has become an AI arms race as both sides rush to deploy AI-enabled systems which are immune to radio interference, making protective jammers useless, and which can find and attack targets on their own.
This arms race is driven by hardware from the world’s biggest company, NVIDIA. Sanctions should prevent Russia from acquiring NVIDIA hardware, but their chips have been found as key components in the latest Russian, with several different types deployed all using NVIDIA hardware.
The Hardware Giant
NVIDIA is the world’s largest company by market value, and the first ever to break the $4 trillion barrier.
The company is incredibly successful, with an estimated an estimated 85% of the global AI chip market, because it makes what everyone wants: powerful hardware to drive AI. The chips, known as GPUs (Graphics Processor Units) or accelerators, differ from the typical computer chip or CPU (Central Processing Unit) in being able to handle lots of small tasks simultaneously rather than applying more power to a few tasks. This is parallel processing, and it is essential for most types of AI, which involve huge datasets.
Their success is reminiscent of the chip wars of the 1980-90s when desktops PCs ground slowly through large spreadsheet calculations and other complex tasks. CPU power was vital and Intel rose to dominance, with ever-faster clock speeds and transistor counts in their 386, 486 and Pentium processors. NVIDIA is doing the same with GPUs in the 2020s, producing every more capable versions and outpacing the competition.
NVDIA has several different families of chips for different applications, including high-power units for data centers and compact, low-power Jetson boards for edge devices like consumer electronics – and drones. Such single-board computers cost just a few hundred dollars.
This is highly capable hardware. In 2021, a team from University of Zurich (UZH) led by Davide Scaramuzza demonstrated an AI system using a Jetson computer on a racing drone which was able to beat world-class human pilots for the first time.
This system relied on external motion trackers to give the drones data, giving them an unfair advantage. But by 2023 , the UZH team had developed an AI system which was able to beat human champions using just onboard sensors and processing.
“We are very excited as this is the first time that AI beats a human in a physical sport designed by and for humans,” Scaramuzza told me at the time.
The feat was replicated recently in Dubai when an AI-enabled drone from TU Delft, having beaten all the other AI drones, competed head-to-head against the human FPV racing winners in an AI vs Human Challenge, and won. Again, the teams used Jetson-based computers.
“Even in our earlier work, we had more than enough computational headroom,” Scaramuzza told me.
The earlier version used the older Jetson TX2. Now drones have the new and more capable Jetson Orin which offers at least ten times as much computing power.
“Some of the algorithms we developed for our drone racers have found their way into companies like Skydio and Zipline, where several of my former students now work,” says Scaramuzza.
While there is no indication that the software developed by UZH is being used Ukraine, others are certainly using AI-enabled systems powered by the same hardware, and which may be equally capable in terms of matching human operators.
Russia’s Digital Predators
Back in 2023 Russian Lancet attack drones – a 35-pound weapon with a reach of 25 miles and a warhead capable of taking out a tanks – were found to have an NVIDIA Jetson TX2 ‘brain.’ Smugglers reportedly ship the chips in small batches labelled as other components, sending them via several third-party countries to reach Russia.
The Jetson drives AI giving Lancet ‘lock on target’ function allowing the operator to designate a target within the field of view. The Lancet then track the target, following it and running into it even if the communication link is lost.
In 2024 this automated system appeared to be performing poorly, for example hitting a shadow next to the target rather than the target itself. For a time the function appeared to be disabled. But it was restored, presumably after software upgrades, and appears to have improved considerably. From data published on Russian weapon performance tracking site LostArmour in 2024 about 30% of Lancet hits involved automated guidance , that figure is now up to almost 60%.
Now two new types of Russian drone have recently been found with the Jetson Orin. A third has entered services which likely uses the same technology.
One of these was a new version of the Shahed attack drone, known as MS001 which in addition to the NVIDIA processor also had a thermal imager and digital modem.
In a LinkedIn post, Ukrainian Major General Vladyslav Klochkov said that this was not simply a drone.
“This is a digital predator. It doesn’t carry coordinates, it thinks,” stated Klochkov
This is an exaggeration – the MS001 also has a satellite navigation system, so clearly it does carry co-ordinates. But unlike the basic Shahed which relies entirely on satellite navigation, the MS001 can identify objects on the ground using its thermal imager and AI and attack them.
The second new drone is the smaller V2U. A report in Ukraine’s Defense Express says this four-winged drone, similar in layout to the U.S. Switchblade 300 but carrying an 8-pound warhead, has a range of over 25 miles. The V2U has a high-resolution camera and a laser rangefinder, which, like the TERCOM system in the Tomahawk cruise missile allows it to navigate by comparing the terrain to a digital map. Like the MS001 it has a digital modem which connects to Ukraine’s cellphone system to communicate with the operator.
Again, the V2U is powered by an NVDIA Jerson Orin, which allows it to fly up and down roads looking for targets.
“They don’t distinguish between military equipment and a civilian bus,” according to one Ukrainian report.
According to other reports cited by Ukrainian electronics expert Serhii Flash the V2U works in teams, with each team member having different color makings on their wings. This likely allows the drones to distinguish each other and carry out their attacks in sequence without conflict and without needing radio communication. According to Flash, the drones are stacked one above another like circling vultures, so for example the ‘blue’ drone waits its turn until the ‘red’ drone has attacked. This is a basic version of swarming behavior but a significant step forward.
Flash also notes that the V2U has limited intelligence when it comes to target discrimination and one attacked a public toilet rather than a vehicle.
The third new AI Russian drone is known as Tyuvik (“Levant sparrowhawk”) which resembles a scaled-down Shahed and is now in mass production. It has a range of 20 miles and carries a 4-pound warhead; it can find and attack striking moving vehicles and its intended targets are armored vehicles. No example has been captured and analyzed yet, but like V2U this flies to a specified location and then finds a target using machine vision. The makers Statim say Tyuvik is built from low-cost commercial components. Given that Russia does not make any suitable AI hardware, this again suggests an off-the-shelf Jetson Orin.
The End Of ‘Dumb’ Drones?
All three drones may use similar software for navigation and target location, like the portable Prism software from FLIR now being integrated on U.S-made attack drones or the open-source software developed by Auterion. (A shipment of 33,000 new Auterion Skynode strike drone systems to Ukraine was announced recently). This type of setup allows new type of drone to be turned into autonomous ‘digital predators’ rapidly and at minimal cost. And as the software for navigation, flight and targeting is improved, the improvements can be shared across every drone type in the fleet. New functions, such as dogfighting or swarming, can be added as needed.
We are at the dawn of the age of AI drones. Thanks to the ready availability of NVIDIA Jetsons, such technology is available to literally everyone. This genie is now very much out of the bottle. In the near future, ‘dumb’ drones lacking onboard AI may be as outdated as biplanes.
