Scaleout Systems, a Swedish startup selected for NATO's Defence Innovator Accelerator for the North Atlantic (DIANA) Challenge Program in 2025, is deploying federated machine-learning models on drones and forward-base hardware to let military units identify and strike targets without relying on a central data center. The company demonstrated the approach in January at the Winter Demo 2026 in Sweden, where a drone using onboard AI autonomously picked out an armored engineering vehicle as its highest-value target and delivered an explosive to it. A follow-on test in June ran at a Swedish Air Force base in Uppsala, where the Swedish military already licenses Scaleout's main software platform.
Founded at Uppsala University in 2018, Scaleout initially built machine-learning tooling for commercial trucks and other vehicles. Russia's full-scale invasion of Ukraine in 2022 pushed the team into defense, where the same federated-learning stack could be applied to sensor-heavy platforms operating far from reliable connectivity.
The pitch is straightforward. Frontier models from OpenAI or Anthropic are too large and too dependent on cloud infrastructure to run on a drone or a field workstation. Scaleout instead trains leaner computer-vision models that fit on embedded devices and edge servers, then uses federated learning to update them from data aggregated across multiple units.
“With the war in Ukraine and a shifting world, we realized that this technology can be very important to operationalize edge data and sensor data for machine learning to make sure that NATO allies have found that strategic advantage.”— Andreas Hellander, Cofounder and CEO of Scaleout Systems
Key facts
- 01Scaleout Systems, founded at Uppsala University in 2018, was selected for NATO's DIANA Challenge Program in 2025 to adapt AI for drone edge hardware.
- 02At the Winter Demo 2026 in Sweden, a drone running Scaleout's onboard AI autonomously identified an armored engineering vehicle and dropped an explosive on it without direct human commands.
- 03In June 2026, Scaleout tested federated learning at a Swedish Air Force base in Uppsala, where a forward computing node kept running inference after losing its central link.
- 04The company is part of BAE Systems Bofors' Affordable Loitering Modular Ammunition (ALMA) project, aimed at building a low-cost autonomous kamikaze drone.
- 05Scaleout uses lean computer-vision models on drone hardware rather than frontier models from OpenAI or Anthropic that require centralized compute.
Drones and drone-operator tablets run local inference on incoming camera and sensor data. They send selective model updates — not raw imagery — back to a platoon or company command post, where a heavier node retrains the shared model and pushes the improved weights back out when bandwidth allows. The design keeps sensitive raw data off the wire and keeps the loop functional even when links to a central lab drop.
That resilience matters on a battlefield where jamming and electronic warfare routinely sever communications, and where large data centers are themselves targets. Ars Technica notes centralized AI infrastructure looked more fragile after data centers were hit during the recent conflict between the US and Iran. Ukrainian forces have already been fielding cheap kamikaze drones with onboard AI to work around exactly these constraints.
Environment drift is the other problem federated learning is meant to address. A recognition model trained on desert imagery will underperform in an urban fight, and vice versa. Scaleout's argument is that pushing model updates within a single day or a single operation — trained on the actual sensor stream coming off deployed units — is the durable edge, not raw model size.
“Models might have been trained in a desert environment, and if we try to deploy them in an urban environment, they're not going to perform well.”— Andreas Hellander, Cofounder and CEO of Scaleout Systems
The Winter Demo 2026 event tied that theory to a live weapon. Scaleout is a participant in the Affordable Loitering Modular Ammunition (ALMA) project, headed by BAE Systems Bofors, which is developing a low-cost autonomous kamikaze drone. In the demo, the drone handled detection, identification and geolocation entirely on dedicated onboard computing, with no external processing. A human operator could still take control, but the drone flew the engagement on its own.
The June test in Uppsala pushed the federated architecture. A forward computing node at the air base continued running its own inference and active-learning routines after losing its link to a central node in Scaleout's lab. When the connection returned, the local model updates synced back — the pattern the company wants to scale across allied forces.
Andreas Hellander, cofounder and CEO of Scaleout Systems, framed the endpoint as cross-border: in principle, member states could share model improvements without sharing the underlying sensor data that produced them. That is the political sell for federated learning inside NATO, where interoperability is a chronic goal and data sovereignty is a chronic blocker.
There are open questions the demonstrations do not answer. Autonomous target selection by a drone that then delivers an explosive is exactly the capability that arms-control researchers, human-rights groups and several European governments have argued should require a human in the loop by default. Scaleout's system preserves the option for operator control but does not require it, and neither the ALMA program nor NATO has published a common rulebook for when autonomous engagement is acceptable. Robustness of small vision models against spoofing, adversarial camouflage and misidentification in cluttered environments is also unresolved in public materials.
The broader signal is where defense AI money and attention are moving. The frontier-lab narrative — larger models, more parameters, bigger clusters — does not describe what NATO militaries are buying for the edge. They are buying small models that fit on a drone, federated pipelines that survive jamming, and vendors like Scaleout that can integrate with existing weapons programs such as ALMA. That is a different market from the one OpenAI and Anthropic are optimizing for, and the companies best positioned to win it are the ones already inside NATO's procurement pipeline before the doctrine catches up to the hardware.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
