Good morning,
After a small hiatus, Sovereign Systems is back with a world-class team to deliver the latest defence tech news more often than before – so our readers keep their finger on the pulse.
This week, we look at a fundamental shift in how defence AI systems are built. Synthetic data is moving from a workaround to core infrastructure—unlocking billion-scale training pipelines and removing one of the sector’s most persistent bottlenecks.
We go inside the architecture making that possible, and what it means for autonomous systems across drones, robotics, and next-generation military platforms.
Alongside this, we highlight the growing momentum across Europe’s defence ecosystem—from new funding instruments accelerating deployment timelines, to the unmissable deftech event of the year, DTM.Defence, bringing together builders, primes, procurement agencies and investors shaping the next wave of European capabilities.
The throughline is clear: defence innovation is becoming faster, more software-defined, and increasingly unconstrained by the physical limits that once slowed it down.
Enjoy the read,
The (newly upgraded) Sovereign Systems Team
EVENT HIGHLIGHT
DTM.Defence Will Be The Unmissable Gathering for European Defence and Dual-Use Tech.
If you’re building, procuring or funding defence and dual-use tech in Europe, this is the room to be in.
On May 20-21 in Berlin, at DTM.Defence, more than 500 senior leaders from NATO armed forces, defence primes, policymakers, defence and dual-use startups and scaleups, investors, and procurement agencies will engage in-person with one mission in mind: to procure, deploy or fund European defence and dual-use innovation.
Beyond the highly curated one-on-one matchmaking, the stage will feature speakers from Europe’s top deftech unicorns and scaleup founders, leading military voices and high-level policy makers, including:
Sven T. Heursch (CDO, HENSOLDT)
Jan-Hendrik Boelens (CEO, Alpine Eagle)
Ricardo Mendes (CEO, TEKEVER)
Max Gulde (CEO, constellr)
Sven Przywarra (Co-CEO, LiveEO)
Sir Ben Wallace (former UK State Secretary)
General Sir Patrick Sanders
Lieutenant General Sir Tom Copinger-Symes
Lieutenant General General (ret.) Chris Badia
Lieutenant General Frank Leidenberger
🪖 A special Signals From the Frontline roundtable series – co-shaped with the NATO Innovation Fund (NIF) and senior defence leaders – will translate frontline needs into clear capability signals for Europe’s defence and dual-use innovators—creating a transparent, repeatable NATO innovation pipeline to accelerate capability development for European defence.
Leading defence primes, neo primes and scaleups at DTM.Defence include
Airbus, Rheinmetall, Thales, Safran, Rolls-Royce, BAE Systems, MBDA, Rohde & Schwarz, Saab, HENSOLDT, TKMS, KNDS, FINCANTIERI, RENK Group, RUAG, TERMA, OHB SE, IABG, DEUTZ, Vitesco Technologies, FEV Group, BWI GmbH, Aumann Group, MDA Space, Aurora Flight Sciences, ST Engineering, Roketsan, Space Norway, Kayser, ARX Robotics, Destinus SA, Constellr, Starlab, Nex Aero, Intelic BV, NIF – and many more.
MODs, procurement agencies and armed forces will also be in the room, including NATO, Bundeswehr, UK Ministry of Defence, Ministero della Difesa, Dutch Ministry of Defense, Danish Ministry of Defence, Belgian Defence Headquarters, Hellenic National Defence General Staff, Armed Forces Sweden, Czech Air Force, British Army, Estonian Defence Forces and many more.
→ Learn more about DTM.Defence
TECHNOLOGY DEEP-DIVE
Inside the Architecture of Billion-Sample Synthetic Data Generation
For teams building computer vision models for drones, autonomous vehicles, robotics, and IoT devices, the hardest question is rarely about the neural network architecture. It is about the data. Synteza.ai, a London-based startup founded in 2024, is building the infrastructure that turns that question from a fieldwork problem into a software problem.
How the Pipeline Works
Synteza’s platform generates training datasets from three source types: photogrammetric captures, generative AI models, and custom 3D assets. Each path produces a different fidelity and flexibility profile, and the platform allows teams to combine them within a single dataset.
Photogrammetric sources begin with real-world objects captured using structured light or photogrammetry, producing meshes with high geometric accuracy. These serve as the physical anchor—the geometry is real, even if everything around it is synthesised. Generative sources use diffusion-based models to produce object variations and scene components that would be prohibitively expensive to capture physically. Custom 3D models allow defence and industrial clients to work from existing CAD assets, maintaining proprietary geometry without any physical scanning.
From those source objects, the platform procedurally generates scene compositions at scale. The parameters that can be swept programmatically include: lighting (time of day, overcast diffuse, point source, infrared simulation), surface materials and textures (wet, weathered, camouflaged), environmental conditions (fog, rain, dust occlusion), camera parameters (focal length, resolution, sensor noise profiles), and physical interaction states (motion blur, partial occlusion, object stacking).
Every generated frame is automatically annotated. The platform outputs bounding boxes, segmentation masks, depth maps, and keypoint labels as standard. Because the scene is procedurally generated, ground truth is exact—there is no human labelling error and no ambiguity at object boundaries. This eliminates the annotation bottleneck entirely.
The Scale Claim: One Billion Datasets Per Day
Synteza states it can generate one billion ready-to-use annotated datasets per day. This is not a throughput number for a single workstation—it reflects a parallelised rendering infrastructure operating across distributed compute. The practical implication is that teams can specify the long tail of edge cases they care about and receive full coverage within hours, rather than planning months-long field collection campaigns.
The more operationally significant capability is compositionality. Each combination of object geometry, lighting state, camera parameter, and environmental condition constitutes a distinct training sample. When you sweep across even a modest parameter grid, the combinatorial product reaches the billion range quickly. The question is not whether the number is achievable—it is whether the generated variation actually corresponds to the distribution shift the deployed model will encounter in the field.
Domain Gap and the Realism Problem
The longstanding objection to synthetic data is domain gap: a model trained on rendered imagery and deployed on a real camera learns features of the renderer, not of the world. Synteza’s approach to closing this gap operates on two levels.
First, physical accuracy. The platform uses physically based rendering (PBR) to model how light interacts with surfaces, including subsurface scattering, specular highlights, and material-specific reflectance. This matters most for detection tasks in challenging illumination—an area where synthetic data has historically performed poorly against real-world captures.
Second, domain randomisation. Rather than trying to make synthetic data look perfectly real, domain randomisation deliberately varies the rendering parameters across a wide range during training. The hypothesis—empirically supported in robotics and autonomous driving research—is that a model trained on sufficiently diverse synthetic conditions generalises to the real domain as a subset of that diversity. Synteza’s parameter sweep architecture is, structurally, a domain randomisation engine at scale.
The Marketplace Layer
Beyond generation, Synteza is building a dataset marketplace where teams can buy and sell generated datasets. This is strategically important for two reasons. First, it lowers the barrier for smaller teams that cannot afford to run the generation pipeline at scale themselves. Second, it creates a network effect: as more domain-specific datasets are generated and listed, the platform becomes the reference source for annotated training data across verticals—a position analogous to what Hugging Face holds for model weights.
The defence and aerospace relevance is direct. Sensor data for military platforms is not available on open datasets. Drones operating in contested environments, autonomous ground vehicles navigating degraded terrain, and maritime surveillance systems processing infrared imagery all require training data that simply does not exist publicly. A synthetic generation pipeline that can produce mission-specific, sensor-specific, scenario-specific datasets removes the data access barrier for the companies building these systems.
Why It Matters for the Deep Tech Ecosystem
The bet Synteza is making is that the synthetic data generation layer will be as infrastructure-critical to the AI era as cloud compute was to the software era. If that bet is right, the companies that own the generation pipelines and marketplace networks will sit upstream of the entire computer vision stack—extracting value every time a model is trained, regardless of which architecture or hardware platform is used.
For the deep tech founders building autonomous systems in defence, mobility, and robotics, the near-term implication is practical: the days of needing a data collection programme before you can start model development are ending. Generation-first workflows—where you specify what you need and synthesise it at scale before any sensor is deployed—are becoming the default. Synteza is one of the companies building the tooling for that shift.
ECOSYSTEM UPDATES
EU Backs Defence Innovators with €115 Million Under AGILE — April 8
Image Credits: Munich Security Conference
EU Backs Defence Innovators with €115 Million Under AGILE — April 8
The European Commission formally launched AGILE, its Programme for Agile and Rapid Defence Innovation, on April 8. The €115 million fund is designed to get disruptive technologies—AI, drones, quantum, robotics—from development to armed forces deployment within one to three years. The application-to-grant timeline is four months, with up to 100% cost coverage and a retroactive clause allowing companies to claim expenses incurred up to three months before the application date. Targeting 20 to 30 projects, AGILE is explicitly designed for SMEs and startups rather than established primes.
Thank you for reading this issue of Sovereign Systems. Here are three ways to stay connected:
1. Become a Guardian of European Deep Tech: Are you a senior leader seeking commercial partnerships with Europe’s leading Deep Tech startups and SMEs? Successful applicants join Deep Tech Momentum’s Guardian network, receiving complimentary VIP access to our flagship event in May. Apply here.
2. Sync the EDTH calendar: Ensure every European Defence Tech Hub hackathon and meetup is on your radar. View the full schedule.
3. Send this newsletter to a friend: They can subscribe here.