Drone Swarms Are Getting Smarter. The Software Underneath Has to Keep Up.

Market numbers for military drones vary. Global Market Insights estimates the military drone market at approximately $18.2 billion in 2025, rising to $66.5 billion by 2035; SNS Insider projects a rise from $16.5 billion to $42.8 billion. Precedence Research puts it closer to $17 billion, growing to $36 billion. Depending which source you use, that's a compound annual growth rate anywhere between 7% and 14%. We'd treat any single one of these figures as directional rather than precise, but the directionality itself was consistent across every source we looked at: this market is getting bigger, and autonomy is the key driver.

What's changing on the ground. The precise capability fuelling that expansion is autonomy that doesn't rely on GPS. The next defence-spending wave is being built around AI-powered drones — increasingly, drones use onboard cameras, sensors and SLAM techniques to navigate relative to their surroundings rather than to a satellite signal. GPS is one of the most vulnerable systems in a contested environment, and a system that stops working the moment it loses signal isn't much of a system. Reporting from Army Recognition and several drone-industry trade publications this year describes a broader shift toward distributed, containerised drone deployment designed to sustain operations precisely under those jammed, GPS-denied conditions.

Swarm coordination is the harder half of the same problem. Getting a handful of drones to share a map, distribute tasks and adapt to a changing situation without a single control node is still, by most technical accounts we reviewed, an open research area rather than a solved one. The challenges cited most consistently: reliable communication over lossy links, collision avoidance in cluttered airspace, and the well-known gap between how a swarm behaves in simulation versus in the field.

Where the friction actually is. If the embedded software running that autonomy can be compromised, none of it matters. Memory vulnerabilities — buffer overflows, use-after-free bugs — have long been a known attack surface in embedded systems, and defence and aerospace programs are no exception. As decision-making moves onto the device itself rather than back to a human operator, the cost of a compromised firmware update or an exploitable memory bug rises accordingly. It's a structural problem that's less headline-grabbing than the drones themselves, which may be exactly why it gets less attention.

Innovation in action. RunSafe Security, a Critical Ventures portfolio company based in the US, works specifically on this layer. Its technology relocates software functions in memory at runtime, closing off the class of memory-based exploits that has historically been difficult to eliminate without rewriting the underlying code. RunSafe's cyber-hardening is deployed across aerospace, defence, energy and industrial-automation programs, and it generates software bills of materials (SBOMs) that give operators visibility into what's actually running on a device — a requirement that's become increasingly central to how governments and primes evaluate embedded software risk. Read our full investment thesis here.

We can also mention Rhizome Labs, a Paris-based portfolio company creating adaptive edge-AI models that keep learning on-device — on drones, sensors and industrial hardware — without having to phone home to a cloud server for retraining. Rhizome has design partners who count Safran among others, giving it a direct line into just the kind of aerospace and dual-use hardware this shift toward on-device autonomy depends on. Through the two companies, you get a fairly complete picture of what needs to be true for autonomous systems to operate safely at the edge: intelligence that lives on the device, and software underneath it that can't be quietly compromised. Read our thesis on Rhizome here.

We invest in defence and dual-use technology because the interesting problems there are rarely about the weapon itself — they're about trust, verification and resilience in systems that increasingly make decisions without a human checking in on every one. That's true of a counter-drone interception system as much as it is of a hospital's medical device fleet. Founders who understand that overlap, and who are building the unglamorous infrastructure layer rather than just the visible product, are the ones we want to meet.

If you're working on embedded security, on-device autonomy, or resilient communications for contested environments, get in touch.