On the night of July 12, the US military deployed explosive-laden surface drones in actual combat for the first time in its history. Three one-way attack vessels made a low-speed, uncontested approach to an Iranian Ghadir-class midget submarine – suspended out of the water on a gantry – and a ship maintenance facility at Bandar Abbas Naval Base before detonating. US Central Command released video of the operation and confirmed on social media it was the «first time American forces have employed sea drones in combat operations.»

The seemingly modest technical detail – the «low-speed, uncontested approach» reported by USNI News – is actually the most revealing. In a congested theater like the Strait of Hormuz, the fact that three small unmanned boats could reach a sensitive target without being intercepted shows that the real advantage is not speed, but stealth and the ability to saturate defense systems. It’s a dynamic we already saw in Houthi attacks in the Red Sea, now institutionalized by a top-tier offensive navy.

For those dealing with defense computing architectures, this episode raises the bar. Suicide drones are nothing new – Houthis have used them for nearly a decade – but the leap comes from integrating them into command chains that demand real-time situational awareness, coordination among heterogeneous units, and, eventually, autonomous decision-making. All of this requires distributed computing power capable of operating in environments with intermittent or no connectivity, where the cloud simply does not exist.

This is where sovereignty enters the picture. Even in their simplest radio-controlled form, naval drones are platforms that collect and process data – radar signals, thermal imagery, telemetry – close to the target. As autonomy grows, the amount of local inference will explode. You cannot offload the decision of whether an unknown vessel is a threat to a remote data center: milliseconds count, and the security of the communication channel is everything. That’s why Western naval modernization programs are pivoting toward on-premise and edge hardware stacks, where models run directly on board, in ruggedized containers with zero third-party dependency.

The structural shift for the defense industry is twofold. On one hand, the miniaturization of computing power – systems that must fit in hulls just a few meters long, resist salt spray, and draw minimal power – is driving ARM and optimized FPGA architectures, not datacenter GPUs. On the other, the sovereignty requirement becomes absolute: every component of the software chain must be verifiable, updatable without external connections, and tamper-resistant. It is no coincidence that tenders for unmanned naval equipment are increasingly specifying end-to-end encryption and air-gapped execution.

What loses ground are approaches that assume always-on connectivity. Operating models that rely on cloud APIs for inference or on centralized orchestrators collide with the reality of an operational environment where an adversary’s jammer is the first actor on the field. The July 12 attack, while not an autonomous operation in the strict sense, is proof that doctrine is moving toward expendable, distributed, low-unit-cost systems – systems that, to work, must carry their own brain, not just a modem.

On the hardware front, the lack of technical specs in the official announcement is no surprise. Armed forces do not disclose component details, but the direction is clear: the next generation of naval drones will have onboard processing capabilities we currently associate with edge servers for industrial IoT, with an emphasis on redundancy and formal software certification. This trajectory intersects directly with the evolution of on-premise LLMs: the same requirements for latency, security, and control that push enterprises to bring models inside their own perimeters apply – multiplied – to a destroyer or a flotilla of USVs.

The lesson for those building computing infrastructure is that the defense market will never seek cloud-dependent solutions for critical functions. Procurement programs are already demanding containerization, support for quantized models, and digitally signed firmware update pipelines that can only be replicated in isolated labs. The Bandar Abbas operation is a starting point: it showed the concept works; now it’s up to the industry to make those platforms increasingly intelligent, without ever breaking the sovereignty constraint.