In the southern Red Sea the Italian Navy, holding tactical command of the European Union’s defensive mission EUNAVFOR ASPIDES established by Council Decision CFSP 2024/583, wins very nearly every engagement it fights. It is also losing the arithmetic. Those are two separate facts, and only one of them makes it into the press releases.
The arithmetic is not a matter of opinion. Work by the RAND Corporation puts the cost of a defensive interceptor at somewhere between 75 and 130 times the cost of the hostile drone it destroys. The doctrinal reply to that ratio is correct and should be stated first: the value of an interceptor is not measured against the price of the drone, but against the crews and the commercial shipping it protects, frequently container vessels worth 150 to 200 million dollars, on routes that keep Italy supplied. Nobody sensible argues for trading a ship to save a missile.
The trap is elsewhere, and it is logistical rather than financial.
The magazine, not the missile
According to the operational profiles of the Italian FREMM frigates, a single coordinated salvo of eight hostile munitions can consume around half of a ship’s available vertical launch cells in two minutes. One engagement can therefore exhaust a frigate’s air defence capacity. And because vertical reloading in the open sea has never reached genuine operational maturity in Western navies, replenishment means leaving the theatre for an equipped base at Djibouti or Augusta.
Read that as the adversary reads it. Saturate the defence, and the defence removes itself from the strait. It does not need to be defeated. It needs to be made to leave, repeatedly, on a schedule somebody else sets.
Behind the ship sits the industrial constraint, which is worse because it cannot be fixed at sea. Average production and stockpiling lead time for the Aster interceptor family runs between 18 and 24 months. Against sustained operational consumption that produces a permanent deficit. European defence industry replenishes on a two year cycle; the militias launch at a rate that exceeds it.
This is the shape of the Millennium Challenge 2002 lesson that keeps being re-learned: a technologically superior force built around a small number of hyper connected platforms can be overwhelmed by quantity when the adversary simply refuses a symmetrical fight and buys cheap, decentralised mass instead.
What the model actually did
Onto that came the story that drove the debate: an interdiction reported by Anthropic’s threat intelligence unit against a Yemeni insurgent cell, which according to that analysis had been using frontier language models to accelerate the development of remotely guided systems.
The distinction here matters more than the headline, and most commentary has flattened it. Frontier autoregressive models have computational and memory requirements that put them nowhere near a launch vehicle. Their role in this case was analytical: compressing the time needed for reverse engineering, software assembly and simulation work, on rented commercial cloud capacity, at a cost close to nothing.
The models did not manufacture anything. They did not produce turbines or warheads. They bought calendar time. In this theatre, calendar time is the weapon, and the cell obtained it by the hour on a civilian service while the alliance opposite acquires it by the decade.
The tactical layer is a separate problem. Guidance on board a vehicle requires small quantised vision models running disconnected on low cost commercial hardware, and the Red Sea punishes exactly that class of component: commercial visual odometry struggles over a moving sea surface, extreme heat and sand degrade cheap sensors and optics. Which is a real limitation and a thin comfort, since the attacking side can afford to lose most of what it launches and the defending side cannot afford to intercept all of it.
What is not the obstacle
There is a fashionable argument that Western navies are slow because lawyers and standards have tied their hands. It is worth disposing of, because it is wrong and because it is expensive.
Legal review of new means of warfare, required by Article 36 of the First Additional Protocol to the Geneva Conventions, and functional safety standards such as DO-178C, STANAG 4671 and MIL-STD-882E, are not bureaucratic decoration. In ballistics and point defence, fire control software has to remain deterministic and formally verifiable. Dropping a probabilistic network capable of confident error into a firing chain, inside a strait dense with civilian traffic, is the fastest available route to killing merchant seafarers with a friendly missile. The Houthis are unconstrained by any of this precisely because they are not accountable to anybody. That asymmetry is not a standard to aspire to.
The obstacle is a multi layered procurement structure that struggles to absorb even the urgency derogations national security law already grants it. The rules permit speed. The apparatus cannot metabolise it.
And underneath that sits a cultural preference legible in the Defence Ministry’s own multiyear planning document: heavy naval infrastructure is what gets funded and defended, embedded software is what gets deferred. In a domain where onboard compute and edge vision models turn over on six month cycles, ten year acquisition planning delivers hulls whose digital architecture was obsolete before commissioning. Buy steel on a twenty year horizon with the computing architecture of twenty years earlier and you have not bought a warship. You have bought a monument with a crew.
The models that already exist
The uncomfortable part for European planners is that none of this requires invention. The United States Fifth Fleet’s Task Force 59 has already demonstrated the value of integrating commercial uncrewed surface vessels with edge vision algorithms for persistent patrol. Within the alliance, the NATO Artificial Intelligence Strategy and the annual REPMUS and Dynamic Messenger exercises have been validating interoperable open swarms for years, in public.
So the capability is proven, the doctrine is written, the exercises have run, and the acquisition pipeline still delivers on the timeline of a bridge rather than a laptop.
For a navy charged with protecting national interests across the wider Mediterranean, the lesson of Bab el-Mandeb is not that safety standards must be relaxed for the sake of readiness. It is that the two were never actually in conflict, and that the opposition between them has been a convenient explanation for an institutional inability to buy software at the speed software exists.
Sea lines of communication will be defended by putting cheap expendable uncrewed systems and on-premise edge computing on the water quickly, alongside the steel rather than instead of it. Everything else is a procurement cycle explaining, with excellent documentation, why the magazine is empty.
Raffaele Di Marzio
All my “insane” books on cybersecurity and governance are here 👇 https://cyberium.limited/bookshelf.html