The real AI arms race is not the one you are reading about.

The most consequential artificial intelligence in modern warfare is not the kind the frontier powers are building. It is the kind their adversaries are buying off the shelf.

Over the past three years, a pattern has emerged across otherwise unrelated theatres. In Ukraine, first-person-view drones costing a few hundred dollars now account for the majority of Russian battlefield losses. In the Red Sea, the United States Navy has expended around 220 interceptors, at an average cost of $4.5 million per round, to defend commercial shipping against Houthi attacks. Earlier this year, an Iranian strike on Prince Sultan Air Base in Saudi Arabia destroyed an E-3 Sentry airborne command aircraft and damaged multiple refuelling tankers, something no adversary had previously managed in the operational history of those airframes. In one coordinated action against targets in the United Arab Emirates, Iran launched 137 missiles and 209 drones in a matter of hours.

Analysts have begun to describe this pattern with a shared term. Precise mass: the use of large numbers of cheap, good-enough weapons to achieve strategic effects that were once the exclusive preserve of small numbers of sophisticated ones. It is the organising idea of the current decade in military affairs, and it rests almost entirely on the technology this essay is concerned with.

The conventional reading of these events is a cost-exchange story. Cheap weapons, expensive defences, an arithmetic that does not work in the long run. The Pentagon’s own acquisition chief has conceded that counter-drone engagements are becoming too expensive even when the least costly available effector is used. This reading is correct as far as it goes, and it has already prompted a modest shift in American and European procurement toward attritable interceptors and directed-energy systems. What it misses is the deeper point.

Two categories of AI, and only one of them matters here

At the other end sits what might be called commoditised AI. Machine vision running on a $400 chipset. Open-source autopilot code. Navigation systems that hold their course well enough when GPS is jammed. Terminal guidance accurate to within a few metres. None of this is new, none of it is classified, and none of it requires a national laboratory to produce. It is the sediment of a decade of consumer drone development, autonomous vehicle research, and open machine learning, now available to any engineering workshop with a soldering iron and an internet connection.

The competitor powers are not trying to match the frontier. They do not need to. A Shahed-136 drone with rudimentary onboard vision and a commodity autopilot is, in operational terms, AI-enabled enough. It will find its target, often enough, at a unit cost of roughly $20,000. Multiply that by a thousand, and the strategic effect is indistinguishable from what a much more sophisticated weapon might have achieved.

This is the real asymmetry of the current decade. The top tier is racing for qualitative edge. The second tier is winning on quantitative saturation using yesterday’s AI, today. Whoever industrialises good-enough AI fastest, rather than whoever builds the most advanced version of it, is setting the shape of modern conflict.

Beyond arithmetic: the attack on decision tempo

Consider what a saturation raid of 300 drones and missiles actually does to a defending force. Each track must be detected, classified, prioritised, and engaged within seconds. Human cognition, however well-trained, does not have the bandwidth to process that volume of information in that window of time. The defending commander is not simply being outgunned. They are being out-thought, by an adversary who has correctly identified that the binding constraint on modern defence is not the number of interceptors available but the number of sound decisions a human being can make per minute.

This is what game theorists would recognise as a regime change in the logic of coercion. For most of the postwar period, credible deterrence required expensive signals. Aircraft carriers, strategic bombers, nuclear submarines on patrol. The cost of producing the signal was precisely what made it believable. Precise mass inverts this logic entirely. The signal is now cheap to produce, and the effect it generates is no longer principally kinetic. It is cognitive. The weapon is aimed at the defender’s capacity to decide, and the kinetic damage is almost a by-product.

The commander’s dichotomy

Once this is understood, a difficult question moves to the centre of the discussion.

A commander facing a modern saturation raid has, in practice, two options. The first is to delegate engagement authority to automated systems. The algorithms make the decisions, applying rules of engagement that were written in peacetime by people who are not in the room. The loop closes fast enough for the defence to function. But the commander has, for those critical minutes, ceased to be a commander in any meaningful sense. They have become an auditor of outcomes generated by software.

The second option is to retain authority, and to insist that every significant decision pass through a human mind. This preserves human agency, which is the foundation of both operational accountability and the international legal frameworks that govern the use of force. But the loop cannot close in time, and the defence is overwhelmed by mass that could not realistically have been processed at human speed. The commander has kept their authority. The price is paid by everyone else.

Neither option is one any serious military officer would freely choose. Both are losses, in different forms. And the choice is being forced on defending forces with increasing frequency, in every theatre where precise mass has established itself.

This is the question that has not yet received the public attention it deserves. It sits well upstream of the usual debates about autonomous weapons and killer robots, because it is not principally a question about what the machines are permitted to do. It is a question about what space remains, at the tempo of modern conflict, for a human being to exercise judgement at all.

The case for decision intelligence

The right response is to change the terms of the question, by investing in the infrastructure that sits underneath the commander. The purpose of a decision intelligence platform, properly understood, is not to make decisions in place of human beings, and not to make them faster for the sake of speed. Its purpose is to fuse the available evidence at the speed it arrives, score its confidence honestly, surface only what a human actually needs to decide, and keep an auditable record of how the decision was reached. Done well, it carves out, within the seconds available, the smallest viable space in which human judgement can still operate. That space is the entire difference between a commander and an auditor.

None of this eliminates the dilemma. There will be engagements in which mass arrives faster than any infrastructure can resolve, and in which some form of delegation will be unavoidable. But the boundary between what must be delegated and what can still be decided is not fixed. It is a function of how well the decision infrastructure is built. A well-designed platform moves that boundary in favour of the human. A poorly designed one, or none at all, cedes ground to the machine by default.

This is the question the advanced democracies ought to be debating, openly and with some urgency. Not whether to build artificial intelligence into their military systems, because that decision has already been made for them by the shape of the current conflicts. The question is how to build it such that the human commander remains the author of their own decisions, at a tempo that would otherwise strip them of authorship altogether. Precise mass has made this an operational problem. Unless it is treated as such, it will become a strategic one, and then a constitutional one, and by then the answer will have been given, by default, in the dark.